Where is everybody? · Working papers

Red team, round 2: the rebuilt model

These are the critics’ reports as filed. The critics were separate runs of Claude given opposing briefs; none is a human specialist. File names refer to the project’s working folder. Where a report cites a brief by subject, such as “the abiogenesis brief”, it means one of the project’s summaries of the published literature. Notes about where working files were stored have been removed; the findings are unedited.

Critic: a statistician

Overall verdict

Claim 1 is the only headline the model earns, and it needs its numbers corrected. Claim 2 ("three live answers, roughly equal") is an artefact of the base priors and should be withdrawn. Under the site convention the base posterior on the six surviving classes equals "delete the contradicted class and renormalise the prior" to within 0.08 in every class (0.03 bits of information). Each of the three answers is a monotone function of one unevidenced prior: P(none other ever) tracks the prior P(N_ever<1); "stuck" tracks the upper bound of W (0.02 to 0.54 across W ranges); "unnoticed" is about 0.9 times p_nonevident. In the project's own 400-draw scan the three are within a factor 2 of each other in 12% of draws (site) and 21% (SSA). Round 1's point-mass objection was moved, not fixed: "unnoticed" requires a median of 10^4.4 independent expansion waves per galaxy to be low-footprint together, and falls from 0.19 to 0.04 if each wave has its own chance of being evident. The base prior w_fast = 0.4 is the abiogenesis brief's posterior after the early-life datum, so L_early counts that datum twice; using the brief's pre-data values (0.09 to 0.23) gives P(none other ever) of 0.64 to 0.74 (site) and 0.45 to 0.61 (SSA) instead of 0.53 and 0.33. The SIA split is again the ratio of prior atoms. The single most important change is to replace claim 2 and Table B's point values with three curves (each answer against its controlling prior), scan bands, and the what-you-must-believe table.

Objections

• fatal / Draft claim 2, 'Three live answers, roughly equal' (fatal for the claim as worded, not for the model)
  • Argument: Rough equality is a property of the base parameter point under SSA (0.33/0.32/0.31). Under site the base is already 0.53/0.25/0.21, a ratio of 2.5. Each answer is controlled by one prior. (a) logit of posterior P(none other ever) regresses on logit of prior P(N_ever<1) with slope 1.03 (R2 0.87 once p_nonevident is added). (b) 'stuck' correlates 0.78 to 0.82 with log W_max. (c) the ratio unnoticed / p_nonevident has median 0.88 (site). Which answer is largest changes with the prior draw and with the anthropic convention.
  • Proposed change: Delete 'roughly equal'. Replace the three point values with (i) 5-50-95% scan bands and (ii) three one-dimensional curves: P(none other ever) against prior P(N_ever<1), 'stuck' against prior P(p_exp<1e-5), 'unnoticed' against p_nonevident. Label the base row 'modeller's judgement'.
  • Expected effect: Bands from the project's 400-draw scan (5/50/95%). Site: none-ever 0.26/0.52/0.74, stuck 0.06/0.17/0.36, unnoticed+hidden 0.09/0.29/0.60. SSA: 0.10/0.30/0.56, 0.05/0.20/0.44, 0.16/0.44/0.77. My wider 240-draw scan: site 0.26/0.57/0.81, 0.10/0.28/0.57, 0.01/0.09/0.37.
  • Tested: Used integrate3_scan.json (my parallel rerun of the same 400 draws reproduces it with maximum difference 0.0). The three answers are within a factor 2 in 12% of draws (site) and 21% (SSA); within a factor 3 in 29% and 43%. Largest of the three: (a) in 76% of site draws, (c) in 61% of SSA draws. A 45-cell biology grid (w_fast 0.1-0.8, m0 weight 0.1-0.6, step_decades max 4-12) gives 'roughly equal' in 31% of cells (site) and 51% (SSA), with site none-ever from 0.17 to 0.80. Scripts scan_analyse.py, sens_bio.py.
• major / The regime atoms p_nonevident = 0.25 and p_here = (1e-3, 0.3), model3.py lines 66-67, 183-185 and 197; draft answer (c) and the 'unnoticed' class
  • Argument: The regime is drawn once per universe, so every sustained expander in that universe is low-footprint, and p_here (median 0.017) does not depend on how many independent expansion waves there were. In 'unnoticed' worlds the posterior median number of independent waves that filled or reached the Galaxy is 10^4.4 (site), 10^4.9 (SSA) and 10^7.6 (SIA); 90% have more than 10 and 67% more than 1,000. The class is therefore a universal-behaviour point mass that is 95-98% immune to the local observation by construction. That is round 1's objection to q_feasible, doom and stealth, relocated. Its posterior is the prior times 1.1 to 2.2: 0.028, 0.087, 0.190, 0.268, 0.350, 0.454 at p_nonevident = 0.03, 0.10, 0.25, 0.40, 0.60, 0.97. The project's own behavioural brief gives 'present but unnoticed' 0.08 (0.01-0.30) of the explanation given at least 10 civilisations; the model's implied share given N_ever >= 10 and the data is 0.41.
  • Proposed change: Split the atom in two: P(low footprint is forced on everyone by physics or economics), which is a legitimate universe-level prior, and a per-origin behaviour with between-world spread, for which evidentness compounds as 1-(1-p_here)^mu. Report 'unnoticed' as a function of the first, and state the number of independent waves that must all be non-evident. If one number is kept, calibrate the atoms to the behavioural brief's conditional shares.
  • Expected effect: Per-wave evidentness: unnoticed 0.19 to 0.04 (site), 0.29 to 0.05-0.06 (SSA), 0.61 to 0.10-0.16 (SIA); P(none other ever) 0.53 to 0.62 (site) and 0.33 to 0.43 (SSA). Brief-calibrated atoms (p_nonevident 0.035, p_stealth 0.055): unnoticed 0.03, hidden 0.055, none-ever 0.61 (site) and 0.41 (SSA).
  • Tested: sens_regime.py (p_nonevident sweep; p_here ranges up to (0.3, 0.99) give unnoticed 0.096; p_harvest_nonev 0.9 gives 0.082; p_here scaled with mu capped at 1e3 and 1e6). diag2.py for the wave counts. In my wider scan the 'per-wave evidentness' indicator is the second-largest driver of unnoticed (standardised logit coefficient -0.73, after log p_nonevident +1.13, R2 0.97), and with it unnoticed never exceeds 0.25 (site 95th percentile 0.13).
• major / w_fast = 0.4, model3.py line 46 ('prior P(lambda >= 1)'), combined with L_early at lines 166-168; draft claim 3
  • Argument: evidence/abiogenesis.md gives 'p_fast central 0.35, range 0.13-0.86', and its own table shows these are the posterior values after the early-life datum (0.33 at a floor of 1e-30; 0.13 to 0.82 across floors). The matching pre-data priors in that table are 0.09 (floor 1e-30) and 0.23 (floor 1e-10). The model uses 0.4 as a prior and then applies L_early, which raises P(fast) to 0.55 (site) and 0.71 (SSA). Round 1 flagged exactly this for w_easy = 0.5; version 3 changed the number to 0.4 and kept the double count. Claim 3's 'Earth's timeline alone pushes toward a populated Galaxy (0.44 to 0.34)' starts from a prior that already contains the timeline.
  • Proposed change: Either set w_fast to a pre-data value (0.09-0.23, with the slow-camp width tied to the same floor) and keep L_early, or keep 0.35-0.4 and switch L_early off for the abiogenesis rate. Document the provenance of every mixture weight and state that none was derived from Earth's timeline or the silence. The m_weights comment cites the timing study; confirm that they were not informed by it (the shift toward fewer steps suggests they were not).
  • Expected effect: P(none other ever), site: 0.53 becomes 0.74, 0.69, 0.64 at w_fast = 0.09, 0.15, 0.23, or 0.64 with timing off. SSA: 0.33 becomes 0.61, 0.53, 0.45, or 0.47 with timing off. At w_fast = 0.15 the three answers are 0.69/0.17/0.13 (site) and 0.53/0.23/0.21 (SSA). Combined with brief-calibrated regime atoms: none-ever 0.75 (site), 0.61 (SSA).
  • Tested: diag3.py, n = 300,000 per run. I found no structural double count among L_early, L_late and the SSA weight: the product L_early x L_late matches the exact joint order-statistic density within 3% for m = 1 to 3, and W[0] x L_late / D_all is the correct joint for (G host, arrival at x). The double count is in the prior's provenance.
• major / W_exp = (1, 30), model3.py line 59; soft spot 2 ('mean prior p_exp is 0.12'); draft answer (b) 'about 0.25 to 0.33'
  • Argument: The mean is the wrong summary. The prior has median p_exp of 0.007, P(p_exp > 0.1) = 0.28, P(p_exp > 1%) = 0.48, P(p_exp < 1e-5) = 0.28 and P(p_exp < 1e-10) = 0.13: an optimist head and a sceptic tail. 'Stuck' is the tail mass P(p_exp < about 1/N_ever) and so is set by W_max: 82% of the stuck posterior has p_exp < 1e-5 (median 1e-8.9). 'Same ignorance structure as the biological links' does not pin the prior down: the single-hard-step analogue W = (2, 8) and the slow-abiogenesis analogue W = (3, 100) differ fourfold in the result. The hierarchical form is a real improvement on round 1's hard floor, but the quoted 0.25-0.33 is the prior restated.
  • Proposed change: Present 'stuck' as a curve against the prior tail mass P(p_exp < 1e-5) (roughly 0.9 times the tail for tails up to 0.4), and print the implied prior quantiles of p_exp instead of the mean. Give named rows: W(1,10), W(1,30), W(1,60), and priors calibrated to P(p_exp > 1%) = 0.7 and 0.1.
  • Expected effect: Stuck (site / SSA / SIA): W(1,4) 0.04/0.05/0.00; W(1,7.8), P(>1%) = 0.70: 0.08/0.09/0.00; W(1,10) 0.10/0.12/0.01; W(1,30) 0.25/0.32/0.33; W(1,60) 0.33/0.43/0.53; W(4,40) 0.39/0.49/0.56; W(9,60), P(>1%) = 0.10: 0.50/0.63/0.76; W(2,8) 0.11/0.12; W(3,100) 0.44/0.57. P(none other ever) moves only 0.62 to 0.39 (site) over the same range; the mass is exchanged mainly with 'unnoticed' (0.34 to 0.04).
  • Tested: sens_W.py and diag4.py, n = 250,000-400,000 per run. Marginal likelihood rises from 0.26 (W 0.3-3) to 0.50 (W 15-60): the data prefer the sceptical expansion prior by a Bayes factor of at most 1.9.
• major / The SIA column of Table B, 'sia' weights at model3.py line 236, and every SIA number in claims 2 and 11
  • Argument: Not fixed from round 1. The SIA posterior median of log10 N_ever is 8.70, the prior's 99.63rd percentile (5% and 95% points sit at the 94th and 99.99th), so the answer is set by the upper edges of Nstars, eta_earth, f_site, f_persist and f_c. The class split is the ratio of prior atoms: taking the prior restricted to its top 0.5% of N_ever and weighting only by the silence terms gives en route 0.018, stuck 0.286, unnoticed 0.640, hidden 0.053, against the SIA 'posterior' of 0.022, 0.319, 0.601, 0.055. Importance sampling is inadequate outside the 3M base run.
  • Proposed change: Report SIA qualitatively (emptiness and rarity are eliminated) and give the split as a table against (p_nonevident, p_stealth, W_max, per-wave evidentness). If numbers are kept, sample the N_ever > 1e6 tail directly (stratified or conditional sampling) and delete SIA from the camps and the scan.
  • Expected effect: SIA unnoticed ranges 0.10 to 0.91 and stuck 0.00 to 0.84 across the single-knob runs: unnoticed 0.18/0.61/0.85 at p_nonevident 0.03/0.25/0.97; stuck 0.00/0.33/0.76 at W(1,4)/(1,30)/(9,60); hidden 0.04-0.06/0.20/0.55 at p_stealth 0.02/0.10/0.40; per-wave evidentness gives stuck 0.71-0.75 and unnoticed 0.10-0.16.
  • Tested: diag2.py at n = 1M: ESS 1,547 (0.15%); the top 100 samples carry 19% of the weight and 829 samples carry half. Camps: ESS 44 (Carter) and 56 (strict ignorance). Scan at n = 100,000: SIA ESS median 175, 5th percentile 36, minimum 12. 'Hidden under 0.05' is already false under SIA in the base run (0.054 at 3M, 0.058 at 1M).
• major / Draft claim 1: '27% of prior mass', 'discriminate among everything else by less than a factor of 2', and Table B's prior column
  • Argument: Four problems. (1) The prior mass of the contradicted class is 0.309 in integrate3_results.json, not 0.27; the draft's prior column sums to 0.96, and the en-route likelihood is 0.41, not 0.44. (2) 'Less than a factor of 2' is 1.96 in the site base run and is not general. (3) Most of that discrimination comes from Earth's timeline, not the silence: silence-only class-mean likelihoods span a factor 1.5 (1.00, 0.93, 0.67, 0.84, 0.87, 0.99), timeline-only 1.97. (4) The 31% removed is itself p_claim (0.73) times prior P(settled) (0.42), a product of priors, and ranges 0.12 to 0.73 across the scan.
  • Proposed change: Correct the table. State the exclusion with its range, give the discrimination among survivors by convention, and add the plain statement that the site posterior equals the prior with the contradicted class deleted to within 0.08. Say that 'not evident from here' survives because it is assumed immune.
  • Expected effect: Spread of class-mean likelihood among the six surviving classes: site base 1.96; scan median 1.72, 95th percentile 2.5, maximum 3.75, under 2 in 74% of draws. SSA base 4.8; scan median 4.1, 5-95% 2.2-7.3, maximum 10.2, under 2 in 1% of draws. SIA unbounded. With per-wave evidentness 'unnoticed' is penalised by a factor 4 to 9.
  • Tested: diag_base.py at n = 1.5M: site posterior against the renormalised prior differs by at most 0.077 (0.029 bits of KL on the six classes); SSA 0.154 (0.22 bits); SIA 0.45 (1.46 bits). Across the scan the same site distance has median 0.04 and 95th percentile 0.079. The excluded class has mean likelihood 0.001-0.011 in all 400 draws and posterior under 0.015 (site).
• moderate / p_stealth = 0.02 and p_nonevident = 0.25 against evidence/behavioral.md's scorecard; the last sentence of claim 2 ('Deliberate hiding stays under 0.05 unless its prior is raised by hand')
  • Argument: Round 1 asked that the behavioural brief's conditional shares (given at least 10 civilisations) be treated as a calibration target. They were not. The model's implied shares given N_ever >= 10 and the data are stuck 0.50, unnoticed 0.41, hidden 0.04 (site). The brief gives non-expansion families 0.45 in total, 'present but unnoticed' 0.08 (0.01-0.30), zoo 0.10 (0.02-0.30). So 0.02 is the bottom of the project's own range for hiding and 0.25 implies five times the brief's central weight for unnoticed. Every prior here is set by hand, and the hidden posterior is the prior times a constant.
  • Proposed change: Replace the sentence with the proportionality: hidden is about 0.9 x p_stealth (site), 1.3 x (SSA), 2 to 2.7 x (SIA). Either calibrate both atoms to the brief or explain the departure.
  • Expected effect: Hidden at p_stealth = 0.02 / 0.05 / 0.10 / 0.20: site 0.018 / 0.042 / 0.081 / 0.148; SSA 0.026 / 0.063 / 0.117 / 0.210; SIA 0.04-0.06 / 0.12 / 0.20 / 0.35. Hidden exceeds 0.05 once p_stealth passes about 0.06 (site), 0.04 (SSA), 0.02 (SIA). At the brief's central 0.10 it is 0.08 to 0.20.
  • Tested: sens_regime.py; conditional shares computed from a 600,000-sample base run. In the scan log(hidden) correlates 0.91 with log(p_stealth) and hidden stays under 0.05 in 83% (site), 71% (SSA) and 50% (SIA) of draws with p_stealth drawn from 0.003 to 0.1.
• moderate / The 'empty universe' row (site 0.20, SSA 0.12) and the missing existence-conditioned column
  • Argument: Round 1 objection 8 was computed (key conditional_on_someone_in_observable_universe in the results file) but left out of the draft. Within the empty class the posterior median expected number of civilisations in the observable universe is 10^-12.4, a quarter of the mass is below 10^-27 and 5% below 10^-56. The 0.20 therefore assumes a universe at least about 10^40 times the observable volume, or treats our own existence as carrying no information. Its posterior-to-prior ratio is 1.04: the row is the prior tail below about 1e-11 per galaxy and nothing else.
  • Proposed change: Add a fourth column, 'at least one civilisation arose within volume V', for V from the observable universe upward, and state the cosmological assumption behind the site and SSA rows.
  • Expected effect: Empty universe (site) is 0.006, 0.045, 0.065, 0.098, 0.138, 0.175, 0.196 when conditioning on someone arising within 10^0, 10^3, 10^5, 10^10, 10^20, 10^40, 10^80 observable volumes; P(none other ever) runs 0.42 to 0.53. SSA: empty 0.003 to 0.119, none-ever 0.245 to 0.333.
  • Tested: Reweighted an 800,000-sample base run by 1 - exp(-N_ever x G x 10^k). The k = 0 row matches the stored value (0.0056).
• moderate / Draft claim 10 (what would change the answer); model3.py lines 214-215 (artefact_null)
  • Argument: (1) The artefact-survey result is arithmetic on an assumed detection probability of 0.5: unnoticed goes from u to u(1-d)/(1-ud), which is 0.104 for u = 0.188 and d = 0.5; the run gives 0.105. No property of a real survey enters. (2) The second-genesis and HWO numbers are specific to the base biology prior. 'Empty universe to 0.03' holds only when emptiness comes from abiogenesis; where hard steps supply it, it survives.
  • Proposed change: Give the artefact row as a function of d and source d, or drop the number. Give the biosignature rows as a range over the biology camps and say that the direction, not the size, is robust.
  • Expected effect: Artefact null: unnoticed 0.173, 0.140, 0.105, 0.046, 0.014 at d = 0.1, 0.3, 0.5, 0.8, 0.95. Second genesis, P(none other ever) before and after: base 0.53 to 0.32; pessimist (w_fast 0.1, steps to 12 decades) 0.83 to 0.61; optimist 0.15 to 0.05; Carter camp 0.91 to 0.82. Empty universe after a second genesis: 0.03 (base), 0.18 (pessimist), 0.31 (Carter). HWO detection is similar: 0.27, 0.60, 0.03, 0.80.
  • Tested: diag2.py (closed form against run for five values of d) and diag3.py (four camps by four observations, n = 300,000).
• moderate / Draft claim 3 (ablation): 'Earth's timeline alone pushes toward a populated Galaxy'; the single ordering of the ablation; the SSA rows of the ablation table
  • Argument: (1) The timeline's net effect is the sum of two opposed terms: early life lowers P(none other ever) (0.44 to 0.30 alone) and late arrival raises it (0.44 to 0.48 alone). The net sign depends on the priors, and in the base it rests on the double-counted w_fast. (2) The ablation is reported in one order, and the survey term's share depends on that order. (3) The surveys look weak partly because harvesting is a universe-level atom (harv is drawn once per universe at line 186), so half of claiming universes are immune to every galaxy survey whatever N_eff is. (4) The SSA rows of the ablation with the local term off are numerically meaningless.
  • Proposed change: Report the early-life and lateness terms separately, give Shapley or all-orders attributions, state the harvest-atom assumption next to 'surveys add little', and delete the SSA ablation rows that have use_local = False.
  • Expected effect: Shapley contributions to P(none other ever) in the base (prior 0.441 to 0.531): early life -0.142, lateness +0.042, nobody here +0.154, surveys +0.035. Surveys alone move it +0.056 and remove only a quarter of the contradicted class (0.308 to 0.233). Across my 240-draw scan the timeline lowers P(none other ever) in 63% of draws (median -0.014, 5-95% -0.087 to +0.051); 'nobody here' raises it in 100% (median +0.20); surveys after that add a median +0.028 (+0.006 to +0.076).
  • Tested: diag2.py (13 term subsets and all 24 orderings from one 1M-sample run) and scan_critic_analyse.py. The SSA ablation rows in integrate3_results.json show evidence 2.8e301 and 'excluded 1.000'.
• moderate / Draft claim 11, 'Anthropics is the largest lever'
  • Argument: True only if SIA is counted. Between the two conventions the draft treats as live (site and SSA) the mean absolute difference in P(none other ever) across the scan is 0.21 (maximum 0.40), while the prior draw moves it from 0.11 to 0.87 within site and 0.03 to 0.79 within SSA (standard deviation 0.15 and 0.14). For 'stuck' and 'hidden' the convention matters less than the prior (mean gap 0.04 and 0.02 against a scan standard deviation of 0.09-0.12 and 0.04-0.05). The convention also flips which camp the data favour: the marginal likelihood is 0.72 (Carter) against 0.27 (biology optimist) under site and 0.044 against 0.109 under SSA, a Bayes factor near 2.5 to 2.7 in opposite directions.
  • Proposed change: Reword to rank the levers: the biological prior and the two late-stage atoms first, the site-SSA choice second (about 0.2 on P(none other ever)), SIA as a separate case that removes emptiness and leaves a split fixed by priors. Add the sentence that no camp is favoured over another by more than about 3 under either convention, and that the direction reverses between them.
  • Expected effect: No model number changes. The claim changes from a ranking that puts anthropics first to one in which it is second to the priors unless SIA is adopted.
  • Tested: scan_analyse.py on integrate3_scan.json; evidence values read from the camps table in integrate3_results.json.
• minor / Draft claim 4 (false precision) and the design of run3.scan (lines 58-79)
  • Argument: (1) '2.5 to 2.9 billion' is the range over expansion speed only. The headstart report gives a factor 6 either way from star count and eta_Earth, and a limit of 1.3e-10 to 5.6e-9. The model's own prior puts habitable-zone rocky planets around G-class stars older than the Sun at 0.9 to 20 billion (5-95%, median 4.2 billion). The headstart referee widens '72.5% (69 to 76%)' to about 0.64-0.76 centred near 0.70. (2) The scan is not centred on the base (prior P(N_ever<1) median 0.33 against 0.44 in the base; W_max median 16 against 30) and holds fixed W_min, p_here, p_harvest_nonev, the f_c floor and the universality assumption.
  • Proposed change: Quote 'about 3 billion (0.5 to 20 billion)' and 'below about 1e-9 (1e-10 to 6e-9)', and 'about 70% (64 to 76%)'. Re-run the scan centred on the base with the extra knobs, and report the bands for the two structures separately.
  • Expected effect: Wider scan, 5/50/95%: site stuck 0.10/0.28/0.57 against 0.06/0.17/0.36 in the project scan; site unnoticed 0.004/0.04/0.30 against 0.07/0.26/0.56; SSA none-ever 0.11/0.35/0.67 against 0.10/0.30/0.56.
  • Tested: Prior quantiles computed from the BASE ranges (1M draws). scan_critic.py, 240 draws at n = 100,000, minimum ESS 7,359 (site) and 4,955 (SSA).

Claims that survive

• The contradicted class is excluded under every prior tried: mean likelihood 0.004 in the base and 0.001-0.011 across all 400 scan draws; posterior under 0.015 (site) in every draw.
• 'Nobody evidently here' raises P(none other ever in the Galaxy) in 100% of 240 wider-scan draws (median +0.20), and the net effect of all observations is positive in 100% (median +0.22, 5-95% +0.07 to +0.38).
• Galaxy surveys add little once the local argument is in (+0.006 to +0.076) and radio SETI adds nothing (base 0.530 to 0.531); p_harvest_claim is not a lever (0.1 to 0.9 moves P(none other ever) from 0.525 to 0.540).
• Hard class boundaries create no artefact. Poisson versions: P(none other ever) 0.532 to 0.519, P(settled) 0.209 to 0.222, stuck 0.250 to 0.273 (site). Moving the N_ever threshold over 0.1-10 gives 0.48-0.585; moving the mu_arr threshold over 0.1-10 gives settled 0.256-0.184. Only 'en route' is threshold-sensitive (0.068, 0.041, 0.017 at x >= 0.001, 0.01, 0.1), and it is small.
• The label 'stuck' is fair: 82% of its posterior has p_exp < 1e-5 and only 5% has p_exp > 1%.
• Importance sampling is adequate for site and SSA: ESS 37% and 19% at n = 1M, top-1,000 samples carry 0.7% and 1.5% of the weight, scan minimum ESS 14,900 and 7,700 at n = 100,000. This does not extend to SIA.
• The SSA denominator integral is numerically sound: refining the time grid 20-fold changes D_all by under 0.4% and no class by more than 1e-4. The 'excluded 0.016' under SSA is real truncation physics, concentrated at mu_arr of 1 to 5.
• No structural double counting among L_early, L_late and the SSA weight: the product matches the exact joint within 3%. The only double count found is in the provenance of w_fast.
• The scan is reproducible: an independent parallel rerun of the 400 draws matches integrate3_scan.json exactly.
• Claim 7's structure holds in the model: the SSA column responds to the planet-lifetime atom (P(none other ever) 0.25, 0.33, 0.38 at p_cap_short 0.2, 0.5, 0.8) and not to the M-dwarf atom (changes of 0.01 or less across p_M_class 0.05-0.6); the site column does not move.
• The direction of the second-origin and biosignature updates is robust across four biology camps (P(none other ever) always falls, by 0.09 to 0.22); only the size is prior-specific.
• Letting the expansion prior depend on the regime (non-evident expansion easier or harder, claiming expansion harder) moves P(none other ever) by at most 0.06 and exchanges about 0.06 between stuck and unnoticed; it creates no new answer.
• The hierarchical expansion prior is a genuine improvement on round 1's hard floors, and the data mildly prefer the sceptical end (marginal likelihood 0.26 to 0.50, a Bayes factor under 2).

Verdict on each headline claim

• Claim 1: reword
  • Reason: The exclusion is robust, but the prior mass is 0.31 not 0.27 (the table's prior column sums to 0.96), the factor of 2 holds only under site (median 1.7, maximum 3.8 across the scan; SSA 4.8 in the base and 2 to 10 across the scan), and most of the discrimination among survivors comes from the timeline, not the silence.
  • Corrected wording: The observations rule out one thing: expansion of the system-claiming kind that would be evident here (31% of the base prior; likelihood 0.004, and 0.001 to 0.011 across 400 prior draws). Among everything else the likelihood varies by about a factor of 2 if Earth is treated as a given site (up to 4 across priors), by 2 to 10 under SSA, and without limit under SIA; the silence terms alone separate the surviving explanations by a factor of 1.5. On the surviving explanations the base posterior is the prior with the contradicted class deleted, to within 0.08. 'Settled but not evident' survives because the model assumes it is nearly immune to the observation, not because it was tested. Any confident single answer is a statement of priors.
• Claim 2: drop
  • Reason: 'Roughly equal' holds in 12% (site) and 21% (SSA) of the project's own scan draws and is false in the site base (0.53/0.25/0.21). Each answer is one prior restated. 'Hiding stays under 0.05' is false under SIA in the base (0.054) and at the behavioural brief's central zoo weight (0.08 to 0.20). Replace it with the wording below.
  • Corrected wording: Three answers survive, and their relative sizes are set by three priors, not by the data. (a) Nobody else has ever arisen in the Galaxy: 0.53 in the modeller's base; 0.26 to 0.74 (site) and 0.10 to 0.56 (SSA) across a joint prior scan; it follows the prior on biological difficulty. (b) Others arose but sustained expansion almost never happens: 0.25 in the base; 0.02 to 0.54 as the prior on expansion is widened from 3 to 60 decades. (c) Settled but not evident from here: 0.19 in the base; about 0.9 times the prior probability that all expansion everywhere is low-footprint, falling to 0.04 if each expansion wave has its own chance of being noticed. Deliberate hiding is about 0.9 (site) to 2.7 (SIA) times whatever prior it is given.
• Claim 3: reword
  • Reason: The timeline's net sign is not robust (negative in 63% of draws; 5-95% -0.09 to +0.05) and in the base it rests on a prior that already contains the early-life datum. The ablation is order-dependent. The other three statements hold in every draw.
  • Corrected wording: Ablation (base priors, site). Early life lowers P(none other ever) (0.44 to 0.30) and late arrival raises it (0.34 with both); the net sign of the timeline depends on the biological prior. The absence of evident settlement here raises it under every prior tried (to 0.51 in the base; median +0.20 across priors). The galaxy surveys add 0.01 to 0.08 after that and 0.06 on their own, in part because the model treats star harvesting as all-or-nothing per universe. Radio SETI adds nothing. Averaged over all orderings the contributions are -0.14 (early life), +0.04 (lateness), +0.15 (nobody here), +0.04 (surveys).
• Claim 4: reword
  • Reason: False precision: 2.5 to 2.9 billion is the range over speed only; the study's own range for the limit is 1.3e-10 to 5.6e-9 and the model's prior on the planet count spans 0.9 to 20 billion. The referees widen 72.5% (69-76%) to about 0.64-0.76 and 11.1 +- 0.5 to 11 +- 1.
  • Corrected wording: Time and distance do not explain the silence (original data). Gaia: stars pass within 1 pc of the Sun about 11 +- 1 times per Myr; about 21,000 (12,000 to 57,000) systems have done so since the Great Oxidation. Stellar ages: about 70% (64 to 76%) of Sun-like stars at the Sun's radius are older than the Sun, by 3 to 4 Gyr on average, so roughly 3 billion (0.5 to 20 billion) habitable-zone rocky planets had time to settle the Galaxy at any speed from 1e-4 c to 0.1 c. If settlement is of the evident kind, the per-planet probability of producing a Galaxy-settling civilisation is below about 1e-9 (1e-10 to 6e-9).
• Claim 5: reword
  • Reason: 'First limits that do not depend on the waste-heat temperature' is contradicted by the tf_kiii report, which presents itself as a 2.5-times tightening of Zackrisson et al. (2015), the same temperature-independent Tully-Fisher method. The referee's TF limits are about 2.3% and 1.4%. In the model the surveys move P(none other ever) by 0.016.
  • Corrected wording: No star-harvested galaxies at any waste-heat temperature (original data). Using optical scaling relations: among about 1e5 early-type galaxies (fundamental plane, new) fewer than about 0.3% have 75% of their starlight intercepted; among about 9,400 discs (Tully-Fisher, extending Zackrisson et al. 2015 about 2.5-fold) fewer than about 2%. Both are systematics-limited by unexplained outliers. In the integrated model these limits change the answer by under 0.02 once the absence of settlement here is counted.
• Claim 6: keep
  • Reason: Outside my lens for the numbers. The model agrees: the SETI term's class-mean likelihood is 0.992 to 1.000 and it moves no class by more than 0.001. Quote the referee's range (a few thousand to about a million) alongside 4e5.
• Claim 7: reword
  • Reason: The structure holds (lifetime atom matters, M-dwarf atom does not), but 0.33 to 0.14 is an SSA-only, base-prior number, the opposite branch is missing, and the site column does not move.
  • Corrected wording: Keep the first three sentences. Replace the last with: In the model this matters only under SSA: learning that planets stay habitable beyond 15 Gyr moves P(none other ever in the Galaxy) from 0.33 to 0.14, learning the opposite moves it to 0.41, and M-dwarf verdicts move it by 0.01 or less; under the site convention none of these changes anything (0.52 to 0.54).
• Claim 8: reword
  • Reason: 'About one bit' needs an object. In the model the timeline is worth a Bayes factor of 1.9 on fast versus slow abiogenesis (0.9 bit on that one question), 0.03 bits on the class partition, and its net effect on P(none other ever) has no robust sign. The base prior already contains the early-life datum.
  • Corrected wording: Earth's timeline is worth about one bit on one question (original reanalysis): it favours fast over slow abiogenesis by about 2 to 1 in the model, with a ceiling of 13 to 1 (Kipping) that falls to 9, 7, 6 with 2, 3, 4 hard steps and ranges 4.6 to 942 across 38 published LUCA datings. It says almost nothing about which explanation of the silence is right. Snyder-Beattie et al.'s tables reproduce only with thresholds e^-k where 10^-k is printed, which halves their claimed weight on extreme rarity. Pacing versus gated hard steps: Bayes factor 0.6 to 3.3.
• Claim 9: keep
  • Reason: Outside my lens. A clustered p of 0.12 is a null result and is worded as one. The referee disputes the encephalisation record behind the dolphin figure and the counting of species as independent lineages, so check 0.73 and '3 to 4 lineages' against the revised report before publishing.
• Claim 10: reword
  • Reason: The artefact-survey number is arithmetic on an assumed 50% detection probability (closed form 0.104, run 0.105). The biosignature numbers are base-specific: empty universe stays at 0.18 to 0.31 in the hard-step camps.
  • Corrected wording: What would change the answer. A second origin of life, or one biosignature among 25 HWO planets, lowers P(none other ever in the Galaxy) under every biological prior tried, by 0.1 to 0.2 (0.53 to about 0.3 in the base); it removes the empty-universe answer only if emptiness comes from abiogenesis (0.20 to 0.03 in the base, but 0.52 to 0.31 with many hard steps). A Solar System artefact survey lowers 'settled but not evident' in proportion to its assumed detection probability d (0.19 to 0.10 at d = 0.5, to 0.05 at d = 0.8). More radio SETI: no change. M-dwarf verdicts: no change.
• Claim 11: reword
  • Reason: Between site and SSA the convention moves P(none other ever) by 0.21 on average, less than the prior range (0.11 to 0.87 within site). Only SIA dominates, and its split is the ratio of prior atoms with an ESS of 0.15%. The convention also reverses which camp the data favour.
  • Corrected wording: Anthropic convention is the second-largest lever and cannot be argued away. Moving from 'Earth is a given site' to SSA lowers P(none other ever in the Galaxy) by about 0.2; the choice of biological prior moves it by more (0.1 to 0.9). The convention also reverses which camp the data favour, by a factor under 3 either way. Under SIA emptiness is ruled out and the answer becomes 'they exist; expansion fails or is not evident', with the split between those fixed entirely by priors.

Code errors

• report/FINDINGS_DRAFT.md / Section B table, prior column and the 'Mean likelihood by class' line; claim 1
  • Problem: The prior column sums to 0.96. The contradicted class has prior 0.309 in integrate3_results.json (my 1.5M-sample run: 0.308), not 0.27, and the en-route mean likelihood is 0.41, not 0.44.
  • Fix: Regenerate the table from integrate3_results.json and change '27% of prior mass' to 31%.
  • Impact: Claim 1 understates the mass removed by 0.04; no posterior changes.
• models/model3.py / line 46, w_fast = 0.4, comment 'prior P(lambda >= 1)'
  • Problem: The value comes from the abiogenesis brief's p_fast 'central 0.35 (0.13-0.86)', which is that brief's posterior after the early-life datum. L_early (lines 166-168) then applies the datum again.
  • Fix: Use the brief's pre-data values (0.09 at a floor of 1e-30, 0.23 at 1e-10) with L_early on, or keep 0.35-0.4 and drop L_early for the abiogenesis rate.
  • Impact: P(none other ever): site 0.53 becomes 0.64-0.74, SSA 0.33 becomes 0.45-0.61.
• models/model3.py / line 235, share_G = W[0] / np.maximum(D_all, 1e-300), used with use_local = False
  • Problem: The SSA numerator's survival factor exists only inside L_local. With the local term switched off the truncated denominator underflows and share_G reaches about 1e301. The SSA rows of the ablation ('Prior', 'timeline only', 'surveys, no local argument') in integrate3_results.json show evidence 1.4e301 to 2.8e301 and 'excluded 1.000'.
  • Fix: Put the pre-emption survival exp(-mu_arr) into the SSA weight explicitly for the claiming regime, independent of use_local, or do not compute SSA when use_local is False. Delete those rows from the results file.
  • Impact: Draft claim 3 quotes only the site ablation, so no headline number is affected; the stored SSA ablation rows are unusable.
• models/model3.py / lines 183-186 and 197 (regime, p_here and harv drawn once per universe)
  • Problem: p_here does not depend on mu_arr and the regime and harvest draws apply to every expander and every galaxy in the universe. This is a modelling choice with a large effect, and it is not stated in the draft.
  • Fix: Add an option for per-origin behaviour: p_here_eff = 1 - (1 - p_here)^mu_arr for the non-evident regime (my model3x.py key p_here_scale_with_mu), and a per-galaxy harvest probability. Report both structures.
  • Impact: Unnoticed 0.19 to 0.04 (site), 0.29 to 0.05-0.06 (SSA), 0.61 to 0.10-0.16 (SIA).
• models/model3.py / lines 214-215, extra_obs == 'artefact_null'
  • Problem: The detection probability 0.5 is hard-coded and unsourced, so the reported effect is u(1-d)/(1-ud) by construction.
  • Fix: Make d a parameter with a stated source and report the row as a function of d.
  • Impact: Unnoticed 0.173 to 0.014 for d from 0.1 to 0.95.
• models/run3.py / scan(), lines 58-79, and the SIA rows of the camps table
  • Problem: At n = 100,000 the SIA effective sample size has median 175 and minimum 12; in the camps it is 44 (Carter) and 56 (strict ignorance). The scan also fixes the lower bound of W_exp at 1, p_here, p_harvest_nonev and the f_c floor, and is centred on a more optimistic prior than BASE (prior P(N_ever<1) median 0.33 against 0.44; W_max median 16 against 30).
  • Fix: Drop SIA from the scan and camps or sample the high-N tail directly. Centre the scan on BASE and add the missing knobs (my scan_critic.py does this for site and SSA).
  • Impact: SIA bands from the scan are noise. Site and SSA bands shift: for example site stuck 5-95% 0.06-0.36 becomes 0.10-0.57.
• models/model3.py / line 164, L_late = 1.0 for gated worlds and for m_hard <= 1
  • Problem: L_late is a density (m x^(m-1), up to 1.9) compared across models with a convention of 1 for the gated model, whose own table (cosmo_tables.py T_DELAY0 = 4.5 Gyr) predicts arrival at exactly Earth's value. P(gated) falling from 0.30 to 0.23 under site is partly this convention.
  • Fix: State the convention, or give the gated model an explicit arrival-time density (for example gate times uniform or log-uniform over the window, as in the timing study, which gives Bayes factors of 1.0-2.4 or 2.9-17 for hard steps).
  • Impact: Small on classes (lateness contributes +0.04 to P(none other ever)); it affects the reported posterior on the number of hard steps (mean 1.15 to 1.71 under site).

Were the round 1 fixes made?

Fixed: (3) late-stage hard floors are replaced by a hierarchical per-universe p_exp, which also represents convergent restraint; (4) the local term now has p_here and the surveys use a limit curve built from the two studies; (9) classes are defined by mechanism and their boundaries are harmless (I tested Poisson versions); (10) the HWO likelihood includes f_site and false positives; (11) the 1/m! factor is in; the base is labelled 'modeller's judgement' in the camps table; a joint scan exists. I also checked the new SSA denominator integral and could not break it.

Moved, not fixed: (7) the point masses. q_feasible, doom and stealth became p_nonevident = 0.25 and p_stealth = 0.02, both universe-wide; 'unnoticed' is the prior times 1.1 to 2.2 exactly as 'infeasible' was the prior times 2.3, and the behavioural brief's conditional shares are still not used as a calibration target (model 0.41 for unnoticed against the brief's 0.08). (2) The single posterior vector became three vectors and a headline (claim 2) that restates the base vector; the scan bands and a must-believe table are not in the draft.

Not fixed: (1) the SIA split is still the ratio of prior atoms, the posterior still sits at the prior's 99.6th percentile of N, and the ESS is 0.15%; the draft labels the split as prior-set in claim 11 but prints two-figure SIA numbers in Table B and claim 2. (5) w_fast = 0.4 is still the abiogenesis brief's posterior used as a prior with L_early applied on top. (8) The existence-conditioned variant is computed in the results file but absent from the draft, so the empty-universe 0.20 still carries an unstated assumption of a universe at least 10^40 times the observable volume. (6) The scan exists but is off-centre, omits several knobs and its SIA rows are noise.

Over-corrected or newly created: the local term now carries the whole inference and the surveys almost none, partly because harvesting became an all-or-nothing atom per universe; the expansion prior's upper bound of 30 decades now decides the 'stuck' class in the way the 1e-4 floor used to decide 'restraint', in the opposite direction (stuck 0.25 to 0.33 where restraint was 0.01); the new artefact-survey row is arithmetic on a hard-coded 0.5.

Missing explanations

• Mixed behaviour within a universe: a fraction of expanders low-footprint and the rest claiming, per origin rather than per universe. Version 3 allows only all-or-nothing, which is what protects 'unnoticed'.
• A catch-all for model misspecification. Any unlisted hypothesis that is immune to the likelihood receives about 1/evidence = 3 times its prior (site), so a seven-way partition that sums to 1 overstates every listed class.
• Existence-conditioned inference in a stated volume, as a fourth column beside site, SSA and SIA.
• Pre-emption as a selection effect under the site convention: if claiming expansion would have prevented our existence, 'nobody evidently here' is not ordinary evidence. The model handles this only inside the SSA denominator and only for the claiming regime.
• Heterogeneity of p_exp and of evidentness across galaxies and over time (metallicity, epoch, host type); the survey term applies the Milky Way's mu_loc and one harvest draw to every surveyed galaxy.
• Synchronised late emergence: the head-start fraction phi is fixed at 0.63-0.69 from the stellar census and is not replaced by the cosmo-table birth dates for worlds with several hard steps, so a late common onset is excluded by construction.
• Economic rather than physical limits on expansion, which belong with per-origin choice, not with a universal atom.

Critic’s own estimate of each answer

• Empty universe: 0.17
• Rare: 0.33
• En route: 0.04
• Stuck: 0.24
• Unnoticed: 0.07
• Hidden: 0.03
• Other: 0.12
• What “other” covers: Mechanisms outside the partition and model misspecification: mixed-behaviour universes, pre-emption selection effects, synchronised late emergence, simulation-type scenarios. These are judgement numbers. They reflect the pre-data abiogenesis prior (which raises the first two classes), a half weight on per-wave evidentness (which lowers unnoticed), and an expansion prior between W(1,10) and W(3,100). The data constrain them only by excluding evident settlement; I would not defend any entry to better than a factor of 2.

Presentation advice

Lead with four exhibits in place of Table B's point values and claim 2.

Exhibit 1, the data's whole contribution. Likelihood by class (site): evident settlement 0.004; all others 0.35 to 0.69. One sentence: on the surviving classes the posterior is the prior with the contradicted class deleted, to within 0.08 (0.03 bits); under SSA the spread is a factor 5 and the shift up to 0.15.

Exhibit 2, three curves, each answer against the prior that controls it. P(none other ever) against prior P(N_ever<1). 'Stuck' against prior P(p_exp<1e-5): 0.04, 0.10, 0.25, 0.33, 0.50 at tail mass 0, 0.08, 0.28, 0.38, 0.75. 'Unnoticed' against p_nonevident: 0.03, 0.09, 0.19, 0.27, 0.35 at 0.03, 0.10, 0.25, 0.40, 0.60, with a second line for per-wave evidentness (0.04 at 0.25).

Exhibit 3, scan bands (5/50/95%, project's 400 draws at n = 100,000). Site: empty universe 0.01/0.11/0.30; rare 0.18/0.36/0.49; en route 0.02/0.04/0.07; stuck 0.06/0.17/0.36; unnoticed 0.07/0.26/0.56; hidden 0.003/0.017/0.10; excluded 0.002/0.004/0.009; none other ever 0.26/0.52/0.74. SSA: empty 0.004/0.05/0.20; rare 0.07/0.21/0.36; en route 0.01/0.04/0.08; stuck 0.05/0.20/0.44; unnoticed 0.12/0.40/0.72; hidden 0.005/0.03/0.14; excluded 0.005/0.016/0.04; none other ever 0.10/0.30/0.56. SIA (ESS median 175, do not publish as numbers): stuck 0.00/0.09/0.52; unnoticed 0.34/0.80/0.98; hidden 0.006/0.05/0.32. My wider scan (240 draws; adds W_min, p_here, p_harvest_nonev, f_c floor, per-wave evidentness). Site: none other ever 0.26/0.57/0.81; stuck 0.10/0.28/0.57; unnoticed 0.004/0.04/0.30; hidden 0.001/0.013/0.23; empty 0.01/0.12/0.34. SSA: 0.11/0.35/0.67; 0.12/0.37/0.69; 0.006/0.06/0.42; 0.002/0.03/0.32; 0.004/0.06/0.25.

Exhibit 4, what you must believe (site unless stated).

• Nobody else ever in the Galaxy above 50%: prior P(N_ever<1) of at least about 0.30 (0.23 if low-footprint settlement is dismissed, 0.44 if p_nonevident is 0.6); under SSA at least about 0.5. In parameter terms: P(fast abiogenesis) of 0.4 or less with hard steps up to 8 decades, or 0.25 or less with any step prior tried; under SSA about 0.15 or less. With the early-life double count removed it is 0.64-0.74 (site) and 0.45-0.61 (SSA).
• Empty observable universe above one third: a slow-abiogenesis prior reaching about 70 decades or more, steps up to about 10 decades, P(fast) of about 0.2 or less, and a universe at least 10^40 times the observable volume. Reached in 3% of scan draws (site) and 0.3% (SSA).
• Civilisations arise but sustained expansion almost never happens, above one third: prior P(p_exp<1e-5) of about 0.4 or more (W up to 60 decades from 1, or a lower bound of 3-4 decades); about 0.3 under SSA. W(1,10) gives 0.10.
• Settled but not evident, above one third: P(all expansion everywhere is low-footprint) of about 0.55 or more (0.32 under SSA), with the chance of noticing a settled Galaxy at or below 0.3 however many waves settled it (median 25,000). With per-wave evidentness it never passes 0.25.
• Deliberately hidden above 10%: prior of about 0.13 or more (site), 0.085 (SSA), 0.04 (SIA).
• Evident settlement: no prior tried rescues it.
• Three roughly equal answers: the base point under SSA; 12-21% of scan draws.

Other points. Never print SIA class shares to two figures. Add the existence-in-volume column. Print the implied quantiles of p_exp (5% 2e-16, 25% 2e-6, median 0.007, 75% 0.13, 95% 0.67) instead of 'mean 0.12'. State next to every ablation number that it is for one ordering and one prior. Say in the text that no biological camp is favoured over another by more than about 3 and that the direction reverses between site and SSA.

Critic: the code audit

Overall verdict

The code is sound: model3.py does what it says to within about 0.015 on every site and SSA class, and no bug I found changes a headline number by more than that. A rerun of the 3M base matches the saved JSON exactly, and my independent re-implementation agrees to 0.0006 (site) and 0.0008 (SSA) on all seven classes. The weak points are in what is printed and claimed, not in the engine. The single most important change is to replace "three live answers, roughly equal" and the two-decimal point estimates with the ranges from the project's own 400-draw scan: the three answers are within a factor of 2 of each other in only 54 of 400 draws (site) and 82 of 400 (SSA), and P(none other ever in the Galaxy) spans 0.26 to 0.74. Second, the SIA column is not converged: a 50-times more efficient sampler gives 0.03 / 0.32 / 0.60 / 0.06 where the draft prints 0.03 / 0.33 / 0.59 / 0.05, the stored Monte Carlo s.d. is three times too small, and camp SIA rows are wrong by up to 0.05. Third, Table B and claim 1 carry transcription errors (prior mass of the excluded class is 0.31, not 0.27). The one modelling inconsistency with a visible effect is the early-life term in the no-hard-steps branch, worth +0.02 (site) to +0.04 (SSA) on P(none). All scripts are in the critic’s working folder (mine.py is the re-implementation; exactchain.py is the exact stage calculation).

Objections

• major / Draft claim 2 ('Three live answers, roughly equal'; 'about 0.5'; 'hiding stays under 0.05') against integrate3_scan.json
  • Argument: The project's own joint prior scan contradicts the wording. Define (a) = empty + rare, (b) = stuck, (c) = unnoticed + hidden. In the base case the site column is 0.50 / 0.25 / 0.21, a spread of 2.4, so it is not 'roughly equal' even there; only the SSA column is (0.31 / 0.32 / 0.31). Across the 400 scan draws the three are within a factor of 2 in 54 draws (site) and 82 (SSA). Which answer leads flips with the convention: (a) leads in 285 of 400 site draws, (c) leads in 266 of 400 SSA draws and 375 of 400 SIA draws. 'Hiding stays under 0.05' is false in the SIA column (0.054 saved, 0.056 converged) and its scan 95th percentile is 0.10 (site) and 0.14 (SSA).
  • Proposed change: Replace with: 'Three answers stay live. Which one leads depends on the anthropic convention and on four priors (number and width of hard steps, P(fast abiogenesis), width of the expansion prior, P(low-footprint settlement)).' Quote scan ranges, not point values: site P(none other ever) 0.26 to 0.74 (5-95%), median 0.52; SSA 0.10 to 0.56, median 0.30. State hiding as 'about twice to three times its prior'.
  • Expected effect: No change to model numbers. The headline changes from a three-way tie to a prior-dependent ordering. Site scan medians: rare 0.36, unnoticed 0.26, stuck 0.17, empty 0.11.
  • Tested: Read integrate3_scan.json (400 rows, site ESS 22k-48k). A regression of site P(none) on the scanned parameters has R2 = 0.94; standardised effects are mean number of steps +0.10, w_fast -0.07, p_nonevident -0.07, step width +0.06. For 'stuck' the expansion-prior width dominates (+0.07 site, +0.10 SSA).
• major / SIA column of Table B, claims 2 and 11; mc_sd['sia']; SIA rows of camps and scan (model3.py line 236, run3.py lines 86, 124, 135-136, scan n=100,000)
  • Argument: Plain prior sampling leaves the SIA weight on a tiny corner of the prior. ESS is 4,702 of 3M in the base run, about 1,400 per 1M and about 640 per 400k. The stored mc_sd (0.006-0.007, six seeds at 400k) is too small by a factor of 3: sixteen fresh seeds at 400k give s.d. 0.020 (stuck), 0.023 (unnoticed), 0.011 (hidden). The JSON disagrees with itself: base 3M has stuck 0.329 and the 1M 'Base' camp has 0.312. Camp SIA rows with ESS 44, 56 and 384 are noise. The scan's SIA column has median ESS 175, and 27% of draws are below 100.
  • Proposed change: Importance-sample for SIA. Draw eta_earth, f_site, f_persist, f_c (and f_gate when m = 0) uniformly in the variable instead of log-uniformly, and multiply the weight by (hi - lo) / (x ln(hi/lo)); this cancels the factor x in W[0] exactly. Draw the fast camp with probability 0.9 and reweight by 0.4/0.9 or 0.6/0.1. For many-step camps also tilt the step probabilities. Recompute mc_sd from at least 16 seeds. Drop the SIA column from the scan or rerun it with the tilted sampler.
  • Expected effect: Base SIA becomes en_route 0.027, stuck 0.316, unnoticed 0.599, hidden 0.056, excluded 0.003, each +/-0.001. Printed to two decimals: 0.03 / 0.32 / 0.60 / 0.06 instead of 0.03 / 0.33 / 0.59 / 0.05. Camp rows move by up to 0.05: Carter hidden 0.020 -> 0.053 and stuck 0.394 -> 0.34; Strict ignorance hidden 0.097 -> 0.058; Rare Earth stuck 0.286 -> 0.316.
  • Tested: sia_tilt.py: 6 x 500k tilted runs, ESS 35,000 each against about 700 plain, per-run s.d. 0.001-0.003; the tilted importance weights average 0.98 and return the plain prior class frequencies within noise. sia_seeds.py: 8 plain seeds at 1M give mean 0.027 / 0.321 / 0.593 / 0.057 with s.d. 0.003 / 0.006 / 0.010 / 0.007. sia_plain400.py: 16 seeds at 400k give mean stuck 0.3165, range 0.281-0.356. sia_camps.py for five camps.
• moderate / FINDINGS_DRAFT.md Table B, the 'mean likelihood by class' line, and claim 1 ('27% of prior mass', 'less than a factor of 2')
  • Argument: Several printed numbers do not match integrate3_results.json. (1) The prior of the excluded class is 0.309, not 0.27; the printed prior column sums to 0.96. (2) SSA excluded is 0.0158, which rounds to 0.02, not 0.01. (3) The mean likelihoods are 0.35 / 0.50 / 0.41 / 0.57 / 0.60 / 0.69, not 0.35 / 0.50 / 0.44 / 0.58 / 0.61 / 0.68. (4) 'Less than a factor of 2' is true by a hair (0.689 / 0.351 = 1.96) and mixes two things: the silence terms alone separate the six surviving classes by only 1.5, and Earth's timeline supplies the rest.
  • Proposed change: Correct the table from the JSON. Reword claim 1: 'The observations remove evident settlement (31% of prior mass, mean likelihood 0.004). Among the other six classes the silence terms differ by at most a factor of 1.5 (0.67 for en route to 1.00 for empty); with Earth's timeline included the spread is 2.0 (0.35 to 0.69).'
  • Expected effect: Presentation only. It strengthens claim 1: the silence itself is even less discriminating than stated.
  • Tested: Read the JSON. test_comp.py (800k samples) gives prior-mean silence-only likelihood by class: empty 1.000, rare 0.931, en_route 0.672, stuck 0.834, unnoticed 0.871, hidden 0.995, excluded 0.006. Timeline-only: 0.351 / 0.541 / 0.624 / 0.686 / 0.696 / 0.693 / 0.694.
• moderate / model3.py lines 111-112 and 165-168: early-life likelihood in the gated (no hard steps) branch
  • Argument: The comment says gated worlds have a fixed delay of 0.75 window after abiogenesis. Given that, and given that we arrive at x_now = 0.70-0.85, life must have started in the first 0-0.10 of the window whatever the rate, so the observed tau of 0.02-0.16 is close to guaranteed. The code instead uses q = P(life by tau | life by 0.25), which is not conditioned on x_now: mean 0.36 for slow abiogenesis and 0.77 even in the fast camp. Slow-abiogenesis gated worlds are therefore penalised about 3-fold by a datum that their own assumption predicts. Separately, worlds with m >= 1 whose steps all came out easy (4.8% of the prior) are physically the same as m = 0 but get the whole window for abiogenesis, no f_gate, and q = tau.
  • Proposed change: Either set L_early = 1 for gated worlds (consistent with the fixed delay), or model the gate as an independent clock and say so. Treat 'm >= 1, no step hard' the same way as m = 0.
  • Expected effect: With L_early = 1 for gated worlds: site empty 0.196 -> 0.227 and P(none other ever) 0.529 -> 0.550; SSA empty 0.118 -> 0.153 and P(none) 0.331 -> 0.372. In the ablation, 'timeline only' gives 0.36 instead of 0.34.
  • Tested: test_misc.py a and c, 500k samples, same seed for both rows. With every numerical fix added as well (final_combo.py, 600k): site P(none) 0.562, SSA 0.379.
• moderate / model3.py class_chain lines 107-122 with the table lookup on lines 154-158: product of per-stage probabilities, 1/h! and the 0.3 threshold; stages that are 'not hard' over a long window
  • Argument: Two approximations bias the civilisation counts upward. (1) Stages with p >= 0.3 are multiplied as if each had the whole window to itself; three stages at p = 0.7 give 0.34 against an exact sequential value of 0.12. (2) For K and M hosts a stage that saturates over the full 20-1000 Gyr window is dropped from the timing index, so the birth table treats it as instantaneous even in the first 13.8 Gyr. An M-dwarf planet 10 Gyr old is then credited with up to thousands of times the completion chance of a G planet of the same age.
  • Proposed change: Compute P(all stages done by time t) exactly as a phase-type distribution: uniformisation with non-negative matrices and repeated squaring, tabulated per sample on a doubling time grid. Use it both for the whole-window probability and for the age-to-date count. exactchain.py does this in about 30 s per 300k samples.
  • Expected effect: Small on the headline, larger than the quoted Monte Carlo error. Site P(none other ever) 0.531 -> 0.544, rare 0.303 -> 0.315, unnoticed 0.188 -> 0.180. SSA P(none) 0.334 -> 0.340. The 'planets stay habitable beyond 15 Gyr' SSA row moves 0.142 -> 0.145.
  • Tested: test_exact.py, 300k samples. Exact / model3 ratio for the whole-window probability in non-gated worlds: median 0.98, 10th percentile 0.36, 1st percentile 0.11. For N_ever to date on M hosts: median 0.71, 10th percentile 0.053, 1st percentile 0.0007; on K hosts 0.79 / 0.14 / 0.014. The exact routine was checked against closed forms for one and two stages (within 2.5% from grid interpolation). test_age.py gives the same direction with the model's own approximation evaluated at planet age (+0.006).
• moderate / model3.py lines 225-236 (D_all, share_G) and the SSA / SIA columns of three ablation rows in integrate3_results.json
  • Argument: The trapezoid on the 281-point grid cannot resolve the denominator when p_exp x N per grid step is large: for 10% of truncated samples D_all is under 0.09 of the closed form, and for 1% it is about 6e-297 of it. share_G then reaches 7e303. In the full model this is harmless, because those samples have L_local = 0. With use_local = False nothing cancels it, so the SSA columns of 'Prior', '+ Earth's timeline only' and '+ galaxy surveys (no local argument)' read excluded = 1.000 with evidence 1e301. An exact D_all does not rescue them (ESS 22-37): under SSA with pre-emption the survival-to-now factor belongs in the numerator, so 'nobody here' cannot be switched off separately in the claiming regime.
  • Proposed change: Use product integration (linear exponent between grid points; exact for the local term, whose closed form is (1 - exp(-p_exp N_alltime)) / p_exp). Put the survival factor into the SSA weight itself. Delete or grey out the SSA and SIA columns of the no-local ablation rows and say why.
  • Expected effect: Base SSA unchanged to four decimals. Three published JSON rows stop reporting nonsense.
  • Tested: test_dall.py. Without the cosmic term the product rule reproduces the closed form to 1.000 at every quantile. Base SSA with grid and exact D_all: identical (0.1198 / 0.1900 / 0.0461 / 0.3217 / 0.2816 / 0.0253 / 0.0155). Samples with D_grid / D_exact below 0.5 carry zero SSA weight in the base case.
• moderate / run3.py line 25 (camp 'Convergence-study steps'), model3.py line 51 tag '[STUDY convergence, timing]', and draft claim 9 ('clustered p = 0.12')
  • Argument: The camp uses m_weights {0: 0.35, 1: 0.39, 2: 0.20, 3: 0.06}. Those are the convergence study's first-pass figures, which its final result.json withdraws in so many words ('P(3 or more) = 0.06 is withdrawn'; compatible range 0 to 4; the author's judgement is 1 to 2 and is labelled as not derived from the data). Claim 9 quotes p = 0.12, which is one Fisher test at one age cut. The final report's own table gives odds ratio 1.44 (0.51-5.00) with permutation p = 0.47, Fisher p = 0.125 at 543 Ma and 0.040 at 1 Ga, and p = 0.008 if only the body-plan rows are collapsed.
  • Proposed change: Rename the camp 'Few hard steps (modeller's reading)' or remove it. Drop the STUDY tag from m_weights. Reword claim 9 with the study's headline statistic and the range across collapses.
  • Expected effect: No change to base numbers. The camp row (site P(none) 0.40) is no longer attributed to a study that withdrew it.
  • Tested: Read studies/convergence/result.json (the stated implications) and report.md lines 14 and 248.
• moderate / Draft claim 5 ('the first limits that do not depend on the waste-heat temperature') and model3.py line 72 (warm_curve tagged 'Lacki 2016; Griffith 2015')
  • Argument: The two K-III studies name their own predecessors: Annis 1999 (57 spirals and 106 ellipticals) and Zackrisson et al. 2015 (1,359 Tully-Fisher discs, below about 3%). Both are optical dimming limits and so are temperature independent. The new results are larger and have an explicit selection function; they are not the first. Separately, the tf_kiii novelty note says Lacki 2016 used Planck and addresses cold waste heat, while the code applies the Lacki covering-fraction curve only when waste heat is at 100-600 K. I did not check the paper itself.
  • Proposed change: Reword to 'the largest temperature-independent limits, and the first from the fundamental plane with a selection function'. Quote the Tully-Fisher limit as 1.8% (2.3% after the referee's corrections). Check which temperature range each infrared curve covers and assign warm_curve and cold_curve accordingly.
  • Expected effect: Wording only for claim 5. If the Planck curve belongs in the cold branch the survey term gets stronger there, but the whole survey term moves P(none) only from 0.514 to 0.530.
  • Tested: Read studies/fp_kiii/result.json and studies/tf_kiii/result.json (the closest prior work each cites) and tf_kiii/referee_report.md. The cold_curve values 130 / 1200 / 540 do equal the sum of 3 / (upper limit) from the two studies (107 + 23, 1034 + 130, 333 + 214).
• minor / model3.py lines 191-198: mu_vis (expanders from elsewhere visible in the survey volume) and the survey term
  • Argument: Three defects. (1) The clamp to ETA_GRID[-1] applies whenever v < 0.075c, which is 75% of en-route samples; mu_vis is then 0.4 down to 2e-7 of the geometric count of domains that intersect the survey volume. (2) kern at t_vis integrates births up to t_vis, so civilisations born after today are counted as visible now. (3) Detection is a hard switch at N_eff >= 100, and the N_eff term uses only home-grown expanders (mu_loc), not galaxies settled from elsewhere (x).
  • Proposed change: Replace both pieces with one expression: expected detections = N_eff x (1 - exp(-(mu_loc + x))) for domains smaller than the survey depth, switching to the count of intersecting domains for larger ones; integrate births to T0 only.
  • Expected effect: Negligible. Site en_route is 0.041 as coded, 0.044 with the term deleted, 0.041 with a sampling-aware version and 0.039 with pure occupancy. No other class moves by more than 0.003.
  • Tested: test_vis.py, 600k samples. Coded / geometric ratio by speed: median 1.9e-7 (1e-4 to 1e-3 c), 1.9e-4 (1e-3 to 1e-2), 0.027 (0.01-0.03), 0.41 (0.03-0.075), 1.0-1.05 above 0.075c.
• minor / model3.py lines 176-178 (T_fill, phi, mu_loc) and the '[STUDY headstart]' table on lines 62-63
  • Argument: N_ever is already 'born by now' through F0 (0.78-0.83 for G hosts, which is the older-than-the-Sun fraction in another guise). Multiplying by phi counts the head start twice. Also, T_fill is confined to 0.3-1.0 Gyr because v >= 1e-4, so only phi = 0.63-0.69 is ever used; the table rows at 2 to 9.5 Gyr are unreachable.
  • Proposed change: Use F(T0 - T_fill) / F(T0) from the same birth tables (about 0.95-0.99), or use W x phi without F0. Not both.
  • Expected effect: mu_loc rises by about 1.5 times. Site: stuck 0.251 -> 0.247, unnoticed 0.189 -> 0.192. SSA: excluded 0.015 -> 0.0125. Nothing else moves by more than 0.002.
  • Tested: test_misc.py a (option phi='F'), 500k samples.
• minor / Section A of FINDINGS_DRAFT.md against the code; cosmo_tables.py line 29; the N_reach soft spot
  • Argument: (1) 'ssa ... at any date': the birth-date density is not in the SSA weight; only the host-class share is. (2) '0 to 4 hard steps (2 to 8 decades each)': each step is log10 p ~ U(-D, 0), so the median step is 1.8 decades, 14% of nominal steps have p >= 0.3 and only 46% are below 0.01. (3) The tables use a 4.5 Gyr delay for no-step worlds; class_chain uses 0.75 x 5.6 = 4.2 Gyr. (4) F at t = 3000 Gyr is 0.947-1.032 for L >= 400 instead of 1. (5) N_reach(0.1c) = 4e5 to 1.5e6 is correct, but the reason given is incomplete.
  • Proposed change: Fix the three descriptions. For N_reach say: the same code at L = 8 Gyr and n = 3-6 returns 1.3e5-3.0e5, exactly the round-1 figure; the 5.6 Gyr window raises it 2-3 times at fixed n, and the top end (1.5e6) comes from n = 1 with civilisations allowed immediately after planet formation. A 2-3 Gyr minimum delay cuts that to 5e5-7e5.
  • Expected effect: Adding the birth-date density to SSA moves no class by more than 0.003. The rest is wording.
  • Tested: cosmo_check.py: 8M-sample Monte Carlo with analytic flat-LCDM cosmology reproduces F0, K0 and b(T0) to 0.1% for L <= 30 and within Monte Carlo noise for longer windows; t(a=1) = 13.797 Gyr, eta0 = 47.10 Gly, future reach 16.56 Gly. nreach.py: L = 8 gives 2.97e5 (n = 3) and 1.28e5 (n = 6); L = 5.6 gives 1.48e6, 8.4e5, 6.0e5, 3.8e5 for n = 1, 2, 3, 6 and 4.1e5 for n = 0. test_comp.py for the SSA density.
• minor / Draft claim 10 and the 'future' block of integrate3_results.json
  • Argument: The size of each update depends on how likely the model thinks the observation is, and that is never shown. From the stored evidences the model gives a 54% chance that an independent origin is confirmed in the Solar System, 19% that 25 HWO planets show a biosignature, and 90% that an artefact survey finds nothing. The artefact row (unnoticed 0.19 -> 0.10) is the hand-set factor 0.5 on line 215 passed through; it is an assumption, not a result.
  • Proposed change: Print the predictive probability beside each row. Label the artefact row 'assuming the survey would find half of all low-footprint presences'. Add the complementary row 'thorough search finds no second origin'.
  • Expected effect: Presentation only.
  • Tested: Ratios of stored evidences to the base evidence 0.3386: 0.182 (second genesis), 0.0636 (HWO some), 0.275 (HWO none), 0.306 (artefact null), 0.334 (technosignature null).

Claims that survive

• integrate3_results.json is exactly reproducible: a rerun of the 3M base with the current model3.py matches every class probability with zero difference, and base_params equals model3.BASE.
• The model does what its description says. My independent re-implementation (own code path, rates instead of probabilities, different sampling order) gives site 0.1955 / 0.3039 / 0.0409 / 0.2498 / 0.1887 / 0.0173 / 0.0038 against saved 0.1958 / 0.3040 / 0.0407 / 0.2503 / 0.1882 / 0.0172 / 0.0039 (max difference 0.0006), and SSA within 0.0008. The only difference above 0.01 is SIA stuck (0.317 against 0.329), which is Monte Carlo noise in the saved run.
• Twelve further rows reproduce within 0.004 on both site and SSA: four camps (Biology optimist, Carter / Snyder-Beattie, Expansion is easy, Low-footprint likely), four ablation rows and four future-observation rows.
• The headline is robust to every numerical choice I could break. Interpolating in L instead of snapping: 0.0002. Hard-step threshold 0.1 or 0.6 instead of 0.3: 0.007. Exact denominator: 0.0000. Any version of the survey visibility term: 0.003. Exact sequential-stage probabilities: 0.013. All numerical fixes together: site 0.200 / 0.312 / 0.038 / 0.242 / 0.187 / 0.017 / 0.004 with P(none) 0.543; SSA P(none) 0.339.
• cosmo_tables.npz is numerically correct: brute-force Monte Carlo with an analytic cosmology reproduces F, kern and b at the 0.1% level for windows up to 30 Gyr, and N_reach(0.1c) = 4.1e5 to 1.48e6 for the G row as claimed.
• Claim 3's numbers: P(none other ever) 0.441 -> 0.344 (timeline) -> 0.514 (nobody here) -> 0.530 (surveys) -> 0.532 (radio SETI). The round-1 complaint that the local term was inert is fixed; now the survey and SETI terms are the near-inert ones, and the draft says so.
• Claim 7's model number: SSA P(none other ever) 0.332 -> 0.143 if planets stay habitable beyond 15 Gyr (0.145 with exact stage probabilities), and 0.41 in the opposite case.
• Claim 10's numbers: second genesis gives P(none) 0.317 and empty 0.030; one HWO biosignature gives 0.273 and 0.038; artefact null takes unnoticed 0.188 -> 0.105; more radio SETI 0.532 -> 0.540.
• The class masks are exclusive and exhaustive, and the hard thresholds do not matter: a soft P(Galaxy settled | data) is 0.211 against 0.210 for the threshold classes (site), and 0.312 against 0.323 (SSA). The 'stuck' label is fair: 91% of its mass has p_exp below 1e-3 and 76% below 1e-6.
• Claim 11's direction, and more strongly than stated: once converged, five biology camps (Base, Carter, Strict ignorance, Rare Earth, and the saved Optimist) all give the same SIA split of about 0.03 / 0.32 / 0.60 / 0.06. Under SIA no biological prior matters at all.
• No clipping, overflow or mask error biases the base case. The 1e-300 and exp(-700) clips are reached only in samples whose likelihood is exactly zero.

Verdict on each headline claim

• Claim 1: reword
  • Reason: The prior mass is 0.309 in the JSON, not 0.27. The 'factor of 2' is 1.96 and comes mostly from Earth's timeline; the silence terms alone give 1.5.
  • Corrected wording: The silence rules out one thing. The observations remove evident, system-claiming expansion (31% of prior mass, mean likelihood 0.004). Among everything else the silence terms differ by at most a factor of 1.5, and by 2.0 once Earth's timeline is included. Any confident single answer is a statement of priors.
• Claim 2: reword
  • Reason: 'Roughly equal' holds only in the SSA column and in 14-20% of the project's own scan draws; site base is 0.50 / 0.25 / 0.21. Hiding is 0.056 under SIA, above the stated 0.05. The SIA figure for 'not evident' is 0.65 with hiding, 0.60 without.
  • Corrected wording: Three answers stay live, and which leads depends on the anthropic convention and four priors. (a) Nobody else has ever arisen in the Galaxy: 0.53 site, 0.33 SSA, 0 SIA; 0.26 to 0.74 across a joint prior scan (site). (b) Others arose but sustained expansion almost never happens: 0.25 to 0.32. (c) Expansion happened and is not evident from here: 0.21 site, 0.31 SSA, 0.65 SIA. Deliberate hiding ends at two to three times whatever prior it is given (0.02 to 0.06 here).
• Claim 3: keep
  • Reason: All five numbers verified against the JSON and by re-implementation. Add one caveat: the timeline-only figure is 0.36 rather than 0.34 if the no-hard-steps branch treats early life consistently with its fixed delay.
  • Corrected wording: As drafted, with '0.34 (0.36 if early life is treated as implied by the fixed delay in worlds without hard steps)'.
• Claim 4: keep
  • Reason: Figures match the flyby and headstart result files. Note for the model section: the head-start study enters the code only as a factor 0.63-0.69, and that factor is double-counted with the birth tables (effect under 0.004).
  • Corrected wording: As drafted; give the speed range as in the study (0.1c down to about 3e-5 c).
• Claim 5: reword
  • Reason: Not the first temperature-independent limits: the studies cite Annis 1999 and Zackrisson et al. 2015 as prior optical-dimming limits. The Tully-Fisher figure is 1.8% in the study and 2.3% after the referee's corrections.
  • Corrected wording: No star-harvested galaxies at any waste-heat temperature. Among about 1e5 early-type galaxies fewer than 0.29% have 75% of their starlight intercepted; among about 9,400 discs fewer than about 2%. These are the largest temperature-independent limits and the first from the fundamental plane with an explicit selection function; they are systematics-limited and insensitive above about 93% interception.
• Claim 6: keep
  • Reason: Matches seti_meta (4.4e5; 113 with 2,000 for modern-only searches). In the model the SETI term is inert, as the draft says (P(none) 0.530 -> 0.532).
  • Corrected wording: As drafted; give the Kardashev-II range as 'about 100 (2,000 if only modern searches are trusted)'.
• Claim 7: reword
  • Reason: The model number is verified (0.332 -> 0.143). It exists only in the SSA column: the site column does not respond at all (0.52), because the site convention has no host-star or date likelihood. The reverse case should be stated too.
  • Corrected wording: As drafted, ending: 'In the model this acts only through the SSA convention: learning that planets stay habitable beyond 15 Gyr moves SSA P(none other ever in the Galaxy) from 0.33 to 0.14, learning the opposite moves it to 0.41, and the site column stays at 0.52 to 0.54 either way.'
• Claim 8: keep
  • Reason: Figures match the timing study. The model's own timeline Bayes factor for fast against slow abiogenesis is 3.1 (P(fast) 0.40 -> 0.67), consistent with 'about one bit' if read as one to two bits.
  • Corrected wording: As drafted, with 'one to two bits'.
• Claim 9: reword
  • Reason: 'Clustered p = 0.12' is one Fisher test at one cut. The final report's headline is odds ratio 1.44 (0.51-5.0), permutation p = 0.47, with Fisher p from 0.04 to 0.125 by cut and 0.008 if only body-plan rows are collapsed. The model's 'Convergence-study steps' camp uses figures the study withdrew.
  • Corrected wording: Convergence data neither support nor exclude many hard steps. Vermeij's 'unique innovations are old' result weakens to an odds ratio of 1.4 (0.5 to 5.0, p = 0.47) when three stem-lineage trait clusters are counted once each, though the answer depends on which rows are collapsed (p = 0.008 to 0.47). Big brains evolved repeatedly; nothing constrains the step from brains to technology.
• Claim 10: reword
  • Reason: All numbers verified. Two additions are needed: the artefact result is a hand-set 50% detection factor passed through, and the model's own predictive probabilities for these observations (54%, 19%, 90%) should be shown.
  • Corrected wording: As drafted, adding: 'The model gives these outcomes prior probabilities of 54% (second origin confirmed), 19% (HWO biosignature) and 90% (artefact null). The artefact row assumes a survey that would find half of all low-footprint presences.'
• Claim 11: reword
  • Reason: Direction verified. It can be stated more sharply, and the SIA numbers need correcting to 0.03 / 0.32 / 0.60 / 0.06.
  • Corrected wording: Anthropics is the largest lever. Under SIA emptiness is ruled out and biology stops mattering: every biology camp gives the same split, about 0.32 'expansion fails', 0.60 'not evident', 0.06 'hidden', which is a function of the expansion and evidentness priors alone.

Code errors

• report/FINDINGS_DRAFT.md / Table B prior column; SSA and SIA rows; 'mean likelihood by class' line; claim 1
  • Problem: Prior of the excluded class printed as 0.27 (JSON 0.309; the column sums to 0.96). SSA excluded printed 0.01 (0.0158). Mean likelihoods printed 0.44 / 0.58 / 0.61 / 0.68 (JSON 0.41 / 0.57 / 0.60 / 0.69). SIA row not converged.
  • Fix: Regenerate the table from the JSON by script; use the tilted-sampler SIA values 0.03 / 0.32 / 0.60 / 0.06.
  • Impact: Claim 1 should say 31%; three SIA cells change in the second decimal.
• models/run3.py / lines 86, 124, 135-136; scan() with n=100,000
  • Problem: SIA is estimated by plain prior sampling. ESS is 4,702 of 3M (base), 44-649 in five camps, median 175 in the scan. mc_sd['sia'] (0.006-0.007 from six seeds) is three times too small.
  • Fix: Tilted proposal for the five log-uniform efficiency factors and the fast-camp switch, as described in the SIA objection; at least 16 seeds for mc_sd.
  • Impact: Base SIA stuck 0.329 -> 0.316, unnoticed 0.590 -> 0.599, hidden 0.054 -> 0.056. Camp rows change by up to 0.05.
• models/model3.py / lines 165-168 (span, q, L_early) with line 111-112
  • Problem: Gated worlds assume a fixed delay of 0.75 window, but L_early is computed as if our arrival time were unknown, penalising slow abiogenesis about 3-fold. 'm >= 1 with no hard step' worlds are treated differently from m = 0.
  • Fix: L_early = 1 for gated worlds, or model the gate as an independent clock; unify the two no-hard-step cases.
  • Impact: Site P(none) +0.021, empty +0.031; SSA P(none) +0.041, empty +0.035.
• models/model3.py / class_chain lines 113-121; lines 154-158 (nidx, F0, K0)
  • Problem: Per-stage probabilities are multiplied with a 1/h! correction only for stages below 0.3; non-hard stages take no time. On long windows a saturated stage is treated as instantaneous in the birth tables, inflating civilisations-to-date on K and M hosts by up to 1,500 times in the tail.
  • Fix: Exact phase-type completion probability per sample (exactchain.py), used for both the window total and the age-to-date count.
  • Impact: Site P(none) 0.531 -> 0.544; SSA 0.334 -> 0.340; no class moves more than 0.012.
• models/model3.py / lines 225-235 (D_all by np.trapezoid on T_GRID; share_G)
  • Problem: The trapezoid is under-resolved when p_exp x N per grid step is large: D_all is as small as 6e-297 of the closed form and share_G reaches 7e303. Harmless with the local term on; it wrecks the SSA columns of the three ablation rows without it (excluded = 1.000, evidence 1e301).
  • Fix: Product integration between grid points; put the survival factor into the SSA weight; drop SSA and SIA from no-local ablation rows.
  • Impact: Base unchanged to four decimals; three JSON rows corrected or removed.
• models/run3.py / line 25, camp 'Convergence-study steps'
  • Problem: Uses P(3 or more steps) = 0.06 and weights 0.35 / 0.39 / 0.20 / 0.06, which the convergence study's final result explicitly withdraws.
  • Fix: Rename as the modeller's own few-steps camp or delete; remove the STUDY tag on m_weights in model3.py line 51.
  • Impact: Labelling of one camp row (site P(none) 0.40).
• models/model3.py / lines 191-198 (d_eta, t_vis, Kv, mu_vis, L_survey)
  • Problem: Clamp at ETA_GRID[-1] for v < 0.075c makes mu_vis wrong by 0.4 down to 2e-7; kern at t_vis includes births after today; hard N_eff >= 100 switch; galaxies settled from elsewhere are missing from the N_eff term.
  • Fix: N_eff x (1 - exp(-(mu_loc + x))) for small domains, intersecting-domain count for large ones, births to T0 only.
  • Impact: En route 0.041 -> 0.039-0.044; nothing else above 0.003.
• models/model3.py / lines 176-178 (T_fill, phi, mu_loc)
  • Problem: phi (0.63-0.69) multiplies N_ever, which already carries the born-by-now fraction F0; the head start is counted twice. Head-start table rows beyond 1 Gyr are never reached.
  • Fix: Use F(T0 - T_fill) / F(T0) from cosmo_tables.
  • Impact: mu_loc x 1.5; class changes at most 0.004.
• models/model3.py / line 236 and docstring line 12 ('origin-date SSA')
  • Problem: The SSA weight is the G-host share only; the birth-date density b_G(T0) is not included, although the docstring and FINDINGS say 'at any date'.
  • Fix: Multiply share_G by TAB['b'][h_G, 0, I0], or change the description.
  • Impact: At most 0.003 on any class.
• models/cosmo_tables.py / line 29 (T_DELAY0 = 4.5) and lines 63-73 (600 midpoints in tau)
  • Problem: The delay for no-step worlds is 4.5 Gyr here and 0.75 x 5.6 = 4.2 Gyr in model3. With L >= 400 the tau grid resolves the star-formation history with only a few points, so F at 3000 Gyr is 0.947-1.032 instead of 1.
  • Fix: Share one constant; sample tau on a grid refined near t - tau < 20 Gyr.
  • Impact: Under 5% on K and M all-time counts; no visible effect on classes.

Missing explanations

• Settlement that is real but not permanent. The flyby study says its limit collapses for presences that fade within about 1 Myr, but the model defines expanders as persistent, so 'came, stayed a while, gone' has no class and no likelihood.
• A host-star and date likelihood under the site convention. Earliness acts only through SSA; the site column's P(M dwarfs habitable) stays at 0.295 against a prior of 0.30, so readers who reject SSA get no earliness evidence at all.
• A minimum evolutionary delay in the birth law. With one hard step the tables let civilisations appear right after planet formation, which sets the top of the N_reach range (1.5e6 at 0.1c; 5e5-7e5 with a 2-3 Gyr floor).
• An SSA 'prior' column. In the claiming regime 'not yet reached' is part of the reference-class normalisation, so the present ablation cannot say how much of the SSA shift is anthropic and how much is observation.

Critic’s own estimate of each answer

• Empty universe: 0.17
• Rare: 0.24
• En route: 0.04
• Stuck: 0.25
• Unnoticed: 0.21
• Hidden: 0.03
• Other: 0.06
• What “other” covers: Outside the class set: transient or abandoned settlement, model misspecification (the site and SSA columns disagree by 0.2 on the main number), and unknown physics. The split is roughly an even mix of the corrected site and SSA columns with a small SIA weight.

Were the round 1 fixes made?

Fixed. (1) The local 'nobody here' term now does the work (0.34 -> 0.51) and the draft says which terms are inert. (2) The galaxy-survey term uses covering-fraction curves built from the two studies (the cold curve equals the sum of 3 / limit from the fundamental-plane and Tully-Fisher results, zero above 0.93). (3) The early-life term is rate-based. (4) 1/h! is applied and only steps below 0.3 count as hard. (5) The HWO likelihood includes f_site. (6) Expanders from other galaxies, an en-route class and a continuous expansion prior replace the three atoms. (7) The SIA weight counts G-hosted civilisations only. (8) grabby.py and sdo.py are superseded by cosmo_tables, which I verified by brute force.

Not fixed. (1) SIA is still plain sampling (round 1 asked for pooling or stratifying); ESS is 4,702, the stored s.d. is three times too small, and the non-comparable SIA 'evidence' (1.1e6) is still saved. (2) The earliness likelihood exists only in the SSA column; the site column still has no host-star or date term. (3) The STUDY convergence tag is still attached to weights the study does not give, and the new 'Convergence-study steps' camp uses the withdrawn first-pass figures. (4) The existence-conditioned variant is computed in the JSON but absent from the draft.

Over-corrected or newly introduced. (1) Counting only 'still hard' steps on long windows makes saturated steps instantaneous, which inflates civilisations-to-date on K and M hosts (effect on classes about 0.01). (2) The gated branch's early-life term contradicts its own fixed delay (+0.02 to +0.04 on P(none)). (3) The head-start factor is now double-counted with the birth tables. (4) The SSA columns of three ablation rows are numerical garbage.

Presentation advice

1. Lead with ranges from the 400-draw scan, not two-decimal base values, and show which four priors drive them (number and width of hard steps, P(fast abiogenesis), expansion-prior width, P(low-footprint)).
2. State the numerical accuracy honestly: site and SSA classes are good to about +/-0.015 once model approximations are included (Monte Carlo error is 0.001); SIA to +/-0.002 only with the tilted sampler.
3. Present SIA as a three-number function of the expansion and evidentness priors, and say that biology drops out entirely. A back-of-envelope check reproduces it: at N of about 1e9, P(p_exp below 1/N) is about 0.16 under the hierarchical prior, against 0.25 for low-footprint and 0.02 for stealth.
4. Split claim 1's likelihood table into 'silence only' and 'timeline only' columns; the silence column (1.00 / 0.93 / 0.67 / 0.83 / 0.87 / 1.00 / 0.006) makes the point better than the combined one.
5. Give the predictive probability beside every future-observation row and mark the artefact row as an assumption.
6. Remove the SSA and SIA columns from ablation rows without the local term, and drop SIA from the scan table unless rerun.
7. Generate every table in the report from the JSON by script; the draft already has four hand-copied errors.

Critic: a specialist in reasoning from our own existence

Overall verdict

The three columns are implemented correctly as far as they go, but the draft's framing of them is not defensible. 'site' is not a neutral or "existence as precondition" position: it is numerically identical (largest class difference 0.005) to SSA with a reference class of civilisations on Sun-like hosts, and the true precondition reading gives a very different table (0.28 on the class the draft says is ruled out). The headline "nobody else ever in the Galaxy: about 0.5" is therefore the answer of the one reference class that maximises it; across the SSA family I get 0.20 to 0.53, under SIA 0.00, and a credence-weighted blend of 0.17 to 0.29. The SIA column's leading answer ("settled but not evident", 0.6) rests on a universe-wide behavioural switch that makes a median of 10^7.7 independent expander origins all invisible; one claiming origin per million cuts it to 0.31. The model still has no observer-weighted (Doomsday) SSA and no constructed-observer class, both requested in round 1, and its "prospects" output (mean_p_exp about 0.09 in all three columns) is a pooled prior echo that hides a spread from 1e-9 to 0.04 for evident expansion. The single most important change: relabel the columns as reference classes, replace "about 0.5" with a school table plus a stated blend (about 0.2 to 0.3), and add the finding that the schools make different testable predictions (P(HWO biosignature among 25 planets) 0.19 / 0.26 / 0.83), which is what stops the fork from being paralysing.

Objections

• major / Draft claim 2 headline '(a) Nobody else has ever arisen in the Galaxy: about 0.5 (0.33 to 0.53 by anthropic convention, 0 under SIA)'; the 'site' label in model3.py docstring (lines 11, 236) and FINDINGS section A
  • Argument: 'site' is presented as a third convention ('Earth is a given site, observations are ordinary evidence') and supplies the headline. It is not distinct from SSA: w_site equals the number of civilisations in our situation divided by the number of Sun-like-host civilisations, because Finf for the G class is 1.000 for every step count. Reweighting confirms it: SSA with a Sun-like-host reference class gives 0.193/0.300/0.043/0.251/0.186/0.017/0.009, none-ever 0.526, against site 0.196/0.304/0.041/0.249/0.189/0.017/0.004, 0.531. So the range '0.33 to 0.53 by anthropic convention' is really two reference classes of one school, and the headline uses the class that gives the largest value. Other defensible SSA classes go lower: civilisations born to date 0.415, all hosts all time 0.332, observer-weighted 0.23 to 0.27, contemporaries 0.20. 'About 0.5' is also not what any blend gives.
  • Proposed change: Rename the columns by reference class: 'SSA, Sun-like hosts' (was site), 'SSA, every civilisation past and future' (was ssa), 'SIA'. Add 'SSA, civilisations to date' and 'SSA, observer-weighted'. Replace 'about 0.5' with the school table and a blend at stated credences. State the SSA range as 0.20 to 0.53.
  • Expected effect: P(no other civilisation ever in the Galaxy): blend 0.22 at my default credences (SIA 0.40, SSA Sun-like 0.15, SSA universal 0.12, SSA to-date 0.08, SSA observer-weighted 0.10, precondition 0.05, other 0.10 left out and renormalised); 0.26 / 0.17 / 0.10 at SIA credence 0.3 / 0.5 / 0.7; 0.29 with the three v3 columns weighted equally. The headline falls from about 0.5 to about 0.2 to 0.3.
  • Tested: 3,000,000 samples (seeds 7, 11-15) from a scratch copy that exports D_all, per-class tables and likelihood terms (model3x.py, a1.py, a4.py). Base reproduced to 3 decimals.
• major / SIA column 'unnoticed 0.59' (claim 2c '0.6 under SIA'); model3.py lines 183-185: regime drawn once per universe, p_here capped at 0.3
  • Argument: Under SIA the posterior sits at N_ever of 10^8.7 per galaxy, so in the 'unnoticed' class the expected number of sustained expanders that have had time to fill the Galaxy has a median of 10^7.7 (5-95%: 10^2 to 10^14). The class survives only because the regime is a universe-wide switch: every one of those origins is low-footprint, and even then the Galaxy has at least a 70% chance of not being evident. That is the 'it only takes one' problem moved into the non-evident regime, and SIA amplifies it because SIA rewards large N. The split is also a prior ratio, as round 1 said of v2: across the draft's own camps SIA 'unnoticed' runs 0.19 to 0.91, and across the 400-draw scan 0.34 to 0.98 (median 0.80).
  • Proposed change: Add a per-origin heterogeneity parameter: a fraction eps of sustained expanders are of the claiming kind even in a 'non-evident' or 'stealth' universe, eps log-uniform 1e-8 to 1, so L_local is multiplied by exp(-eps*mu_arr) in those regimes. Report the SIA column as 'others certainly exist; which barrier is set by the prior' and show the grid, not a single 0.6.
  • Expected effect: SIA: unnoticed 0.61 falls to 0.47 (eps 1e-8), 0.31 (1e-6), 0.19 (1e-4), 0.23 (log-uniform); stuck rises 0.31 to 0.69. Site: unnoticed 0.19 to 0.09, none-ever 0.53 to 0.60. SSA: unnoticed 0.28 to 0.13, none-ever 0.33 to 0.41. With eps log-uniform, all three schools agree that, given others arose, 'expansion fails' beats 'present but unnoticed' by about 3 to 1 (0.74 / 0.24 / 0.02 for stuck / unnoticed / hidden in each).
  • Tested: a2.py on 3M samples, and an importance-sampled SIA run (a5.py, ESS 7,015): base SIA 0.025/0.318/0.603/0.051/0.003; eps 1e-6 gives stuck 0.61, unnoticed 0.31; eps log-uniform gives 0.69 / 0.23.
• major / No observer-weighted SSA branch; output mean_p_exp (model3.py line 265) proposed as 'prospects for humanity'
  • Argument: The 'ssa' column counts each civilisation once however large it becomes. That is Hanson's convention, not Bostrom's SSA over observers, and it is the one flavour of SSA that switches the Doomsday argument off. Round 1 objection 5 asked for the observer-weighted branch; it is still missing. Separately, mean_p_exp is 0.092 / 0.087 / 0.092 in the three columns against a prior of 0.120, so it looks school-independent. It is not: it pools regimes. Under SIA the probability of becoming an evident, system-claiming expander is 1.8e-9; the whole 0.09 comes from the non-evident and stealth regimes, where p_exp is unconstrained and returns its prior mean. Under site and ssa the claiming part is 0.040 and 0.027.
  • Proposed change: Add a column 'SSA, observer-weighted' with weight ssa/(1 + p_exp*R), R in {1e3, 1e6, 1e10}. Replace mean_p_exp with two outputs per school: E[p_exp and claiming regime] and E[p_exp, any kind], plus P(p_exp < 1e-6). Do not headline a single 'prospects' number; say it spans nine orders of magnitude by school and is prior plus anthropics, not data.
  • Expected effect: Observer-weighted SSA: R=1e3 gives stuck 0.53, unnoticed 0.13, none-ever 0.27, E[p_exp] 8e-4; R=1e6 gives stuck 0.66, unnoticed 0.05, none-ever 0.23, E[p_exp] below 1e-6, P(p_exp<1e-6) 0.96. Evident-expansion prospects by school: Sun-like SSA 0.040, universal SSA 0.027, SIA 2e-9, observer-weighted SSA 6e-7 to 3e-4. If only claiming expanders are populous, R=1e6 gives unnoticed 0.37 and E[p_exp, claiming] of zero.
  • Tested: a1.py, a3.py, a4.py on 3M samples; SIA figure confirmed by the importance-sampled run (1.76e-9).
• major / SIA column has no constructed-observer class and no presumptuous-philosopher caveat (round 1 objection 6, not acted on); claim 11
  • Argument: SIA weights by observers in our evidential situation whoever made them. The same factor that gives 'empty = 0.00' favours hypotheses in which civilisations create copies of our situation (ancestor simulations, cultivated or sheltered biospheres). With prior s that this happens at multiplicity K per evolved civilisation, the SIA posterior that we are constructed is sK/(1 + sK): 0.5 at s=0.001, K=1,000 and 0.91 at s=0.01, K=1,000. The SIA column itself claims a median of 5,600 technological civilisations in the Galaxy now, which makes a small s hard to defend. The column is also almost entirely a priori: 96% of its mass comes from the 6.5% of prior samples with N_ever above 1e7, the weight ratio against 'no other in the Galaxy' is at least 10^8.7, and the empirical terms discriminate among surviving classes by less than a factor of 2.
  • Proposed change: Add an explicit 'other: constructed observers' share to the SIA column as a function of s*K, and two sentences naming the presumptuous-philosopher objection as the reason a reader may discount the column. State the SIA posterior on N_ever (10^7.1 to 10^9.9) and that it is the top of the prior range, not a measurement.
  • Expected effect: Within SIA, 'other' takes 0.5 to 0.9 of the column for s*K of 1 to 10. In a 40% SIA blend that is 0.2 to 0.36 on 'other', drawn from stuck and unnoticed. I use a more cautious 0.15 in my own posterior.
  • Tested: Closed-form for the constructed share; SIA mass concentration and N_ever, N_now quantiles from a2.py and a4.py.
• moderate / Draft claim 1 ('The observations remove expansion that claims systems and would be evident here: 27% of prior mass, likelihood 0.004'); L_local in the claiming regime (model3.py line 197) versus the truncated SSA denominator (lines 225-234)
  • Argument: The model is inconsistent about what claiming means. The SSA denominator assumes claimers pre-empt later civilisations (births are truncated). If so, finding ourselves unclaimed is entailed by our existence and is not an observation. What removes the class is then anthropic typicality, which is exactly the contested step. Under a strict precondition reading (L_local = 1 in the claiming regime) the class keeps 0.28 and none-ever falls to 0.40; with a reference class of contemporaries only it gets 0.385. Those readings have a known defect: 90% of that mass has expected arrivals above 10 (median 10^4.4), meaning we would be in a galaxy spared at odds of e^-10 or worse. Any leak in pre-emption restores the penalty. Also, the draft table's prior for this class is wrong: the JSON says 0.309, not 0.27, and the draft prior column sums to 0.96.
  • Proposed change: Reword claim 1: the class is excluded at 0.3% to 1.6% in every school whose reference class spans time (SIA and all SSA variants), by typicality rather than by observation, and survives only under a bare-precondition or contemporaries-only reading that also licenses freak observers. Rename the class 'settled by claimers; we would not be here'. Correct 27% to 31%. Add a p_preempt parameter so the dependence is visible.
  • Expected effect: Excluded class: 0.004 (as coded), 0.014 at p_preempt 0.9, 0.025 at 0.99, 0.044 at 0.999999, 0.039 if only luck up to e^-10 is allowed, 0.28 at exactly 1. None-ever moves 0.53 to 0.51-0.52 except at exactly 1 (0.40).
  • Tested: a1.py and a6.py on 3M samples.
• moderate / 'ssa' column definition and draft claim 7 last sentence ('learning that planets stay habitable beyond 15 Gyr moves the SSA column's P(none) from 0.33 to 0.14')
  • Argument: The whole difference between 'ssa' and 'site' is one factor, W_G / D_all: how typical a Sun-like-host civilisation of today is among all civilisations including those on K and M dwarfs up to 3,000 Gyr hence. That factor alone, before any observation, moves none-ever from 0.44 to 0.28, because it penalises hard steps (posterior mean 1.07 against 1.71) and long-lived planets. The column therefore draws astrophysical conclusions from typicality: P(planets lose habitability within 15 Gyr) 0.72 from a prior of 0.50, P(M dwarfs habitable) 0.20 from 0.30. This is the mirror image of the presumptuous philosopher and should be disclosed. The 0.33 to 0.14 shift exists only for this reference class.
  • Proposed change: Label the column 'every civilisation that will ever exist, on any star'. Next to claim 7 give the same row for the other reference classes, and say that the earliness puzzle exists only under SSA with future long-lived-host civilisations in the class; it vanishes under SIA and under a Sun-like-host class.
  • Expected effect: Planets shown long-lived, P(none-ever): universal SSA 0.33 to 0.14; SSA to-date 0.42 to 0.38; Sun-like SSA 0.53 to 0.52; SIA no change. No school's base table changes; the claim's scope narrows.
  • Tested: a3.py: typicality-only weight gives none-ever 0.283; conditional rows as quoted. Algebra check: ssa weight equals sia weight divided by D_all, so there is no double counting with L_late or L_early.
• moderate / Draft claims 10 and 11 ('cannot be argued away'); round 1 advice that no observation moves credence between columns
  • Argument: The schools make different predictions about observations the draft already models, so the fork can be scored. Predictive probabilities from the evidence ratios in integrate3_results.json (future rows over the Base camp, same seed): at least one biosignature among 25 HWO planets 0.19 (site), 0.26 (ssa), 0.83 (sia); an independent origin of life in the Solar System 0.54, 0.69, 0.96; an Earth-level technosignature among the nearest million stars 0.01, 0.03, 0.30; a thorough artefact survey finding something 0.10, 0.14, 0.30. Claim 10 quotes only site-column shifts, and its 'more radio SETI: no change' is false for the SIA column, where technosignature and artefact searches are the informative ones.
  • Proposed change: Rewrite claim 10 as a school-by-observation table that gives both the predictive probability and the resulting shift. Rewrite claim 11: anthropics is the largest lever, it moves one number (whether anyone else has arisen), and each school stakes a different bet on HWO. A reader who treats schools as hypotheses can update between them at about 4 to 5 to 1 per HWO outcome; say that this step is itself contested.
  • Expected effect: No change to class tables. An HWO null favours the SSA columns over SIA by 0.81:0.17, about 4.9 to 1; a detection favours SIA by 0.83:0.19, about 4.4 to 1.
  • Tested: Computed from integrate3_results.json evidence values. SIA future rows have ESS 437 to 1,472, so the SIA predictive values carry roughly +-0.03.
• moderate / 'Empty universe' 0.20 (site) and 0.12 (ssa); run3.py line 117 existence variant, computed but absent from the draft
  • Argument: A universe with N_ever of 1e-36 per galaxy contains us with negligible probability unless the whole universe is vastly larger than the observable part. Any school that conditions on someone existing multiplies by 1 - exp(-N*G_tot). The JSON variant uses G_tot equal to the observable universe and gives empty 0.006, but the draft does not mention it, and G_tot is not a parameter. Full non-indexical conditioning is the general form: weight 1 - exp(-G_eff * n_E), which reproduces SIA for small universes and gives no update at all for infinite ones.
  • Proposed change: Add G_tot as a hyperparameter, print 'empty universe' as a function of it, and lead with the merged 'no other ever in the Galaxy' figure, which is stable. Mention FNC as the bridge between the SIA and SSA columns.
  • Expected effect: Site column, empty universe: 0.006 (G_tot 1e11), 0.05 (1e15), 0.10 (1e22), 0.16 (1e40), 0.18 (1e60); none-ever only 0.42 to 0.52. FNC: none-ever 0.00 (G_eff 1e-5), 0.28 (1e5), 0.44 (1e11), 0.52 (1e20), 0.60 (1e100).
  • Tested: a3.py on 3M samples.
• moderate / run3.py ABLATION rows with use_local=False, SSA column (integrate3_results.json ablation: 'Prior', '+ Earth's timeline only', '+ galaxy surveys (no local argument)')
  • Argument: With the local term off, the SSA numerator loses the survival factor but the denominator is still truncated by expanders in the claiming regime, so W_G/D_all explodes for heavily settled universes. The stored SSA results for those rows are 100% 'excluded' with none-ever 0.000 and ESS about 8,000 of 3M. Claim 3 quotes only the site column, so the draft text is unaffected, but the JSON is wrong and the SSA ablation cannot be reported.
  • Proposed change: When use_local is False use the untruncated denominator (D_all = N_alltime), and add a 'typicality only' row so the SSA ablation starts from its real baseline.
  • Expected effect: Corrected SSA ablation, none-ever: typicality only 0.28; plus timeline 0.18; plus surveys without local 0.22; all observations 0.33. For SSA the sequence is 0.44, 0.28, 0.18, 0.33, quite different from the site sequence in claim 3.
  • Tested: a3.py: reproduced the artefact (excluded 1.000, ESS 8,143) and the corrected rows.
• minor / SIA Monte Carlo (FINDINGS section D; run3.py scan with n=100,000; camps)
  • Argument: Base SIA ESS is 4,700 of 3M, which is adequate: my six 500k chunks have s.d. 0.011 on stuck and 0.009 on unnoticed, and an independent importance-sampled run agrees (0.318 / 0.603). But the 400-draw scan has SIA ESS median 175, minimum 12, 58% under 200, and two camps have ESS 44 and 56. Any SIA statistic from the scan or those camps is noise.
  • Proposed change: Sample the SIA column from a tilted proposal and reweight. A proposal with w_fast 0.95 and m_weights {0: 0.6, 1: 0.25, 2: 0.1, 3: 0.04, 4: 0.01} raised ESS per sample 3.4 times and reproduced site and ssa to 3 decimals. Suppress SIA cells with ESS under 500.
  • Expected effect: Base SIA numbers unchanged within +-0.02; scan and camp SIA cells become usable.
  • Tested: a5.py (ESS 7,015 from 1.5M); integrate3_scan.json ESS distribution.
• minor / model3.py line 178 (mu_loc uses phi) versus line 233 (denominator survival uses no phi and no fill delay)
  • Argument: In the claiming regime the numerator's survival is exp(-(x + phip_expN_ever)) with phi between 0.63 and 0.69, while the denominator integrates exp(-(x(t) + p_exp*N(t))). The same quantity is computed two ways.
  • Proposed change: Use one survival function for both, with the head-start factor applied consistently.
  • Expected effect: SSA 'excluded' 0.016 falls to 0.009; no other class moves by more than 0.004.
  • Tested: a3.py.

Verdict on each headline claim

• Claim 1: reword
  • Reason: What removes the claiming class is typicality given our existence, not an observation, if claimers pre-empt (which the SSA denominator assumes). It holds at 0.3% to 1.6% in every school with a time-spanning reference class and fails (0.28) under a bare-precondition reading. The prior mass is 31%, not 27%.
  • Corrected wording: One class is eliminated in every mainstream school: expansion that claims systems and reaches most sites (31% of prior mass, 0.3% to 1.6% afterwards). If such settlers would have prevented us, this is a typicality argument and not an observation; it fails only under a reading that treats our existence as a bare precondition. Among everything else the observations discriminate by less than a factor of 2, so any confident single answer is a statement of priors and of anthropic school.
• Claim 2: reword
  • Reason: 'About 0.5' is the Sun-like-host SSA value, the highest of the family. SSA variants span 0.20 to 0.53, SIA gives 0, blends give 0.17 to 0.29. 'Unnoticed 0.6 under SIA' depends on a universal behaviour switch and on prior ratios (0.19 to 0.91 across camps).
  • Corrected wording: Three live answers. (a) Nobody else has ever arisen in the Galaxy: 0.2 to 0.5 under self-sampling depending on reference class, 0 under self-indication, about 0.2 to 0.3 blended. (b) Others arose and sustained expansion almost never happens: 0.25 to 0.33 as modelled, 0.3 to 0.7 if expanders are not assumed to behave identically. (c) Expansion happened and is not evident: 0.2 to 0.6 as modelled, 0.1 to 0.25 without that assumption. Deliberate hiding stays under 0.06.
• Claim 3: reword
  • Reason: The sequence is for the Sun-like-host SSA column only. Under universal SSA the typicality factor alone moves 0.44 to 0.28 before any data; under SIA the timeline does nothing (P(fast) is already 0.99). The stored SSA ablation rows without the local term are artefacts.
  • Corrected wording: Ablation (Sun-like-host self-sampling): Earth's timeline 0.44 to 0.34; absence of evident settlement here back to 0.51; galaxy surveys 0.53; radio SETI nothing. Under universal self-sampling the sequence is 0.44, 0.28 (typicality alone), 0.18, 0.33. Under self-indication none of these terms moves the answer from 0.
• Claim 4: reword
  • Reason: The 1e-9 bound treats non-settlement as ordinary evidence. The headstart report itself (line 246) says it becomes an anthropic statement if settlement would have prevented us; the draft drops that caveat.
  • Corrected wording: Keep the Gaia and stellar-age numbers. Final sentence: If settlement is of the evident kind, the per-planet probability of producing a Galaxy-settling civilisation is below about 1e-9; if such settlement would have pre-empted us, the bound is a typicality statement that holds under self-indication and time-spanning self-sampling, not a direct observation.
• Claim 5: keep
  • Reason: Not school-dependent. Note only that under a selection-aware reading the surveys are the sole ordinary evidence about claiming expanders, and in the model they are weak (mean survey likelihood in the excluded class 0.63).
• Claim 6: reword
  • Reason: True as a statement about the record. But under SIA the model implies a median of 5,600 extant civilisations in the Galaxy and a 30% chance of an Earth-level technosignature among the nearest million stars, so SETI is the informative programme in that column.
  • Corrected wording: Add: under self-indication the model expects thousands of extant civilisations in the Galaxy (5-95%: 4 to 1.7e7) and gives a 30% chance that a search of the nearest million stars at Earth-level sensitivity succeeds, against 1% to 3% under self-sampling.
• Claim 7: reword
  • Reason: Earliness is a puzzle only under self-sampling with future long-lived-host civilisations in the reference class. The 0.33 to 0.14 shift is specific to that class.
  • Corrected wording: Replace the last sentence: learning that planets stay habitable beyond 15 Gyr moves P(none other ever in the Galaxy) from 0.33 to 0.14 if the reference class is every civilisation that will ever exist, from 0.42 to 0.38 if it is civilisations to date, and not at all for a Sun-like-host class (0.53 to 0.52) or under self-indication. Add: that column also concludes from typicality alone that planets probably lose habitability within 15 Gyr (0.72 from a prior of 0.50).
• Claim 8: reword
  • Reason: 'About one bit' is correct for the non-SIA columns. Under SIA the existence weight has already pushed P(fast abiogenesis) to 0.99 and the timeline adds nothing.
  • Corrected wording: Add one clause: this is the weight of the timeline under self-sampling; under self-indication our existence already puts P(fast abiogenesis) at 0.99 and the timeline is redundant.
• Claim 9: keep
  • Reason: Not affected by the anthropic fork, except that the universal-SSA column independently pushes toward fewer hard steps (posterior mean 1.07 against 1.71), which is typicality and not the convergence data.
• Claim 10: reword
  • Reason: Quotes site-column shifts only. The schools differ both in what they predict and in what would move them; 'more radio SETI: no change' is false under SIA.
  • Corrected wording: Give a table by school. P(at least one biosignature among 25 HWO planets): 0.19 / 0.26 / 0.83 (Sun-like SSA / universal SSA / SIA); second origin in the Solar System: 0.54 / 0.69 / 0.96. If seen, P(none other ever) falls 0.53 to 0.27-0.32 and 0.33 to 0.12-0.16; SIA unchanged. Artefact survey null: unnoticed 0.19 to 0.10, 0.28 to 0.16, 0.60 to 0.43. Technosignature searches matter only in the SIA column.
• Claim 11: reword
  • Reason: There are two levers (does existence count; which reference class), the fork is scoreable by HWO at about 4 to 5 to 1, the SIA column needs the constructed-observer and presumptuous-philosopher caveats, and the observer-weighted SSA school is missing.
  • Corrected wording: Anthropic reasoning is the largest lever and it moves one number. If Earth is treated as a random habitable planet that happened to produce us (self-indication), emptiness is ruled out and the answer is 'they exist; expansion fails or is not evident', with that split set by priors. If Earth is singled out because it produced us (self-sampling), P(alone in the Galaxy) is 0.2 to 0.5 by reference class. The two make different bets on HWO (0.83 against 0.19 to 0.26), so the fork is testable.

Claims that survive

• The SIA weight is implemented as intended: w_site * W_G is the expected number of civilisations in our evidential situation (Sun-like host, our planet age and arrival fraction via L_late, early life, unreached, null surveys). 'SIA restricted to unreached civilisations' is what the code already does.
• No double counting between the SSA weight and the Carter timing terms. Algebraically ssa = sia / D_all; L_late * W_G is the density of Sun-like-host civilisations at our date and W_G / D_all is the host and date share. Each factor appears once.
• Not truncating births in the non-evident and stealth regimes is consistent with how those regimes are defined (settlers claim no planets; hiders leave living planets alone), and numerator and denominator agree there.
• P(no other civilisation ever in the Galaxy) = 0.000 under SIA in every variant I ran, including the heterogeneity variants and an independent importance-sampled run (ESS 7,015). 'Earth as a random Sun-like-star planet' gives the same table (0.000/0.000/0.029/0.349/0.569/0.051/0.003), so the result does not depend on counting planets.
• Base SIA class numbers are stable to about +-0.02: pooled 3M run 0.024/0.305/0.614/0.054, importance-sampled 0.025/0.318/0.603/0.051, draft 0.03/0.33/0.59/0.05.
• Evident claiming settlement of most sites is at or below 1.6% in every school whose reference class spans time (SIA 0.3%, Sun-like SSA 0.4% to 0.9%, universal SSA 1.6%, observer-weighted SSA 0.7% to 1.4%).
• Deliberate hiding stays at or below 0.054 in every school and every variant at the base prior (0.004 to 0.054).
• Expanders already en route stay between 0.024 and 0.065 in every school and variant.
• Given that others have arisen in the Galaxy, the split between 'expansion fails' and 'present but unnoticed' is set by the priors on low-footprint settlement and on W_exp in every school; with heterogeneous expanders all three schools give about 0.74 / 0.24 / 0.02.
• Claim 1's second half survives: among the surviving classes the empirical likelihood terms discriminate by less than a factor of 2 (0.35 to 0.68), in every school.

Code errors

• models/model3.py with models/run3.py / model3.py lines 225-236 (D_all truncation) combined with run3.py ABLATION rows that set use_local=False
  • Problem: With the local term off, the SSA numerator has no survival factor but the denominator is still truncated in the claiming regime, so W_G/D_all diverges for heavily settled universes. The stored SSA results for 'Prior', '+ Earth's timeline only' and '+ galaxy surveys (no local argument)' are 100% 'excluded' with none-ever 0.000.
  • Fix: If use_local is False, set D_all = N_alltime (no truncation). Add a 'typicality only' ablation row for the SSA column.
  • Impact: Corrected SSA none-ever: 0.28 (typicality only), 0.18 (plus timeline), 0.22 (plus surveys, no local). Draft text unaffected because claim 3 quotes site only; the JSON is wrong.
• report/FINDINGS_DRAFT.md / Section B table, prior column and SSA column; claim 1
  • Problem: Prior for 'settled in a way we would see' is printed as 0.27; integrate3_results.json base.prior.excluded is 0.309 and the printed prior column sums to 0.96. The SSA entry for the same class is printed 0.01; the JSON value is 0.016.
  • Fix: Print 0.31 and 0.02; change '27% of prior mass' to 31% in claim 1.
  • Impact: Presentation only.
• models/model3.py / line 178 (mu_loc = N_ever * phi * p_exp) versus line 233 (nlogS uses p_exp * sum(W*F), no phi)
  • Problem: Survival against local expanders is computed with the head-start factor in the SSA numerator and without it in the denominator.
  • Fix: Use one survival function in both places.
  • Impact: SSA 'excluded' 0.016 becomes 0.009; other classes move by under 0.004.
• models/model3.py / lines 256-266 (mean_p_exp, p_pexp_gt_1pct)
  • Problem: The proposed 'prospects' output pools the claiming, non-evident and stealth regimes. Under SIA it reports 0.085 to 0.092 while the probability of evident, claiming expansion is 1.8e-9.
  • Fix: Report E[p_exp * claiming], E[p_exp * (non-evident or stealth)] and P(p_exp < 1e-6) separately, and add the observer-weighted SSA variant.
  • Impact: Evident-expansion prospects by school: 0.040 / 0.027 / 2e-9, and below 3e-4 under observer-weighted SSA, instead of an apparent 0.09 everywhere.
• models/run3.py / line 117 (conditional_on_someone_in_observable_universe)
  • Problem: The existence factor uses the observable-universe galaxy count as the size of the whole universe and is attached to the site weights only. It is a single point on a curve and is not carried into the draft.
  • Fix: Make G_tot a parameter, tabulate 1e11 to 1e60, and compute it for the SSA column as well.
  • Impact: Site 'empty universe' 0.006 at 1e11, 0.10 at 1e22, 0.16 at 1e40, 0.18 at 1e60.
• models/run3.py / scan() with n=100,000 per draw; CAMPS at n=1,000,000
  • Problem: SIA effective sample size is too small outside the base run: scan median 175, minimum 12, 58% under 200; 'Carter / Snyder-Beattie' camp 44, 'Strict ignorance' camp 56.
  • Fix: Importance-sample the SIA column (a proposal with w_fast=0.95 and m_weights tilted to few steps gave 3.4 times the ESS per sample and reproduced site and ssa to 3 decimals), or suppress SIA cells with ESS below 500.
  • Impact: SIA scan ranges and those two camp rows are currently noise.
• models/model3.py / line 197 (L_local = 1 - (-expm1(-mu_arr)) * p_here) and line 257 ('evidence')
  • Problem: For the claiming regime the expression loses all precision above mu_arr of about 36 (returns exactly 0 or multiples of 1.1e-16). 'evidence' is stored under one key for three schools whose values (0.34, 0.066, 1.1e6) are not comparable.
  • Fix: Use exp(-mu_arr) directly where p_here is 1. Tag evidence by school and document that only within-school ratios are meaningful.
  • Impact: Negligible at base settings; matters for any variant that divides by a truncated denominator.

Missing explanations

• Constructed observers (ancestor simulations, cultivated or sheltered biospheres). Under SIA this class takes 0.5 to 0.9 of the column once prior times multiplicity reaches 1 to 10, and it changes what the astronomical likelihoods mean.
• Universal doomsday as an inference (Gerig, Olum and Vilenkin 2013, cited in gaps_and_corrections.md and still unused): observer-weighted SSA gives stuck 0.53 to 0.75 and drives the expansion probability below 1e-3.
• 'Unclaimed remnant': claiming expansion is common, most galaxies are taken, and we exist only because ours was spared. Zero in the site column by construction; 0.03 to 0.04 if moderate luck (up to e^-10) is allowed and pre-emption is near-total.
• Heterogeneous expanders: most are low-footprint but a small fraction claim systems. The model has only all-or-nothing regimes per universe; a mixed population removes most of 'unnoticed' under SIA.
• Common but briefly detectable civilisations. The SIA column implies a median of 5,600 extant in the Galaxy with detectable lifetimes around 1e5 years; the class scheme has nowhere to put this and files it under 'stuck' or 'unnoticed'.
• The decision-theoretic reading (Armstrong): total-welfare agents should act on the SIA column, average-welfare or self-interested agents on an SSA column. This lets a reader choose a column by purpose without settling the philosophy.

Critic’s own estimate of each answer

• Empty universe: 0.07
• Rare: 0.12
• En route: 0.04
• Stuck: 0.375
• Unnoticed: 0.225
• Hidden: 0.02
• Other: 0.15
• What “other” covers: Constructed observers (simulation or cultivated biosphere), which SIA-style reasoning favours; failure of the anthropic framework itself; and about 0.02 for 'unclaimed remnant' (claiming expansion is common and our galaxy was spared). Built from school credences SIA 0.40, SSA Sun-like hosts 0.15, SSA universal 0.12, SSA to-date 0.08, SSA observer-weighted 0.10, bare precondition 0.05, other 0.10, averaging the as-coded tables with the heterogeneous-expander variant.

Were the round 1 fixes made?

Fixed: (9) early-life likelihood is now rate-based; (11) SIA and SSA count Sun-like-host civilisations only; (12) intergalactic pre-emption via the LambdaCDM kernel; (8) the SSA column now carries the earliness argument and future-observation rows are run in all three schools. Partly fixed: (1) three co-equal columns exist, but claim 2 still headlines the 'site' value and 'site' still charges for pre-emption without an existence weight; I find it equals SSA with a Sun-like-host class, so it is a legitimate school under the wrong name. (2) Claim 11 concedes the SIA split is prior-set, yet claim 2 still prints 0.6 and no grid is shown. (4) An existence-conditioned variant is computed at one universe size and left out of the draft. (7) SIA still sits at the prior ceiling (N_ever 10^8.7) with no importance sampling. Not fixed: (3) no p_preempt parameter, and the local term is now load-bearing (0.34 to 0.51), which is the exposure round 1 warned about; (5) no observer-weighted SSA and no usable prospects axis; (6) no constructed-observer class and no presumptuous-philosopher caveat; (10) 'evidence' still shares one key across schools. Over-corrected: the SSA column now folds host typicality and far-future earliness into one undisclosed factor that moves the answer from 0.44 to 0.28 before any data; and the non-evident regime as a universe-wide switch gives SIA an escape that needs about 10^7.7 expander origins to be uniformly invisible.

Presentation advice

1. Put the fork as one plain question: "Is Earth a random habitable planet that happened to produce us, or is it singled out because it produced us?" The first reading is self-indication: our existence is strong evidence that civilisations are common. The second is self-sampling: it is no evidence, and the answer then depends on whom we compare ourselves with.

2. Say that the fork moves one number, P(no other civilisation has ever arisen in the Galaxy), and print it per school with my suggested credences:

Blend over the computed schools: 0.22 (0.26 / 0.17 / 0.10 at SIA credence 0.3 / 0.5 / 0.7). Full blended class table: empty 0.08, rare 0.13, en route 0.04, stuck 0.33, unnoticed 0.38, hidden 0.03, claimed-remnant 0.02; with heterogeneous expanders 0.09 / 0.15 / 0.05 / 0.53 / 0.15 / 0.01 / 0.03. Fix the credences as inputs and label the blend as the project's all-things-considered figure.

3. Lead with what no school disputes: evident claiming settlement of most sites is under 2%; deliberate hiding under 6%; expanders en route 2% to 6%; and, given others arose, the choice between 'expansion fails' and 'unnoticed' is a prior, the same in every school.

4. Do not print a seven-way SIA split as a finding. Print "others certainly exist; which barrier is set by the prior" with the range (unnoticed 0.19 to 0.91 across camps), and add two sentences: SIA reaches this from the armchair (96% of its mass is the top 6.5% of the prior), and the same logic favours constructed observers.

5. Add the bet: P(at least one biosignature among 25 HWO planets) is 0.19 / 0.26 / 0.83 and P(second origin in the Solar System) 0.54 / 0.69 / 0.96 by school. This is the answer to "paralysing": the columns are rival forecasts that the next two decades can score.

6. Drop 'prospects for humanity' as a single number. If kept, give P(becoming an evident, system-claiming expander): 0.04 (Sun-like SSA), 0.03 (universal SSA), about 1e-9 (SIA), below 3e-4 (observer-weighted SSA), and say it is prior plus anthropics.

7. Give each SSA column its own caveat: the universal class infers planetary lifetimes from typicality (0.72 that planets die within 15 Gyr); the Sun-like class discards the host-star evidence the project's own M-dwarf study measures; 'empty universe' needs a universe of 1e40 or more galaxy-equivalents.

Citations from memory, not re-checked for this review: Bostrom 2002 (Anthropic Bias; presumptuous philosopher), Neal 2006 (full non-indexical conditioning), Carlsmith 2021 (SIA versus SSA essays), Armstrong (anthropic decision theory), Grace 2010 (SIA doomsday), Gerig, Olum and Vilenkin 2013. The PhilPapers 2020 Sleeping Beauty split (roughly 28% thirder, 19% halfer, remainder undecided, as I recall it) is my basis for putting SIA-like views near 0.4 and not treating either school as the default.

Critic: the camp that holds “it only takes one”

Overall verdict

From this lens the draft headline "three live answers, roughly equal" is not defensible as worded; "any confident answer is a statement of priors" is, and more strongly than the draft says. Two of the three answers are hand-set prior masses passed through: "not evident" is the world-level coin p_nonevident = 0.25 (posterior 0.19, about 0.75 times the prior), and 0.16 of the 0.25 for "stuck" is the prior tail P(p_exp < 1e-8) = 0.17 multiplied by 1.7; under SIA the whole column is the normalised ratio of those masses. The non-evident class is a monoculture: its posterior worlds contain a median of 27,000 independently arisen expander lineages (interquartile range 218 to 6 million; 5 million Milky-Way-born under SIA), all assumed low-footprint by a single draw, with the chance of being evident capped at 0.3 however many there are. If each lineage chooses for itself, "not evident" falls from 0.19 / 0.28 / 0.59 to 0.04 / 0.05 / 0.06 (site / SSA / SIA) and P(no other civilisation ever in the Galaxy) rises from 0.53 to 0.62. That is the single most important change: make evidentness a per-lineage property plus an explicit common-cause prior, and report "stuck" split into a universal barrier (which also stops us) and "a few others, none happened to expand". My own position is weaker than Hart's: with priors I would defend, P(none other ever) is 0.61 (site) and 0.42 (SSA), ranging 0.48 to 0.74 as the two common-cause priors are varied, and it reaches 0.87 only if expansion is also assumed easy. Human-anchoring the expansion prior changed almost nothing numerically (stuck 0.250 to 0.251), because the model's 30-decade tail already behaves like a barrier prior of about 0.17; that objection is about labelling and derivation, not numbers. All scratch code and logs are in the critic’s working folder (model3x.py reproduces the base case exactly with hooks off).

Objections

• fatal / Draft claims 2(c) and 11; model3.py lines 183-185 and 197: one world-level regime draw (p_nonevident = 0.25) and L_local = 1 - (1 - exp(-mu_arr)) * p_here with p_here capped at 0.3
  • Argument: Fatal to the claim as worded, not to the model. Evidentness is drawn once per universe and applied to every expander lineage in it, and the penalty saturates at p_here <= 0.3 however many lineages arrive. In the posterior 'unnoticed' class the number of independently arisen expander lineages that have reached us (mu_arr) has quantiles 10/25/50/75/90% = 9.6 / 218 / 27,000 / 6.2 million / 6.7 billion; Milky-Way-born alone the median is 400 (site), 2,600 (SSA) and 4.8 million (SIA). So the class mostly asserts that thousands to millions of separately evolved species each became persistent expanders and every one chose a footprint we cannot see. That is the monocultural assumption the project's own behavioural brief warns against; the brief prescribes (1 - e)^N for independent choices plus one common-cause Bernoulli, and the model implements neither for this regime. For the geometric-mean world (82,000 lineages) to keep a likelihood of 0.5, fewer than 1 lineage in 118,000 may be of the evident kind. There is also a logical point about selection: a lineage that 'claims no planets' by definition cannot exclude a later lineage that does, so whether a galaxy is evidently settled is decided by whether any claiming lineage exists, which is exp(-mu * c), not a coin shared by all.
  • Proposed change: Make evidentness per lineage. Minimal version: in the non-evident regime use L_local = exp(-mu_arr * p_here). Better version: keep a small common-cause prior q_cc for 'every viable form of expansion is low-footprint', and in all other worlds draw c = fraction of independent lineages that are of the claiming kind (uniform 0-1 is neutral) with L_local = exp(-mu_arr * c) * exp(-mu_arr * (1 - c) * p_here). Classify a world as 'unnoticed' only if mu_arr * c < 1 and mu_arr * (1 - c) >= 1. Report the lineage count of the class next to its probability.
  • Expected effect: 'Unnoticed' falls by a factor of 3 to 10: site 0.19 to 0.04-0.06, SSA 0.28 to 0.05-0.09, SIA 0.59 to 0.06-0.23. P(no other civilisation ever in the Galaxy) rises 0.53 to 0.61-0.62 (site) and 0.33 to 0.41-0.42 (SSA). 'Stuck' rises to 0.29-0.30 (site) and 0.41 (SSA) because it absorbs the freed mass. The silence then discriminates against multi-lineage non-evident settlement by a factor of 6 (class mean likelihood 0.60 to 0.10).
  • Tested: 3M samples each. Per-origin only: site 0.232/0.360/0.048/0.297/0.038/0.020/0.005, none-ever 0.621; SSA 0.153/0.245/0.060/0.416/0.047/0.033/0.046, none-ever 0.422; SIA 0/0/0.057/0.754/0.060/0.123/0.006 (ESS 1,925). Lineage mix with c uniform and q_cc 0.05: site unnoticed 0.054, none-ever 0.609; SSA 0.090, 0.414; SIA stuck 0.598, unnoticed 0.248, hidden 0.099. Reweighting the base unnoticed class by a per-lineage evident probability c leaves 0.42 (c = 1e-4), 0.30 (1e-3), 0.19 (1e-2), 0.08 (0.1), 0.024 (0.5), 0.038 (c uniform) of it.
• major / Draft claim 2(b) and soft spot 2; model3.py line 59 and 172-173: W_exp = (1, 30), described as 'the same ignorance structure as the biological links'
  • Argument: It is not the same structure. Each hard step gets 2 to 8 decades; expansion gets up to 30 with no 'easy' camp. The prior that results puts 0.34 below 1e-4, 0.24 below 1e-6, 0.17 below 1e-8 and 0.13 below 1e-10. A per-civilisation probability of 1e-10 or less cannot be a contingent matter of survival or will; it can only be a universal barrier. So the tail is a barrier atom in disguise, and the 'stuck' result is that atom restated: of the 0.250 posterior, 0.161 has p_exp < 1e-8 (prior 0.173 times P(N_ever >= 1) = 0.097, times mean likelihood over evidence = 1.66), 0.044 is 'more than 100 civilisations, no barrier' and 0.046 is 'a handful, none happened to expand'. Under SIA 0.324 of 0.329 is the tail. Round 1 objected that the infeasible/doom/hidden split was input, not output; version 3 replaced the three atoms with a tail mass and a regime coin, and the same objection holds.
  • Proposed change: Either give expansion the hard-step width, W in (2, 8), as the literal reading of 'same structure', or state the barrier explicitly: q_barrier = P(self-sustaining interstellar settlement is infeasible for everyone) with a written derivation and range, plus a bounded continuous part. Report 'stuck' in three rows: universal barrier, many civilisations without a barrier, few civilisations. Show prior and posterior side by side for the barrier row.
  • Expected effect: With W in (2, 8): stuck 0.25 to 0.11 (site), 0.32 to 0.12 (SSA), 0.33 to 0.001 (SIA); none-ever 0.53 to 0.60 (site). With an explicit barrier atom the barrier part of 'stuck' is about 1.7 times the prior chosen: stuck 0.14-0.20 for 0.05, 0.23-0.32 for 0.15, 0.34-0.40 for 0.30. The 'few others, none expanded' part stays near 0.05 and the soft part in total near 0.09 under every variant.
  • Tested: 1M samples. W (2,8): site 0.229/0.326/0.043/0.107/0.265/0.024/0.005; SSA stuck 0.117; SIA stuck 0.001, unnoticed 0.908. W (1,8): stuck 0.082. W (1,12): 0.124. Barrier atom q with log-uniform (1e-4,1) / (1e-6,1) / (1e-8,1) otherwise: q = 0.05 gives stuck 0.143 / 0.195 / 0.258; q = 0.15 gives 0.234 / 0.272 / 0.320; q = 0.30 gives 0.344 / 0.368 / 0.399. Tail masses checked with 2M prior draws.
• major / Draft claim 2(b) wording and soft spot 2: p_exp is applied to every civilisation including us, with no output for humanity's own prospects
  • Argument: The model's posterior implies, without saying so, that the chance humanity ever becomes a sustained expander has median 4e-4, is below one in a million with probability 0.36 (site), 0.39 (SSA), 0.43 (SIA), and below 1e-10 with probability 0.21. Inside 'stuck' the median is 5e-11 (site) and 8e-16 (SIA). So 'sustained expansion almost never happens' is a prediction of our own failure, the Great Filter ahead of us, presented as a mild behavioural tendency. This is where symmetric ignorance with abiogenesis fails: for abiogenesis our existence is guaranteed by selection whatever the rate, but nothing selects us on our future expansion, so we are an ordinary sample at the threshold and can inspect the chain. The project's own longevity brief does that: P(survive to capability) Beta(2.5, 1.2), mean 0.68; P(expand given survival) log-uniform 1e-3 to 1; 'draw it per civilization, not once for the Galaxy'. That product has median 0.02 and puts 0.0002 below 1e-4, where the model's prior puts 0.34. Motive is not the bottleneck either, since factions that want to expand already exist; the low tail can only be infeasibility or universal extinction, and published hazard estimates supply at most one decade. Round-1 Hansonian objection 7 asked for this output and it was not added.
  • Proposed change: Add outputs: P(humanity becomes a sustained expander) overall and by class, and P(p_exp < 1e-6). Relabel 'stuck' as 'a universal barrier stops sustained expansion, for us too' plus 'few others, none happened to expand'. Replace the 30-decade tail by q_barrier (I would defend 0.15, range 0.05-0.30) plus a human-anchored continuous part: survival (0.27-0.97) times attempt (0.5-1) times self-sustaining given no barrier (1e-3 to 1), or log-uniform 1e-5 to 1.
  • Expected effect: Almost none numerically at q_barrier = 0.15: stuck 0.250 to 0.251 (log-uniform) or 0.230 (product); none-ever 0.53 to 0.53-0.54. The model's tail is already equivalent to a barrier prior of about 0.17, so the change is in derivation and labelling. It matters only if the barrier is argued down: at 0.05, stuck 0.14-0.20 and none-ever 0.56-0.58.
  • Tested: 3M samples: barrier 0.15 plus log-uniform (1e-5,1): site 0.196/0.307/0.036/0.251/0.189/0.017/0.004. Barrier 0.15 plus the three-factor human product: 0.202/0.305/0.034/0.230/0.207/0.019/0.004, soft 'many civilisations' part 0.044 to 0.013. Longevity-brief product sampled with 2M draws: mean 0.098, median 0.020, 5th percentile 8.8e-4. Posterior p_exp quantiles from a 2M base run.
• major / Soft spot 1: p_nonevident = 0.25 (model3.py line 66), and whether selection among lineages should move the regime prior
  • Argument: 0.25 has no derivation and sits at the top of the project's own range: the behavioural brief gives 'present but unnoticed' 0.08 (0.01-0.30) conditional on at least 10 civilisations. It was set to meet two round-1 critics who asked for 0.3-0.5, with no critic on the other side. The joint scan shows the output tracks it: correlation of 'unnoticed' with p_nonevident is 0.68 (site), 0.72 (SSA), slope 0.64 and 0.81 per unit. Deliberate stealth gets 0.02 on the grounds that it needs one permanent hegemon, yet a universal low footprint with no enforcer gets 12 times more; in a world with thousands of lineages the enforced version is the more coherent of the two. What a common-cause version requires: (i) biological or large-scale settlement is infeasible for everyone, and (ii) machine lineages universally gain nothing from more matter or energy, against the general expectation that resources are useful for almost any goal. I put that at roughly (0.3-0.5) x (0.05-0.2), so 0.02-0.10, central 0.05. That is a judgement, and a sceptic's 0.25 is not refutable; but it should be stated as the prior on a universal cause, not as 'low-footprint settlement'.
  • Proposed change: Set the common-cause low-footprint prior to 0.05 (range 0.02-0.10) with the decomposition written out, keep stealth at 0.02, and let all other worlds be per-lineage mixtures (previous objection). Publish a two-way table of P(none other ever) against the barrier prior and the common-cause prior.
  • Expected effect: With the per-lineage structure, 'unnoticed' is close to the common-cause prior: 0.19 / 0.093 / 0.055 / 0.029 (site) for 0.25 / 0.10 / 0.05 / 0.02. Across barrier 0.05-0.30 and common cause 0.02-0.25, P(none other ever) spans 0.48-0.71 (site) and 0.29-0.53 (SSA): the headline is close to linear in two hand-set priors.
  • Tested: 1M samples per row. SSA unnoticed 0.286 / 0.149 / 0.091 / 0.051; SIA 0.596 / 0.381 / 0.234 / 0.110. Shape of the per-lineage prior barely matters at common cause 0.05: Beta(2,1) 0.048, uniform 0.055, Beta(0.5,0.5) 0.061, Beta(1,3) 0.066, Beta(1,9) 0.083. Grid of 12 cells at 400k samples (grid_out.json). Scan statistics computed from integrate3_scan.json.
• moderate / p_here = log-uniform (1e-3, 0.3), model3.py line 67: not tied to the project's own Gaia and Dyson-swarm numbers
  • Argument: The flyby study counts 21,000 systems within 1 pc and 2.1 million within 10 pc since the Great Oxidation. In a settled Galaxy, p_here <= 0.3 is the statement that the chance a passing settled system leaves any evident, lasting presence here is below 1.7e-5 (1 pc) or 1.7e-7 (10 pc); at the prior's geometric mean (0.017) it is 8e-7 and 8e-9. The same regime must also keep heavily shrouded stars below 4.8e-4 (covering above 50%) and 8.6e-5 (above 90%) of stars within 300 pc, and leave the Sun unharvested. None of these numbers enters the model; the regime is given a likelihood of at least 0.7 (mean 0.95) by fiat. The cap also means elapsed time is irrelevant: a Galaxy settled for 4 Gyr is no more evident than one settled yesterday.
  • Proposed change: Derive p_here from the exposure: p_here = 1 - exp(-N_pass(d) x f_occ x x_ev), with N_pass from the flyby study, f_occ the settled fraction of systems and x_ev the per-passage chance of an evident lasting presence, each with a stated prior; and add the Suazo fractions as a likelihood on the fraction of settled systems that are heavily shrouded. State in the report what per-system restraint the regime needs.
  • Expected effect: Unnoticed 0.19 to 0.04-0.11 (site), 0.28 to 0.06-0.17 (SSA), depending on the priors for f_occ and x_ev. Merely removing the 0.3 cap does little (0.173), which shows the cap is not the main problem; the world-level structure is.
  • Tested: 1M samples. N_pass 2.1e4, f_occ log-uniform (1e-3,1), x_ev (1e-6,1): unnoticed 0.108 (site), 0.169 (SSA), 0.433 (SIA), none-ever 0.580. f_occ (1e-2,1), x_ev (1e-4,1): 0.037, 0.059, 0.182, none-ever 0.623. p_here log-uniform (1e-3,1): 0.173. The Dyson-swarm term was not implemented.
• moderate / model3.py lines 172-173: one p_exp per universe shared by every civilisation; coherence of 'stuck' worlds with thousands to millions of civilisations
  • Argument: In the 'stuck' posterior 55% of the mass has more than 1,000 civilisations ever in the Galaxy, 17% more than a million (SSA 70% and 28%; SIA median 400 million), with a median N x p of 4e-8. All of them share one propensity. Civilisations differ in biology, planet, lifespan and culture, and the number of expanders is set by the population mean, which the most expansion-prone minority dominates. The round-1 expansion sceptic found this himself ('it only takes one has real force unless the lower tail is heavy') and the longevity brief says to draw the factor per civilisation. Version 3 dropped scatter and adopted the heavy tail. With scatter, a low mean can only come from a barrier that binds everyone, which is the point: many-civilisation 'stuck' worlds are coherent only as a common cause.
  • Proposed change: Treat the drawn p_exp as the median civilisation and add log-normal scatter between civilisations (0.5-3 dex, itself uncertain), using the truncated-lognormal mean in mu; exempt only an explicit barrier atom. Or keep the shared value and say in the report that it is the population mean and that values below about 1e-8 mean a universal barrier.
  • Expected effect: Stuck 0.25 to 0.17-0.18 (site), 0.32 to 0.22-0.24 (SSA), 0.33 to 0.20-0.23 (SIA) for 0.5-3 or 1-2 dex; to 0.10 / 0.12 / 0.07 for 2-4 dex. None-ever rises 0.53 to 0.56-0.60. The freed mass goes mostly to 'unnoticed' unless the per-lineage fix is also made.
  • Tested: 1M samples each on the base prior: scatter uniform 0.5-3 dex: site 0.218/0.307/0.039/0.169/0.241/0.022/0.004; 1-2 dex: stuck 0.183; 2-4 dex: stuck 0.096, none-ever 0.598. Class composition from a 2M base run.
• moderate / run3.py CAMPS (lines 22-34) and claim 2(a) 'about 0.5 (0.33 to 0.53)'
  • Argument: Each late-stage camp changes one lever, which was a round-1 request, but the two Hart-side levers are never crossed. Alone, each only moves mass between the two escape classes: 'expansion is easy' cuts stuck to 0.045 and unnoticed rises to 0.317; 'settlement would be obvious' cuts unnoticed to 0.028 and stuck rises to 0.298. A reader concludes that P(none other ever) cannot pass about 0.64 on these biology priors. That is an artefact of the camp design.
  • Proposed change: Add the crossed camp (W_exp (1,4), p_nonevident 0.03, p_stealth 0.003) and quote the range of P(none other ever) across late-stage camps including it.
  • Expected effect: Crossed camp: none-ever 0.87 (site), 0.74 (SSA). The stated range for answer (a) becomes roughly 0.43-0.87 (site) across late-stage camps, not 0.33-0.53.
  • Tested: 1M samples: site 0.369/0.455/0.048/0.056/0.057/0.006/0.009; SSA 0.302/0.382/0.061/0.086/0.112/0.012/0.045. SIA has ESS 88 there and is not usable.
• moderate / Draft claim 1: 'rules out one thing and one thing only ... discriminates among everything else by less than a factor of 2'; base table prior column
  • Argument: True by construction, not a finding. The two late-stage survivors are the two classes given parameters that no observation in the model can reach: p_here <= 0.3 guarantees L_local >= 0.7 in the non-evident regime, and the expansion tail has no observable consequence. With per-lineage evidentness the silence does discriminate: the mean likelihood of 'unnoticed' is 0.10 against 0.50-0.57 for rare and stuck. Separately, the table prints the prior for 'settled in a way we would see' as 0.27 and claim 1 repeats '27% of prior mass'; integrate3_results.json has 0.309, and the printed prior column sums to 0.96. The SSA entry for the same row is 0.016, printed as 0.01.
  • Proposed change: Reword as conditional on the structure, correct 0.27 to 0.31 and 0.01 to 0.02, and add a column giving each class's mean likelihood from the local term alone.
  • Expected effect: No change to base numbers. Removes the impression that the equal treatment of the survivors is something the data showed.
  • Tested: Mean likelihood by class from 3M runs: base unnoticed 0.603; per-origin 0.103; lineage mix with common cause 0.05: 0.466 (survivors are then mostly common-cause worlds). Prior column read from integrate3_results.json base.prior.
• moderate / Draft claim 10 and run3.py FUTURE; model3.py lines 214-215 (artefact_null)
  • Argument: The list omits the observation that moves the expansion side most: evidence on whether self-sustaining interstellar settlement is feasible, which we will produce ourselves (closed self-replicating industry, an interstellar precursor, a self-sufficient off-world settlement). The artefact-survey row rests on an unexplained factor: a null multiplies non-evident settled worlds by 0.5 regardless of how many lineages are present, so '0.19 to 0.10' is that 0.5 restated.
  • Proposed change: Add rows 'expansion shown feasible for us (p_exp > 1e-3)' and 'shown infeasible (p_exp < 1e-8)'. Make the artefact term per lineage, or state the 0.5 as an assumption with a range.
  • Expected effect: Base model, feasibility shown: stuck 0.25 to 0.05, unnoticed 0.19 to 0.31, none-ever 0.53 to 0.62 (site). With my preferred structure: none-ever 0.61 to 0.78 (site), 0.42 to 0.62 (SSA). Barrier shown: stuck 0.58-0.63, none-ever 0.35-0.37. This lever is larger than the artefact survey and comparable to a second genesis.
  • Tested: Conditioning 1M-sample posteriors on p_exp thresholds (diag3.py). Base site, p_exp > 1e-6: 0.234/0.326/0.042/0.092/0.276/0.025/0.005. The artefact term itself was not re-implemented.
• minor / Draft claim 11, SIA column: 'they exist; expansion fails or is not evident, with the split set by priors'
  • Argument: Correct but understated on two counts. First, the 0.59 for 'not evident' needs a median of 4.8 million independent Milky-Way-born expander lineages, all low-footprint; with per-lineage evidentness SIA gives stuck 0.60-0.75, hidden 0.10-0.12, unnoticed 0.06-0.25. Second, the split is not merely 'set by priors', it is the priors: with atoms 0.15 (barrier), 0.05 (common-cause low footprint), 0.02 (stealth) the SIA posterior is 0.678 / 0.215 / 0.086, and 0.15 : 0.05 : 0.02 normalises to 0.68 : 0.23 : 0.09. SIA plus non-exclusivity is the 'filter is ahead' result: P(humanity's p_exp < 1e-6) is 0.70-0.89 and the median is 1e-12 to 1e-17.
  • Proposed change: State the SIA result as 'civilisations are common and a universal cause, most likely a barrier that will also stop us, prevents evident expansion', give the atom ratio, and report P(p_exp < 1e-6) under SIA.
  • Expected effect: Wording only at base. SIA unnoticed 0.59 to 0.06-0.25 and stuck 0.33 to 0.60-0.75 if the per-lineage fix is made.
  • Tested: 3M samples; SIA ESS 1,900-2,600, so about +/-0.02. Per-origin: stuck 0.754, P(p_exp < 1e-6) 0.89. Preferred package: 0.678 / 0.215 / 0.086, P(p_exp < 1e-6) 0.70. Crossed with the biology-optimist camp (ESS 42,000): 0.680 / 0.203 / 0.092.
• minor / 'Stuck' has no observable consequence in the model (round-1 expansion-sceptic objection 10, not fixed)
  • Argument: I tried to give it one from the project's data and failed, which should be reported. Non-expanding civilisations in systems that pass within 1 or 10 pc could still hop here; the flyby census bounds that. But civilisation lifetimes in the model are short relative to the interval between passages, so the expected number of passages by a living civilisation is tiny, and the term changes nothing. The published Dyson-swarm fractions would bite only if N_now times the swarm-building fraction exceeded about a million. So the barrier part of 'stuck' can be tested only through our own engineering, and 'stuck' with many long-lived stay-at-home civilisations is the classic SETI scenario that the search record does not constrain.
  • Proposed change: Say this explicitly in the report next to the 'stuck' row, and list what would test it (previous objection).
  • Expected effect: None on the numbers: every class moved by 0.003 or less.
  • Tested: Hop likelihood exp(-N_pass x N_now/N_stars x x_hop): 1 pc with N_pass 2.1e4 and x_hop log-uniform (1e-4,1): site stuck 0.250 to 0.250; 10 pc with 2.1e6 and (1e-6,1): 0.250, SIA 0.311 against 0.312.
• minor / Where my own case is weakest (requested)
  • Argument: (1) A machine lineage with error-corrected replication makes a single lineage's low footprint one design decision copied 1e11 times, not 1e11 choices; so I multiply only over independent origins, and the class survives where there are fewer than about 10 lineages (10% of its mass). (2) 'Nobody evidently here' is strong against planet-claimers and star-harvesters and weak against small dormant probes; artefact searches of the Solar System are crude, so my flyby-based sharpening bounds evident presence only. (3) The human anchor is one case, intent is not a reproduction number above 1, and the one real interstellar programme stalled; my barrier prior of 0.15 is a judgement no better founded than a sceptic's 0.30, and the result is nearly linear in it. (4) Per-lineage independence assumes separately evolved species do not converge on restraint; if they do, that is a common cause and my 0.05 for it is again a judgement. (5) Hart's argument is silent when N x p is below 1: 'a few others, none happened to expand' keeps about 0.05-0.10 under everything I tried. (6) Under SIA my structure does not give 'alone'; it gives a barrier ahead of us.
  • Proposed change: None; these limits should be stated alongside the non-exclusive variant if it is adopted.
  • Expected effect: Bounds what the lens can deliver: P(none other ever) 0.61 (site) and 0.42 (SSA) with my preferred priors, not the near-certainty Hart claimed.

Claims that survive

• Evident, system-claiming expansion is excluded: prior 0.31, mean likelihood 0.004, under every variant I ran.
• Time and distance do not explain the silence (claim 4). The Gaia encounter rate and the head-start census stand, and the per-planet limit can be stated more usefully: (probability per planet) x (fraction of lineages that are evident) < about 1e-9, so even if only 1 lineage in 1,000 were evident the limit is 1e-6.
• Any confident single answer is a statement of priors. This is stronger than the draft says: P(none other ever in the Galaxy) moves almost linearly, 0.48 to 0.71 (site), with two hand-set priors (universal barrier 0.05-0.30, universal low footprint 0.02-0.25).
• Deliberate hiding stays small: 0.02-0.03 in site and SSA under every structure; it rises to 0.09-0.12 under SIA only as a prior ratio.
• Radio SETI adds nothing to the class probabilities, and the galaxy surveys add little once the local term is in (0.59 to 0.61 under my structure; 0.51 to 0.53 in the base).
• Anthropics is the largest lever: SIA sends 'none other ever' to 0.00 in every variant, including mine.
• A soft 'stuck' component is robust and legitimate: worlds with a few to a few thousand other civilisations where N x p < 1 hold 0.05-0.10 under every prior I tried. Hart's argument does not remove them.
• 'Stuck' as a whole could not be pushed below 0.14-0.18 (site) without both narrowing the expansion prior to hard-step width and setting the barrier prior to 0.05.
• Human-anchoring the continuous part of the expansion prior does not change the base numbers (stuck 0.250 to 0.251 at barrier 0.15): the model's tail is numerically equivalent to a barrier prior of about 0.17, so the draft's stuck figure is not inflated relative to that prior, only mislabelled.
• The flyby census does not constrain non-expanding civilisations (hop term tested, no effect), so the draft is right that the data bound willingness and persistence, not existence.

Verdict on each headline claim

• Claim 1: reword
  • Reason: True only because the two surviving late-stage classes were given parameters the data cannot reach (p_here capped at 0.3 regardless of lineage count; an expansion tail with no observable consequence). The prior for the excluded class is 0.31 in the results file, not 0.27.
  • Corrected wording: The silence removes expansion that claims systems and would be evident here (31% of prior mass, likelihood 0.004). Whether it also removes 'settled but not evident' depends on one modelling choice: if non-evidence is a property of the whole universe, nothing else is discriminated by more than a factor of 2; if each independently arising expander lineage chooses its own footprint, 'settled but not evident' is disfavoured by a further factor of about 6, because the surviving worlds contain a median of about 27,000 such lineages. Hypotheses carried by a universal common cause are untouched by the data and are set by their priors.
• Claim 2: reword
  • Reason: (b) and (c) are prior masses passed through: 0.16 of the 0.25 for stuck is the prior tail P(p_exp < 1e-8) = 0.17 times 1.7, and unnoticed is about 0.75 times p_nonevident. 'Roughly equal' reflects the chosen 0.17 and 0.25, and (c) requires thousands of lineages to share one footprint. The range for (a) omits the crossed Hart camp.
  • Corrected wording: Three answers survive, and their shares are set mainly by two priors the data do not touch. (a) Nobody else has ever arisen in the Galaxy: 0.5 to 0.7 (site), 0.3 to 0.55 (SSA), 0 under SIA; up to 0.87 if expansion is assumed easy and obvious. (b) A universal barrier stops sustained expansion, for us too: about 1.7 times whatever prior it is given (0.16 at the model's implicit 0.17; 0.08 to 0.30 across priors of 0.05 to 0.30), plus about 0.09 for 'a few others arose and none happened to expand'. (c) Settled but not evident: close to the prior placed on a universal low-footprint cause (0.19 for 0.25; 0.03 to 0.09 for 0.02 to 0.10) and negligible otherwise. Deliberate hiding stays under 0.05.
• Claim 3: reword
  • Reason: Direction is right, but the size of the local term depends on the evidentness structure, and three SSA rows of the ablation table in the results file are numerically invalid (excluded = 1.000, evidence 2.8e301).
  • Corrected wording: Earth's timeline alone pushes toward a populated Galaxy (P(none other ever) 0.44 to 0.34); the absence of evident settlement here pushes back, to 0.51 if non-evidence is world-level and to 0.59 if it is per lineage; the galaxy surveys add 0.02 to 0.03; radio SETI adds nothing. (Site convention; the SSA ablation rows without the local term must be recomputed.)
• Claim 4: keep
  • Reason: Tried to break it from the other side and could not; it is the strongest part of the draft. Add the evident-fraction form and note that the model's p_here does not yet use these numbers.
  • Corrected wording: Add: the limit applies to the product (per-planet probability) x (fraction of settling lineages that are evident), so it stays below 1e-6 per planet even if only 1 lineage in 1,000 is evident; and in a settled Galaxy, 21,000 passages within 1 pc mean the chance that a passing settled system leaves an evident lasting presence must be below about 2e-5.
• Claim 5: keep
  • Reason: From this lens it supports the absence of harvesting at the 75% level in other galaxies. It says nothing about low-footprint settlement, which the draft already implies; say it explicitly.
• Claim 6: keep
  • Reason: Agrees with the hop-term null I found: the record does not constrain non-expanding civilisations. No objection from this lens.
• Claim 10: reword
  • Reason: The artefact figure is an assumed factor of 0.5 restated, and the list omits the largest expansion-side lever, which is evidence on feasibility from our own engineering.
  • Corrected wording: What would change the answer. A second origin of life or one biosignature among 25 HWO planets: P(none other ever) 0.53 to about 0.3. Evidence that self-sustaining interstellar settlement is feasible for us: 'stuck' 0.25 to 0.05 and P(none other ever) 0.53 to 0.62 (0.61 to 0.78 if evidentness is per lineage); evidence that it is infeasible: 'stuck' to about 0.6. A thorough Solar System artefact survey finding nothing: 'unnoticed' falls by an assumed factor of about 2. More radio SETI: no change.
• Claim 11: reword
  • Reason: The SIA split is exactly the normalised ratio of the prior atoms, and the 0.59 for 'not evident' needs millions of lineages sharing one footprint. With per-lineage evidentness SIA gives a universal barrier.
  • Corrected wording: Anthropics is the largest lever and cannot be argued away. Under SIA emptiness is ruled out and civilisations are common (median of order 1e8 ever per galaxy), so some universal cause must prevent evident expansion. Which cause is not identified: the split is the ratio of the priors on a universal barrier, a universal low footprint and deliberate hiding (0.68 / 0.22 / 0.09 for priors of 0.15 / 0.05 / 0.02). On any of these priors SIA makes it likely (0.4 to 0.9) that humanity's own chance of sustained expansion is below one in a million.

Code errors

• models/model3.py / lines 183-185 and 197 (regime draw, p_here, L_local)
  • Problem: Evidentness is one draw per universe and L_local = 1 - (1 - exp(-mu_arr)) * p_here saturates at 1 - p_here >= 0.7 however many independent expander lineages arrive. Posterior 'unnoticed' worlds have a median of 27,000 arriving lineages.
  • Fix: Per-lineage form: L_local = exp(-mu_arr * c) * exp(-mu_arr * (1 - c) * p_here) with c the claiming fraction (1 in pure claiming worlds, 0 only under an explicit common-cause prior), and classify by mu_arr * c and mu_arr * (1 - c). Working patch in make_x.py (here_mode 'per_origin' and 'mix').
  • Impact: Unnoticed 0.19 to 0.04-0.06 (site), 0.28 to 0.05-0.09 (SSA), 0.59 to 0.06-0.25 (SIA); none-ever 0.53 to 0.61-0.62.
• models/model3.py / lines 225-236 (D_all and share_G), as used by run3.py ABLATION
  • Problem: share_G = W[0] / max(D_all, 1e-300). In claiming worlds where pre-emption drives D_all to zero the ratio is about 1e303. With use_local on, L_local underflows and hides it; with use_local off nothing cancels it. 1.1% of samples then carry 100% of the SSA weight.
  • Fix: Apply the survival factor S(T0) in the SSA numerator independently of use_local, or compute the ratio in logs and floor it consistently; mark SSA as undefined for ablations without the local term.
  • Impact: Three SSA rows in integrate3_results.json ablation ('Prior (no observations)', '+ Earth's timeline only', '+ galaxy surveys (no local argument)') read excluded = 1.000 with evidence 2.8e301 and ESS about 5,000. Site rows and the full-data SSA row are unaffected.
• report/FINDINGS_DRAFT.md / section B table, last row, and claim 1
  • Problem: Prior for 'settled in a way we would see' is printed as 0.27; integrate3_results.json base.prior.excluded is 0.309. The printed prior column sums to 0.96. The SSA entry 0.016 is printed as 0.01.
  • Fix: Use 0.31 in the table and in claim 1 ('31% of prior mass'); print 0.02 for SSA.
  • Impact: Presentation only.
• models/model3.py / line 59 (W_exp) and lines 172-173
  • Problem: Comment says 'same ignorance structure as the biological links', but hard steps use 2-8 decades and expansion 1-30. One p_exp is shared by all civilisations in a universe, with no scatter and no output for humanity's own probability.
  • Fix: Either W_exp = (2, 8), or an explicit barrier atom plus a bounded continuous part; add between-civilisation scatter; output P(humanity expands) and P(p_exp < 1e-6) overall and by class.
  • Impact: W (2,8): stuck 0.25 to 0.11 (site). Scatter 0.5-3 dex: 0.25 to 0.17.
• models/model3.py / line 178 against line 233
  • Problem: L_local discounts locally born expanders by the fill-time factor phi (mu_loc = N_ever * phi * p_exp), while the SSA survival function uses p_exp * sum(W * F) with no fill delay. The same expander pre-empts instantly in the denominator and after a delay in the numerator.
  • Fix: Use the same lag in both, for example evaluate F at t minus T_fill in the denominator.
  • Impact: Not tested. Direction: SSA denominators slightly too small in claiming worlds, so SSA over-weights them a little (SSA 'excluded' is 0.016 against 0.004 for site).
• models/model3.py / lines 214-215 (extra_obs 'artefact_null')
  • Problem: A null artefact survey multiplies non-evident settled worlds by a fixed 0.5 regardless of the number of lineages, and leaves every other world untouched. The factor has no stated basis.
  • Fix: State it as an assumption with a range, or tie it to the per-lineage structure.
  • Impact: The draft's '0.19 to 0.10' is this factor restated.
• models/run3.py / CAMPS, lines 22-34
  • Problem: 'Expansion is easy' and 'Settlement would be obvious' are run only separately, so mass moves between stuck and unnoticed and the Hart-side range of P(none other ever) is understated.
  • Fix: Add a crossed camp: W_exp (1,4), p_nonevident 0.03, p_stealth 0.003.
  • Impact: Crossed camp gives none-ever 0.87 (site), 0.74 (SSA), against 0.62 and 0.64 for the two alone.

Missing explanations

• A common-cause barrier is one class in the model but three different claims: physical infeasibility of self-sustaining settlement, universal extinction before capability, and universal economic non-viability. They predict different things (no artefacts anywhere; ruins and dead technosignatures; long-lived stay-at-home civilisations) and different futures for us. The report should separate them and say which the 0.16 is meant to be.
• Pre-emption or suppression by the first expander (Brin's deadly-probes family, which he judged non-exclusive; cited from memory). The model lets claiming expanders truncate later births only in the SSA denominator; there is no class for 'a first lineage prevents others from arising or expanding', which is a single-cause explanation and the mirror image of the zoo.
• Correlated timing: a galaxy-wide or cosmic common cause that makes nobody much older than us (Annis; Cirkovic and Vukotic, as cited in the behavioural brief). Round 1 flagged it; the cosmic birth-date law is now in, but there is still no variable for a regulation mechanism, which is the only way 'we are first' becomes a common-cause explanation.
• Mixed-lineage galaxies. The model has no world in which some lineages claim and others do not; every universe is all one kind. That is the hypothesis the non-exclusivity argument is about, and it is the one my 'mix' variant adds.
• Frontier selection. Among competing lineages the faster-reproducing, more resource-using variant takes the frontier, and non-claiming lineages cannot hold territory against it (Hanson's 'Burning the Cosmic Commons', cited from memory). The model has no dynamics between lineages at all.
• Came and went. A past wave that filled the Galaxy and vanished is folded into 'unnoticed' with the same p_here; with thousands of lineages it needs every one to vanish, so it should be costed per lineage too.

Critic’s own estimate of each answer

• Empty universe: 0.24
• Rare: 0.37
• En route: 0.05
• Stuck: 0.25
• Unnoticed: 0.05
• Hidden: 0.03
• Other: 0.01
• What “other” covers: Outside the taxonomy: pre-emption by a first expander, correlated timing, simulation, model misspecification. Basis: average of my preferred and hawkish packages, mostly the site convention with some SSA weight and little SIA. Of the 0.25 for stuck, about 0.16 is a universal barrier that would also stop us and about 0.09 is 'a few others, none happened to expand'.

Were the round 1 fixes made?

Round 1 had no critic on this side, so I checked the adjacent ones. Fixed: expanders from other galaxies and an 'en route' class (Hansonian 1, 9); the floors on will and harvesting removed (sceptic 1, SETI optimist 2); an explicit evidentness term and an 'unnoticed' class (sceptic 2, optimist 3); a per-term ablation table; camps that change one lever each (sceptic 8c). Not fixed: (i) 'the split among universal-barrier classes is input, not output' (sceptic 4, Hansonian 6, optimist 9): the three atoms were replaced by a tail mass P(p_exp < 1e-8) = 0.17 and a regime coin of 0.25, and the stuck barrier part (0.161) is still prior times likelihood over evidence, with the SIA column exactly the atom ratio; (ii) Hansonian 7, an output for humanity's own prospects, is absent, and the prior moved further from the human-calibrated one in the longevity brief; (iii) sceptic 10, the barrier class has no observable consequence, still true, and my attempt to add one from the flyby data had no effect; (iv) the camps are no longer confounded with biology, but the two Hart-side levers are never crossed. Over-corrected: (a) the will floor of 1e-4 became a 30-decade prior; the sceptic asked for a range 'mirroring hard_step_decades', which in version 3 is 2-8, so expansion is now wider than any hard step; (b) low-footprint settlement was given 0.25 with the evident chance capped at 0.3, against 0.08 (0.01-0.30) in the project's behavioural brief, and as a world-level coin where that brief prescribes (1 - e)^N plus a common-cause Bernoulli; (c) the round-1 sceptic's own test showed that scatter between civilisations shrinks the barrier classes, and version 3 dropped scatter; (d) 'restraint is crushed' was replaced by 'nothing but evident expansion is discriminated', which is the opposite overstatement.

Presentation advice

Show the reader the three priors that carry the late-stage result as a prior-to-posterior table: universal barrier (0.17 implicit, posterior part of stuck 0.16), universal low footprint (0.25, posterior 0.19), deliberate hiding (0.02, posterior 0.02), for site, SSA and SIA, and say that the SIA column is their normalised ratio. Add one two-way table or figure: barrier prior (0.05-0.30) against low-footprint prior (0.02-0.25), shaded by P(none other ever); from my grid it runs 0.48 to 0.71 (site) and 0.29 to 0.53 (SSA). Split 'stuck' into three rows (universal barrier 0.161, many civilisations without a barrier 0.044, few civilisations 0.046) and rename the first 'a universal barrier stops sustained expansion, for us too'. Next to 'unnoticed', print how many independent expander lineages the class contains (median 27,000; 4.8 million Milky-Way-born under SIA) and the one-line requirement: fewer than about 1 lineage in 100,000 may be of the evident kind. Report P(humanity becomes a sustained expander) and P(p_exp < 1e-6) overall and by class; a reader who sees 'stuck 0.25' should be told it carries a median of 5e-11 for us. State the head-start limit in the form (per-planet probability) x (evident fraction) < 1e-9, and the flyby result as 'in a settled Galaxy the chance that a passing settled system leaves an evident presence must be below 2e-5'. Present the per-lineage variant as a second column beside the base table, not as a footnote; the table I would print is: preferred (per-lineage evidentness, low-footprint common cause 0.05, stealth 0.02, barrier 0.15, otherwise log-uniform 1e-5 to 1) site 0.228 / 0.358 / 0.043 / 0.290 / 0.055 / 0.020 / 0.005, none-ever 0.614; SSA 0.151 / 0.243 / 0.047 / 0.407 / 0.093 / 0.034 / 0.025, none-ever 0.419; SIA 0 / 0 / 0.021 / 0.678 / 0.215 / 0.086 / 0.001. Hawkish (0.02, barrier 0.05, log-uniform 1e-4 to 1, claiming favoured): site 0.290 / 0.414 / 0.050 / 0.179 / 0.030 / 0.030 / 0.008, none-ever 0.740; SSA none-ever 0.569. Concessive (0.10, barrier 0.30, log-uniform 1e-6 to 1): site 0.182 / 0.311 / 0.035 / 0.394 / 0.062 / 0.013 / 0.004, none-ever 0.512; SSA 0.318. Crossed with the biology-optimist camp the preferred package gives none-ever 0.12 and stuck 0.62, of which 0.57 is the barrier, which is the clearest way to show that optimistic biology plus non-exclusivity means a filter ahead. Correct the 0.27 in the base table to 0.31.

Critic: an expansion sceptic

Overall verdict

Version 3 delivered most of what this lens asked for in round 1, and under the site convention the three-family picture is defensible: P(no other civilisation ever in the Galaxy) stayed between 0.47 and 0.63 in every variant I ran. But the draft overstates what the data and the original studies do. The posterior is the prior with one class deleted, then tilted by Earth's timeline: that two-step calculation reproduces every class to within 0.02, the Tully-Fisher and fundamental-plane limits carry exactly zero weight (removing them or making them 100 times stronger changes nothing at three decimals), and the fill-time and head-start inputs move no class by more than 0.012. Two structural choices carry the headline split and both are over-corrections toward my own camp. First, 'not evident' is a universe-wide rule with prior 0.25; allowing a minority of expanders to be of the claiming kind, with the same ignorance prior used for expansion itself, cuts 'unnoticed' from 0.19 to 0.075 (site) and from 0.59 to 0.23 (SIA). Second, claiming settlement is assumed evident with probability exactly 1; under SSA a leak of one unclaimed site in a million brings the 'contradicted' class back at 0.12 to 0.33. Claim 4 quotes numbers its own referee withdrew and omits its conditions, and claim 5 calls limits 'first' that Annis 1999 and Zackrisson 2015 already had. The single most important change is to replace the two universe-level switches (claiming or non-evident; evident with certainty) with continuous quantities, a claiming fraction and a pocket fraction, with the SSA reference class made consistent, and to present 'stuck', 'unnoticed' and 'hidden' as one prior-split family.

Objections

• major / model3.py line 185 (p_here = 1.0 in the claiming regime) with lines 225-236 (SSA denominator D_all); SSA column of table B; claims 1 and 2
  • Argument: The claiming regime is evident with probability exactly 1, so a claimed Galaxy has no unclaimed pockets at all. Landis voids, Carroll-Nellenback steady states (their P of about 0.1 that Earth is unvisited) and settlers skipping G dwarfs are therefore representable only as the separate 'low-footprint' regime. Under the site convention a small leak is harmless. Under SSA it is decisive: in a claimed universe nearly every civilisation that ever exists arises in a pocket and sees what we see, so the likelihood of our observation saturates once the pocket fraction exceeds about 1/mu. The model's D_all truncates all births after the first expander and counts no pocket civilisations, so the SSA column is valid only for a pocket fraction of exactly zero. Change p_here to U(0.99,1) without fixing D_all and the SSA column collapses to excluded = 1.000 (ESS about 2,000).
  • Proposed change: Give the claiming regime a pocket fraction (log-uniform 1e-6 to 0.1 is what I would defend) and add (1 - p_here) x (N_alltime - D_trunc) to D_all for claiming worlds. Relabel class 7 'claimed Galaxy; we are in an unclaimed pocket' and report it. State in claim 1 that the exclusion is a site-convention result.
  • Expected effect: Site column unchanged (class 7 goes 0.004 to 0.007). SSA column: class 7 goes 0.016 to 0.12-0.33; P(none other ever) 0.33 to 0.23-0.30; stuck 0.32 to 0.21-0.29; unnoticed 0.28 to 0.18-0.25. SIA: class 7 0.003 to 0.014.
  • Tested: Patched copy (reproduces base to 4 decimals), n = 1e6. Pockets counted in D_all: p_here U(0.99,1) gives SSA class 7 = 0.325, none-ever 0.233; U(0.9,1) gives 0.352; leak LU(1e-6,1e-2) gives 0.263; leak LU(1e-12,1e-6) gives 0.120; leak LU(1e-6,0.1) gives SSA row 0.087/0.139/0.034/0.235/0.204/0.019/0.282. Same leak without the denominator fix: SSA class 7 = 1.000, ESS 567.
• major / model3.py line 184: regime drawn once per universe (p_nonevident = 0.25, p_stealth = 0.02); class 'unnoticed'; claims 2(c) and 11
  • Argument: Round 1 asked for a 'settled but unnoticed' class. Version 3 supplies it as a universal law: in 25% of universes every sustained expander, in every galaxy, for all time, is low-footprint. That is the monocultural assumption in my own camp's favour. The class then pays almost nothing: mean p_here in it is 0.047, only 2.9% of it is in the harvesting branch, and its posterior/prior ratio (1.78) equals that of 'stuck' (1.69) and 'hidden' (2.0). So 0.19 is 0.25 passed through, exactly as 'hidden' is 0.02 passed through; the draft says this of hiding but not of 'unnoticed'. Under SIA the class median has about 4e7 expander origins, none of the claiming kind. Also, 17% of the class (site) is a Galaxy reached only by expanders from other galaxies, and the label says 'or visited' although visits are not modelled.
  • Proposed change: Keep p_nonevident as the probability that low-footprint is the norm, but let a fraction q_c of expanders be of the claiming kind with log10 q_c ~ U(-W2, 0), W2 log-uniform 2 to 40 (the same structure as p_exp). Multiply the local likelihood by exp(-mu x q_c). Report 'unnoticed' with and without this term, and say it is prior-determined.
  • Expected effect: Site: unnoticed 0.19 to 0.075 (W2 1-30), 0.12 (W2 4-40) or 0.03 (W2 1-6); P(none other ever) 0.53 to 0.58-0.63; stuck 0.25 to 0.27-0.30. SSA: unnoticed 0.28 to 0.11-0.17. SIA: unnoticed 0.59 to 0.23-0.36, stuck 0.33 to 0.52-0.62, hidden 0.05 to 0.09-0.10.
  • Tested: claim_mix option, n = 1e6 (site, SSA) and 3e6 (SIA). W2 LU(1,30): site 0.225/0.346/0.046/0.284/0.075/0.019/0.004, none-ever 0.600; SSA unnoticed 0.110; SIA 0.619 stuck / 0.228 unnoticed / 0.102 hidden (ESS 2,550). W2 LU(4,40): site unnoticed 0.116, SIA 0.355. W2 LU(1,6): site 0.029. Three seeds at n = 4e5 agree to 0.002. Also p_nonevident 0.1 / 0.5: unnoticed 0.086 / 0.310.
• major / Draft claim 4 (wording, numbers and conditions)
  • Argument: (a) The numbers are the pre-referee ones. The headstart referee wrote that '72.5% (69 to 76%)' does not survive; the revised report gives 0.70 with systematic range 0.64-0.78 and mean head start 3.5-4.4 Gyr. The flyby referee asks for 11 +- 1 per Myr, the range 12,000-57,000 beside 21,000, and the 0.47-0.67 Gyr eras (4,200-6,000 systems) because Proterozoic oxygen may not have been remotely detectable. (b) 'Below about 1e-9' is the fiducial-count figure; the study's own marginalised limit is 3.4e-9 to 7.9e-9. (c) Conditions are missing. 'Any speed from 1e-4 c' holds only if the front is transport-limited: with 30 km/s ships that needs hops of 1 pc or more and a settleable fraction above a few percent (flyby referee: f = 0.03 with the rendezvous cap gives 0.7 Gyr at 1 pc and 2.7 Gyr at 0.5 pc; below 0.14 pc the Galaxy is not filled in 1 Gyr). The limit also needs settlement to persist for gigayears (x < 0.27 for a 1 Myr presence, against 1.4e-4), to leave no pockets, and an Earth-clock emergence model. (d) It is a limit on the whole planet-to-expander chain, which a reader will take as a limit on willingness to expand. (e) 'Original data' oversells: the referees class both studies as reproductions (Lineweaver 2001, Legassick 2015; Zuckerman 2026 for the passing-star count). (f) Neither study moves the posterior.
  • Proposed change: Rewrite with the refereed numbers and explicit conditions (wording under claim verdicts). Add one sentence: in the model, a fill time of 5 Gyr for every world changes no class by more than 0.012.
  • Expected effect: No change to the table. The limit becomes 'about 1e-9 for the fiducial planet count, 3e-9 to 8e-9 marginalised, and only for claiming, persistent, space-filling settlement', a regime with posterior weight 0.56 in the model.
  • Tested: T_fill = 5 Gyr for all worlds: stuck 0.250 to 0.260, none-ever 0.532 to 0.527 (SSA stuck 0.323 to 0.338). T_fill x LU(1,30) in 30% / 100% of worlds: stuck 0.254 / 0.262. Hard-step check from cosmo_tables: of civilisations on Sun-like hosts arisen by now, 93-96% arose more than 1 Gyr ago for 0 to 6 hard steps, so late births do not rescue 'not enough time'. Model posterior in the claiming regime: 95% upper quantile of expander origins per habitable-zone rocky planet 7e-12 (site).
• major / model3.py lines 59 and 172-173: one p_exp prior (W log-uniform 1-30) for every regime; class 'stuck'; claim 2(b)
  • Argument: Round 1 asked for feasibility split by mode because feasibility and evidentness are anticorrelated: settlers who claim every system are the kind that would be evident and the kind least likely to sustain 36 doublings and gigayear persistence; sparse machine presence is the reverse. Version 3 draws p_exp independently of regime, so both get prior median 0.7%, P(p_exp > 0.1) = 0.28 and P(p_exp > 0.01) = 0.48. For claiming settlement that is the optimist corner. In the other direction, 'stuck' is the deep tail of the same prior: in the class the median p_exp is 5e-11, a quarter of it lies below 1e-15, and 76% below 1e-6; under SIA the median world has 4e8 civilisations per galaxy and p_exp 1e-15. That is a common-cause barrier in everything but name, with effective prior P(p_exp < 1/N) = 0.52, 0.34, 0.24, 0.13 for N = 1e2, 1e4, 1e6, 1e10. The label 'almost never happens' also covers 8% few-civilisation worlds with ordinary propensity and 10% with expected expander origins of 0.1-1. The class share is set by the tail depth alone.
  • Proposed change: Make the prior regime-specific. I would defend: claiming regime log10 p ~ U(-0.5 - W, -0.5) with W log-uniform 2-40; non-evident regime W log-uniform 1-10. Describe 'stuck' as 'a shared barrier: fewer than one civilisation in N sustains expansion' and print the tail-depth row (W_max 8, 12, 20, 30, 60) beside the headline.
  • Expected effect: Regime-specific priors: stuck 0.25 to 0.26-0.34, unnoticed 0.19 to 0.16-0.21, none-ever 0.53 to 0.47-0.50. Tail depth alone: stuck 0.08 (W_max 8), 0.12 (12), 0.19 (20), 0.25 (30), 0.33 (60), with unnoticed moving the other way from 0.29 to 0.15.
  • Tested: n = 1e6. Claim W LU(4,40) with non-evident W LU(1,6): site stuck 0.294, unnoticed 0.208, none-ever 0.475 (SIA 0.399 / 0.532). Claim W LU(4,40) only: stuck 0.337, unnoticed 0.158. Non-evident W LU(1,6) only: stuck 0.198, unnoticed 0.247. Claim W LU(2,30) capped at 0.1 with non-evident LU(1,10): 0.262 / 0.208. Stuck-class composition from a 2e6 keep run (diag1.py).
• major / model3.py lines 72-73 (warm_curve, cold_curve), line 198; claim 5; the survey share in claim 3
  • Argument: (a) warm_curve is mis-sourced. Its 0.85 to 1e6 point and its points below 50% come from Lacki 2016, a Planck search for galaxies shrouded at just above the CMB temperature (abstract re-read). The model applies them in the 100-600 K branch. G-HAT (Griffith et al. 2015) supports about 1e5 galaxies at 85% or more, about 5e3 between 50% and 85% (50 natural-looking sources above 50%), and nothing below 50%, where ordinary dust re-radiates that much. Because f_cov is log-uniform 0.01-1, 85% of harvesting worlds sit below 50%, so the whole survey effect comes from the unsupported part of the curve. (b) cold_curve adds bright early types to HI-rich discs (1,200 = 3/0.0029 + 3/0.018), uses the pre-referee Tully-Fisher value, treats searches with 16 and about 20 unresolved outliers as zero-count samples, and is tagged 'any temperature' although the Tully-Fisher part fails above about 300 K. None of that matters, because the curve has zero weight. (c) Claim 5 is wrong in places. 'First limits that do not depend on waste-heat temperature': Annis 1999 and Zackrisson et al. 2015 already had them (tf_kiii referee). 'About 1e5 early types' and '9,400 discs': the effective samples at 75% dimming are 14,853 and about 1,400, with V >= 120 km/s only. 'No star-harvested galaxies': 5 fundamental-plane outliers at 75% or more and 20-22 Tully-Fisher candidates are unexplained. The limits assume grey dimming of light; enclosing every dwarf star removes only 20-40% of the light, where the limit is 7-17% or absent.
  • Proposed change: Use three temperature bands: warm 100-600 K ((0.5, 4.8e3), (0.85, 1e5)); intermediate, scaling relations only, for discs ((0.75, 130), (0.9, 214)); very cold (below about 20 K) Lacki's curve. Reword claim 5 (wording under claim verdicts) and state that these limits do not move the posterior.
  • Expected effect: Survey contribution to P(none other ever) falls from +0.016 to +0.003; stuck 0.250 to 0.262; en route 0.041 to 0.047. Nothing else moves.
  • Tested: Surveys off: none-ever 0.516. Warm curve restricted to 50% and above: 0.519. Warm curve removed: 0.517. Warm 0.85 point set to 1e5 only: 0.532 (no change). Cold curve removed: 0.531, every class within 0.001. Cold curve x100: 0.532, no change. Discs-only referee cold curve: no change. p_warm 0.3: 0.524. p_harvest 0.9 / 0.5: 0.578, unnoticed 0.149. In the excluded class the mean survey likelihood is 0.63 and only 37% of it has survey likelihood below 0.05.
• moderate / Claims 1 and 2: arithmetic, attribution, and the quoted ranges
  • Argument: (a) The prior column of table B sums to 0.96. The excluded class has prior 0.309 in integrate3_results.json, not 0.27; '27% of prior mass' should read 31%. (b) 'Discriminate among everything else by less than a factor of 2' credits the silence with what Earth's timeline does. Prior-weighted mean of the silence terms (local x survey x SETI) is 1.00 empty, 0.93 rare, 0.83 stuck, 0.87 unnoticed, 1.00 hidden, 0.66 en route; the factor-2 spread comes from the early-life term (0.27 for empty against 0.60 for hidden). (c) The posterior is a two-step calculation: delete the excluded class, multiply by the timeline. (d) 'Roughly equal' does not describe 0.53 / 0.25 / 0.19. (e) The ranges in claim 2 span anthropic conventions only. The project's own 400-draw scan gives site 5-95% ranges of 0.26-0.74 (none ever), 0.06-0.36 (stuck), 0.07-0.56 (unnoticed); base 'stuck' is at the scan's 80th percentile and base 'unnoticed' at its 32nd. (f) The draft never reports P(another civilisation in the Galaxy now): 0.22 site, 0.39 SSA, 0.98 SIA.
  • Proposed change: Fix the prior column. Lead with the decomposition row (prior; minus excluded; times timeline; all terms). Reword claims 1 and 2 as given under claim verdicts. Quote scan ranges beside the base numbers and add the 'now' probability.
  • Expected effect: No change to the model. P(none other ever) reads 0.44 prior, 0.61 after deleting the contradicted class, 0.505 after the timeline, 0.531 with all terms.
  • Tested: 2e6 keep run: prior minus excluded times timeline gives 0.176/0.293/0.055/0.267/0.194/0.016 against the full 0.197/0.303/0.041/0.249/0.188/0.017. Scan percentiles computed from integrate3_scan.json.
• moderate / model3.py line 178 (mu_loc = N_ever x phi x p_exp, permanent once filled); definition of 'sustained, persistent expander'
  • Argument: Round 1 asked for finite persistence. Version 3 defines it away: 'persistent' is part of the definition of p_exp, and every expander with enough head start is present and evident today and for ever, in the local term and in the survey term. A wave that came and went is therefore not a class. It sits unseen inside 'stuck', whose label says expansion never happened. The same holds for bounded bubbles and for a wave still in progress, which is 1.3% of 'stuck' because T_fill is capped at 1 Gyr. Whether a departed claiming wave would have pre-empted us or left traces is a separate unknown the model should carry.
  • Proposed change: Sample a persistence time T_p (log-uniform 1 Myr to 10 Gyr); count as evident only waves with head start less than T_fill + T_p, plus a probability p_trace (I would use 0.3, range 0-0.5) that a departed wave pre-empted us or left evidence. Add the class 'a wave came through and is gone'. The SSA reference class needs matching treatment.
  • Expected effect: Site: new class 0.07-0.10 if departed waves leave no trace, 0.01 at p_trace = 0.5; P(none other ever) 0.53 to 0.48-0.50; other classes shrink in proportion. With my package (p_trace 0.3) the class is 0.015.
  • Tested: persist option, n = 1e6, site column. LU(1 Myr, 10 Gyr), no trace: class 0.067, none-ever 0.500. LU(10 kyr, 10 Gyr): 0.103. LU(1 Myr, 1 Gyr): 0.082. p_trace 0.5: 0.009. The SSA column is not usable for these runs (ESS 2 to 4,000) because D_all assumes permanent pre-emption.
• moderate / Claim 10 and model3.py lines 212-215 (techno_null_1e6, artefact_null)
  • Argument: (a) 'Unnoticed 0.19 to 0.10' is the assumed 50% survey efficiency, not a result: 0.188 x 0.5 / (1 - 0.094) = 0.104. The efficiency is also independent of p_here, so a presence that had a 0.1% chance of being evident is found half the time. (b) 'More radio SETI: no change' is true only of the survey modelled, a million stars. An Earth-level null for the whole Galaxy is among the most informative observations available to the model. (c) The two futures are inconsistent: the artefact survey can see settlers, but the technosignature survey of the nearest million stars counts only quiet civilisations, although a settled Galaxy has occupied systems among those stars. (d) Resolving the Tully-Fisher and fundamental-plane candidates is not listed and would change nothing.
  • Proposed change: State the artefact result as conditional on the assumed efficiency and tie it to p_here. Add the Galaxy-wide technosignature census as a future observation and let nearby-star surveys act on the non-evident regime. Say that resolving the scaling-relation candidates would not move the posterior.
  • Expected effect: Galaxy-wide Earth-level null: P(none other ever) 0.53 to 0.68, stuck 0.25 to 0.19, unnoticed 0.19 to 0.11, P(another civilisation now) 0.22 to 0.02 (SSA none-ever 0.33 to 0.54). Million-star null that also sees settled systems at 50%: unnoticed 0.19 to 0.10.
  • Tested: extra_obs techno_null_galaxy: site 0.254/0.390/0.042/0.192/0.109/0.010/0.003; SIA P(now) 0.98 to 0.40. techno_null_1e6_settled: unnoticed 0.102. Model's own artefact_null: 0.105.
• moderate / p_exp as a single number per universe (the question put to this lens)
  • Argument: Missing between-civilisation scatter is not an error by itself. Each civilisation already expands independently with probability p_exp, and for a Poisson count only the population mean matters, so scatter at fixed mean changes nothing. It matters through what the prior is taken to describe. If W log-uniform 1-30 is a prior on a typical civilisation's propensity, the mean is dominated by outliers (a factor 10^(1.15 s^2) for s dex of scatter) and 'it only takes one' gains force. This cuts against my camp and should be shown.
  • Proposed change: State that p_exp is the population mean over all civilisations and factions, not the typical value. Show one row with the prior read as a median with 0.5-1.5 dex scatter.
  • Expected effect: Prior read as median with 0.5-1.5 dex scatter: stuck 0.25 to 0.215, unnoticed 0.19 to 0.21, none-ever 0.53 to 0.55. With 1-2.5 dex: stuck 0.167, unnoticed 0.243, none-ever 0.566. The deep tail is untouched, so 'stuck' never falls below about 0.15.
  • Tested: scatter option (mean of a lognormal capped at 1), n = 1e6: 0.207/0.304/0.039/0.215/0.212/0.019 and 0.219/0.307/0.039/0.167/0.243/0.022 (site); SSA stuck 0.280 and 0.219.
• minor / model3.py lines 62-63, 176-178: T_fill = max(0.3, 1e-4/v) and the head-start table
  • Argument: (a) Double count. N_ever already counts only civilisations arisen by now; phi(0) = 0.72 is the fraction of Sun-like stars older than the Sun, so multiplying N_ever by phi applies the 'older than us' cut twice. The right factor is phi(T)/phi(0): 0.96 at 0.3 Gyr, 0.875 at 1 Gyr. (b) The table is star ages on an Earth clock, while the same model's birth-date tables give 0.94-0.98 at 0.3-1 Gyr for Sun-like hosts and 0.73-0.92 for long-window hosts with three hard steps; the local and cosmic terms use different birth laws. (c) The table is the LAMOST primary curve (0.725), not the adopted 0.70 (0.64-0.78). (d) T_fill has no hop-range or settleable-fraction dependence and is capped at 1 Gyr, so the round-1 request was only partly met. All four are immaterial in version 3.
  • Proposed change: Use the model's own F tables per host class at T0 - T_fill for mu_loc, or at least divide phi by phi(0). Drop the claim that the head-start table is a model input that matters, and say the fill time is immaterial.
  • Expected effect: mu_loc rises by a factor 1.1-1.5. Stuck 0.250 to 0.246, unnoticed 0.188 to 0.191, SSA excluded 0.016 to 0.012.
  • Tested: phi_norm and phi_tables options: identical results, 0.198/0.304/0.040/0.246/0.191/0.017. Fill-time variants as in the claim-4 objection.
• minor / model3.py line 78 (c_det = 1e-12 to 1e-5), line 199 (N_now); claim 6
  • Argument: The range matches seti_meta for one behaviour at a time: 1e-11 is the generous leakage bound (0.72 detections if every star hosted one) and 6.9e-6 is an always-on isotropic 1e17 W beacon in 1-10 GHz. As a prior on a population it assumes fewer than about 2e-4 of civilisations run Kardashev-II beacons (detection chance 0.05 in 1-10 GHz, up to 0.5 on the 21-cm line) and excludes the literal-leakage and radar cases (1e-17 or zero). The term is inert either way, which is the SETI optimist's own position. Claim 6's '100 to 2,000' for Kardashev-II beacons is the referee's phrasing; the revised study gives 113 (55 to 18,000), with 2,000-9,000 from post-2015 searches alone.
  • Proposed change: Replace c_det by a mixture: beacon fraction log-uniform 1e-4 to 0.3 with detection chance 1e-6 to 0.5, else 1e-17 to 1e-11. In claim 6 write 'roughly 100 to 10,000'.
  • Expected effect: Class changes under 0.01 unless nearly everyone beacons. The interpretation stays: the radio record constrains behaviour.
  • Tested: c_det 1e-17 to 1e-11: none-ever 0.530. 1e-8 to 1e-2: 0.550, stuck 0.245. 1e-5 to 0.5 (everyone runs a 1e17 W beacon or stronger): 0.589, stuck 0.232, P(now) 0.14.
• minor / Class 'stuck' has no observational consequence; visible-domain rule 'N_eff >= 100' (line 198)
  • Argument: Round 1 objected that the barrier class could not be tested. I checked whether the project's own flyby census tests it: 21,000 systems passed within 1 pc, so quiet civilisations that hop once and leave a lasting presence are bounded. It does not bite. I also checked the rule that any expander domain whose visible edge lies inside the survey sphere counts as detected whenever N_eff >= 100; it ignores how much of the survey volume the domain overlaps, but it is immaterial.
  • Proposed change: None needed to the numbers. Say that 'stuck' is tested only by a Galaxy-wide technosignature census, and replace the threshold rule with an overlap weight for tidiness.
  • Expected effect: None on the table.
  • Tested: Hop term exp(-2.1e4 x N_now/N_stars x y): y log-uniform 1e-4 to 1 and y = 1 both leave every class within 0.001. Visible-domain rule off: en route 0.041 to 0.044. Overlap-weighted rule: identical at 3 decimals.

Claims that survive

• Under the site convention, 'no other civilisation has ever arisen in the Galaxy' remains the largest single family in every variant I ran: 0.47 (claiming settlement made much harder) to 0.63 (claiming minority allowed), against 0.53 at base.
• Time is not the binding constraint inside this model. A 5 Gyr fill time for every world moves 'stuck' from 0.250 to 0.260; replacing the star-age table with the model's own birth-date tables changes nothing; 93-96% of civilisations on Sun-like hosts that have arisen by now arose more than 1 Gyr ago for zero to six hard steps.
• Radio SETI carries no weight under any detection prior from 1e-17 to 1e-2 per civilisation (P(none other ever) 0.530-0.550). The technosignature record constrains behaviour, as claim 6 says.
• The galaxy-survey term is immaterial to the answer: removing it entirely moves P(none other ever) from 0.532 to 0.516. The visible-domain rule is also immaterial.
• 'Stuck' is not contradicted by the project's own flyby census: a short-hop term on quiet civilisations changes no class by more than 0.001 even if every passing civilisation hops.
• Between-civilisation scatter at a fixed mean propensity has no effect. Read as scatter around a typical value it reduces 'stuck' to 0.17-0.22 but never below about 0.15.
• Deliberate hiding stays between 0.010 and 0.024 (site) unless its prior is raised.
• The ablation numbers in claim 3 match integrate3_results.json for the site convention (0.441, 0.344, 0.514, 0.530, 0.532).
• Under SIA an empty Galaxy is ruled out and 0.93-0.97 of the mass is 'they exist; expansion fails or is not evident' in every variant; only the split moves.
• Monte Carlo noise is negligible for site and SSA (three seeds agree to 0.002); my patched copy reproduces the base table to four decimals.
• The 'Expansion sceptic' camp now changes one lever only, and the ablation table requested in round 1 exists.

Verdict on each headline claim

• Claim 1: reword
  • Reason: Prior mass is 31%, not 27% (table column sums to 0.96). The class is removed by construction (claiming settlement is assumed evident with certainty) and only under the site convention; under SSA any pocket fraction above about 1e-6 restores it at 0.12-0.33. The factor-2 spread among survivors comes from Earth's timeline; the silence terms differ by at most 1.2 among the main classes (1.5 for en route).
  • Corrected wording: Under the site convention the absence of evident settlement in the Solar System removes one class: worlds in which settlers of the claiming, space-filling, persistent kind have reached the Galaxy (31% of prior mass, likelihood 0.004, because such settlement is assumed evident with certainty). Nothing else in the silence discriminates among the remaining classes by more than a factor of about 1.2 (1.5 for 'en route'); the factor-2 spread in their likelihoods comes from Earth's timeline. Any confident single answer is a statement of priors. Under SSA the exclusion holds only if claimed galaxies leave fewer than about one site in a million unclaimed.
• Claim 2: reword
  • Reason: 0.53 / 0.25 / 0.19 is not 'roughly equal'. The quoted ranges span anthropic conventions only; the project's own prior scan is far wider. (b), (c) and hiding are all prior pass-throughs with the same multiplier (1.7-2.0); singling out hiding as 'raised by hand' is asymmetric. (c) depends on a universe-wide monoculture.
  • Corrected wording: Three families remain, and their shares are set by priors. Site convention, base priors: (a) no other civilisation has ever arisen in the Galaxy, 0.53; (b) civilisations arose but a shared barrier keeps any from sustaining expansion, 0.25; (c) the Galaxy is settled in a way not evident from here, 0.19; deliberate hiding, 0.02. Across a 400-draw scan of the priors the 5-95% ranges are 0.26-0.74, 0.06-0.36 and 0.07-0.56. SSA moves about 0.2 from (a) to (b) and (c); SIA removes (a). (b), (c) and hiding are each their prior times the same factor of 1.7-2.0. (c) is 0.19 because low-footprint settlement was given prior 0.25 as a universal rule; it falls to 0.03-0.12 if a minority of expanders may be of the claiming kind. The chance that another civilisation exists in the Galaxy now is 0.22 (site) to 0.39 (SSA).
• Claim 3: reword
  • Reason: Numbers are right for the site convention. The survey step (+0.016) comes entirely from the part of the warm-waste-heat curve taken from a cold-shroud paper; with the curve restricted to what G-HAT measured it is +0.003, and the Tully-Fisher and fundamental-plane limits contribute 0.000. The stored SSA columns for the three ablation rows without the local term are degenerate.
  • Corrected wording: Ablation (site convention). Earth's timeline alone pushes toward a populated Galaxy (P(none other ever) 0.44 to 0.34); the absence of evident settlement here pushes back (to 0.51); the galaxy surveys add at most 0.02, and 0.003 once the infrared limit is restricted to covering fractions G-HAT actually tested; the optical scaling-relation limits and radio SETI add nothing.
• Claim 4: reword
  • Reason: Uses pre-referee numbers (72.5% (69-76%), 11.1 +- 0.5, 1e-9), omits the conditions (transport-limited front, gigayear persistence, no pockets, Earth-clock emergence), presents a whole-chain limit as if it bore on expansion alone, and calls reproductions 'original data'. The first sentence survives inside the model.
  • Corrected wording: Time and distance do not explain the silence, under stated conditions (new measurements reproducing known results). Gaia: star systems pass within 1 pc of the Sun at 11 +- 1 per Myr; about 21,000 (12,000-57,000) have done so since the Great Oxidation, 4,000-6,000 in the last 0.5-0.7 Gyr. Stellar ages: about 70% (64-78%) of Sun-like stars now in the solar cylinder are older than the Sun, by 3.5-4.4 Gyr on average. A settlement front faster than about 3e-5 c, which for 30 km/s ships needs hops of 1 pc or more and more than a few percent of systems settleable, fills the Galaxy in 0.25-1 Gyr, well inside that head start; in the model even a 5 Gyr fill time changes no class by more than 0.012. If settlement is also of the kind that claims every system, persists for gigayears and would be evident here, then fewer than about 1e-9 of habitable-zone rocky planets (3e-9 to 8e-9 once the planet count's uncertainty is included) have produced a Galaxy-settling civilisation. That is a limit on the whole chain from planet to expander, not on the willingness to expand.
• Claim 5: reword
  • Reason: 'First limits that do not depend on waste-heat temperature' is false (Annis 1999; Zackrisson et al. 2015). Sample sizes quoted are the parents, not the effective samples. Unexplained outliers exist (5 and about 20). The Tully-Fisher figure is 2-3% after refereeing and is valid only below about 300 K. The limits are on grey dimming of light. They carry zero weight in the model.
  • Corrected wording: No confirmed star-harvested galaxy in optical scaling-relation searches (original data, extending Annis 1999 and Zackrisson et al. 2015). For grey, galaxy-wide dimming: among the roughly 15,000 bright early-type galaxies that would stay in the sample, fewer than 0.3% (0.24-0.45%; 1.4% under a pessimistic noise model) have lost 75% of their light, with 5 unexplained outliers at that level; among about 1,400 effective HI-rich discs with rotation above 120 km/s, fewer than 2-3%, with about 20 unresolved candidates. The early-type limit holds at any waste-heat temperature, the disc limit below about 300 K. Both are systematics-limited, say little below half the light (enclosing every dwarf star removes only 20-40%), and do not move the integrated posterior.
• Claim 6: keep
  • Reason: Consistent with the revised seti_meta report and with the model, where the term is inert for any detection prior from 1e-17 to 1e-2. One number should be widened: the revised Kardashev-II range in 1-10 GHz is 113 (55 to 18,000), 2,000-9,000 from post-2015 searches alone.
  • Corrected wording: ...Kardashev-II beacons to roughly 100 to 10,000 in 1-10 GHz (from a handful to about 5,000 on the 21-cm line, depending on whether legacy surveys are trusted)...
• Claim 10: reword
  • Reason: The artefact number is the assumed 50% efficiency, by construction. 'More radio SETI: no change' holds only for a million-star survey; a Galaxy-wide Earth-level null moves P(none other ever) from 0.53 to 0.68. The technosignature future ignores settled systems while the artefact future does not. Resolving the scaling-relation candidates is missing and would change nothing.
  • Corrected wording: What would change the answer. A second origin of life, or one biosignature among 25 HWO planets: P(none other ever in the Galaxy) 0.53 to about 0.3, 'empty universe' 0.20 to 0.03. A Solar System artefact survey finding nothing lowers 'unnoticed' in proportion to its assumed efficiency (0.19 to 0.10 at 50%). Radio SETI of the present kind (a million stars at Earth level): no change; an Earth-level census of the whole Galaxy finding nothing would move P(none other ever) to 0.68 and the chance of another civilisation now from 0.22 to 0.02. Resolving the Tully-Fisher and fundamental-plane candidates as natural: no change. M-dwarf verdicts: no change except through the earliness argument.
• Claim 11: reword
  • Reason: Direction is right and robust. The 0.6 for 'not evident' under SIA depends on the universe-wide non-evident rule: with a claiming minority allowed the split is 0.52-0.62 stuck, 0.23-0.36 unnoticed, 0.09-0.10 hidden. The SSA column has its own fragility (pockets) that the claim should name.
  • Corrected wording: Anthropics is the largest lever and it cannot be argued away. Under SIA emptiness is ruled out and at least 0.93 of the mass is 'they exist; expansion fails or is not evident'. The split is prior: 0.33 / 0.59 / 0.05 (barrier / not evident / hidden) at base, 0.62 / 0.23 / 0.10 if a minority of expanders may be of the claiming kind. Under SSA the answer also depends on whether claimed galaxies leave any unclaimed pockets.

Code errors

• models/model3.py / line 72, warm_curve=((0.01, 3e3), (0.1, 1e5), (0.85, 1e6))
  • Problem: The 1e6 point and the sub-50% points are Lacki 2016 numbers for galaxies shrouded at just above the CMB temperature, applied in the 100-600 K branch. G-HAT supports about 1e5 at 85% or more, about 5e3 at 50-85%, and nothing below 50%.
  • Fix: Warm branch ((0.5, 4.8e3), (0.85, 1e5)); move Lacki's curve to a separate very-cold branch.
  • Impact: The whole survey effect comes from the unsupported points: P(none other ever) contribution +0.016 becomes +0.003; stuck 0.250 to 0.262.
• models/model3.py / line 73, cold_curve, tag [STUDY tf_kiii + fp_kiii]
  • Problem: Adds bright early types to HI-rich discs (1,200 = 3/0.0029 + 3/0.018), uses the pre-referee Tully-Fisher 1.8% at 0.75 (referee 2.3%), encodes searches with 16 and about 20 unresolved outliers as zero-count samples, and is labelled 'any temperature' although the Tully-Fisher part fails above about 300 K.
  • Fix: Use discs only for a Milky-Way-like galaxy ((0.75, 130), (0.9, 214)), or a Poisson likelihood with the observed outliers and unknown background; relabel.
  • Impact: None on the posterior: removing the curve or multiplying it by 100 changes no class at three decimals. The tag implies a weight the studies do not have.
• models/model3.py / lines 185 and 225-236 (p_here = 1.0 for the claiming regime; D_all and share_G)
  • Problem: The SSA denominator truncates all births after the first expander, with no fill delay and no pocket civilisations. It is consistent with the numerator only when the local term is on and p_here is exactly 1. share_G is not a fraction in truncated universes.
  • Fix: Add pocket civilisations to D_all ((1 - p_here) x untruncated births), apply the same fill delay as phi(T_fill), and skip SSA when use_local is False.
  • Impact: With a pocket fraction of 1e-6 to 0.1, SSA class 7 goes 0.016 to 0.28 and SSA P(none other ever) 0.33 to 0.25. Without the fix any p_here below 1 gives SSA excluded = 1.000.
• models/integrate3_results.json / ablation rows 'Prior (no observations)', "+ Earth's timeline only", '+ galaxy surveys (no local argument)': schools.ssa
  • Problem: SSA posteriors are degenerate (excluded = 1.000, ESS 2,800-5,200) because D_all assumes pre-emption while the local term is switched off.
  • Fix: Do not store or quote SSA for rows with use_local = False, or make the denominator follow the switches.
  • Impact: The draft quotes site only, so no headline changes; an SSA ablation table cannot be built from this file.
• models/model3.py / lines 62-63 and 176-178 (headstart table; mu_loc = N_ever * phi * p_exp)
  • Problem: N_ever already counts only civilisations arisen by now, and phi(0) = 0.72 applies 'older than the Sun' a second time. The table is Earth-clock star ages while the cosmic term uses hard-step birth tables. The table is the LAMOST primary curve, not the adopted 0.70.
  • Fix: mu_loc = p_exp x sum over classes of W x F(T0 - T_fill) from cosmo_tables, or divide phi by phi(0).
  • Impact: mu_loc low by a factor 1.1-1.5; stuck 0.250 to 0.246, unnoticed 0.188 to 0.191.
• models/model3.py / lines 212-215 (techno_null_1e6, artefact_null)
  • Problem: artefact_null applies a fixed 0.5 regardless of p_here, so the reported 0.19 to 0.10 is the assumption. techno_null_1e6 counts only quiet civilisations and ignores occupied systems in a settled Galaxy.
  • Fix: Tie survey efficiency to p_here or sample it (0.1-0.7); let nearby-star technosignature nulls act on the non-evident regime; add a Galaxy-wide census.
  • Impact: Galaxy-wide null: P(none other ever) 0.53 to 0.68. Million-star null seeing settled systems at 50%: unnoticed 0.19 to 0.10.
• report/FINDINGS_DRAFT.md / section B table, prior column and the mean-likelihood line; claim 1
  • Problem: Prior column sums to 0.96: 'settled in a way we would see' is 0.309 in the results file, not 0.27. Mean likelihoods are slightly off against the 3e6 run (en route 0.41 not 0.44; hidden 0.69 not 0.68). SSA excluded is 0.016, shown as 0.01.
  • Fix: Regenerate the table from integrate3_results.json base; change '27% of prior mass' to 31%.
  • Impact: Presentation only.
• models/model3.py / lines 239-246, class 'stuck'
  • Problem: Threshold on expectations: worlds with expected arrivals of 0.1-1 (10% of the class) and few-civilisation worlds with ordinary propensity (8%) are labelled 'sustained expansion almost never happens'; there is no class for a wave in progress or one that came and went.
  • Fix: Split by cause: p_exp below 1/N_ever (barrier) versus small N_ever; report the median p_exp of the class (5e-11).
  • Impact: Relabelling of about 0.04 of posterior mass.

Missing explanations

• A claimed Galaxy with unclaimed pockets: Landis percolation voids, Carroll-Nellenback steady states, settlers who skip Sun-like stars. Representable only as the separate 'low-footprint' regime; under SSA it would hold 0.12-0.33.
• A wave that came through and is gone, beyond the 1-10 Myr evidence horizon of the geological record: 0.07-0.10 (site) if departed waves leave no trace, about 0.01-0.02 if they often do.
• Bounded bubbles from sub-critical waves: present only inside 'stuck', with no local or survey consequence modelled.
• Mixed populations: mostly low-footprint expanders with a rare claiming minority. The model allows only all-or-nothing per universe.
• Probe-only exploration without settlement: the class label says 'or visited' but no visit term exists.
• A wave still in progress inside the Galaxy: 1.3% of 'stuck' and unlabelled, because the fill time is capped at 1 Gyr.
• The classic SETI position as a reported quantity: quiet civilisations present now. The model has it (0.22 site, 0.39 SSA) but no headline claim states it.
• Harvesting that takes the faint stars first: under 40% of the light even if every dwarf is enclosed, which no survey in the model constrains.

Were the round 1 fixes made?

Item by item against the round-1 requests from this lens.

1. Continuous treatment of self-sustaining, persistent, evident expansion: partly delivered. The floor on the will to expand and the feasibility switch are gone, replaced by one hierarchical p_exp prior, which is what I asked for as a minimum. Not delivered: any reproduction-number structure, bubble reach, or split by mode; p_exp is independent of regime, and 'persistent' and 'evident' are defined in, not modelled. In effect it is still a barrier with prior P(p_exp < 1/N) of 0.24 at N = 1e6.
2. A class for 'settled or probed but unnoticed': delivered and over-corrected. It is a universe-wide rule with prior 0.25 and mean evidentness 0.047, so it passes through at 1.78 times its prior; a claiming minority cuts it to 0.03-0.12.
3. Finite persistence: not delivered. Tested here: 0.07-0.10 of site mass if departed waves leave no trace.
4. Hop-range-dependent fill time: under-delivered (speed-dependent only, capped at 1 Gyr) but immaterial in version 3. The round-1 expectation of +0.08 on restraint no longer holds because the propensity prior now spans dozens of decades; a 5 Gyr fill time moves 'stuck' by 0.01.
5. Survey limits as curves: delivered in form. The warm curve is mis-sourced and the cold curve is a sum of unlike samples, but the term went from carrying the whole result in version 2 to +0.016 (+0.003 after the fix). The round-1 objection that the local term was inert is reversed: it now does all the work, by the assumption p_here = 1.
6. Misleading study tags: the flyby tag on feasibility is gone and the head-start study now has a report. Remaining: the cold-curve tag implies weight the studies do not have (zero), and the warm curve cites Lacki 2016 for the wrong temperature band. Also fixed: per-term ablation table; hand-entered barrier Bernoullis removed; the sceptic camp changes one lever. Not fixed: the barrier class still has no observational test except a Galaxy-wide technosignature census. New in version 3: the SSA column is valid only when claiming settlement is evident with probability exactly 1.

Critic’s own estimate of each answer

• Empty universe: 0.15
• Rare: 0.26
• En route: 0.04
• Stuck: 0.29
• Unnoticed: 0.15
• Hidden: 0.02
• Other: 0.09
• What “other” covers: A Galaxy that has been claimed or crossed by settlers without that being evident here for non-deliberate reasons: we sit in an unclaimed pocket (voids, steady state, skipped star type), or a wave came through and is gone. Weighted about two parts site to one part SSA from my package run (site 0.182/0.307/0.047/0.287/0.146/0.024/0.007; SSA 0.087/0.151/0.041/0.288/0.156/0.029/0.248). Priors behind it: claiming 0.62, low-footprint 0.35, stealth 0.03; claiming-regime log10 p_exp uniform on (-0.5 - W, -0.5) with W log-uniform 2-40; low-footprint W log-uniform 1-10; claiming minority within the low-footprint regime with W2 log-uniform 2-40; pocket fraction log-uniform 1e-6 to 0.1 with pocket civilisations in the SSA reference class; persistence log-uniform 1 Myr to 10 Gyr with trace probability 0.3; corrected survey curves.

Presentation advice

Lead with the four-row decomposition, not the class table: prior; minus the contradicted class; times Earth's timeline; all terms (P(none other ever) 0.44, 0.61, 0.505, 0.53). It shows the reader in one glance that one assumption and one data set do the work.

Say plainly what the original studies did and did not do. The head-start, flyby, Tully-Fisher and fundamental-plane studies close loopholes (not enough time; cold waste heat) but move the posterior by at most 0.012 and 0.000. A sophisticated reader will find this out; it is better stated by the authors.

Show 'stuck', 'unnoticed' and 'hidden' as one family, 'common but not expanding or not evident' (0.45 site), with a prior-determined split, and give each class as a prior-posterior pair. Quote the scan's 5-95% ranges next to every base number.

Rename 'stuck' to something like 'a shared barrier stops sustained expansion' and print the median propensity inside it (5e-11 per civilisation; 1e-15 under SIA) so nobody reads it as restraint by choice.

Add two rows beside the headline table: the claiming-minority row (unnoticed 0.19 to 0.075) and the tail-depth row for the expansion prior (stuck 0.08 to 0.33 as the maximum width goes from 8 to 60 decades).

Flag the SSA column as conditional on claimed galaxies having no unclaimed pockets, or fix the reference class and show both.

Report P(another civilisation in the Galaxy now): 0.22 site, 0.39 SSA. It is the number a SETI reader wants and the draft omits it.

In claims 4 and 5 use the refereed numbers and name the prior work (Lineweaver 2001, Zuckerman 2026, Annis 1999, Zackrisson et al. 2015).

Critic: an astrobiologist

Overall verdict

Version 3 delivered most of what my lens asked for in round 1, and the site column is robust to every biological correction I could run (0.50 to 0.58 for "no other civilisation ever in the Galaxy"). Three things are not defensible as drafted. First, the SSA column, and with it claim 7 and the "0.33 to 0.53 by anthropic convention" in claim 2, is an artefact of assuming that a K-dwarf planet is about 9 times (interquartile 3 to 44) more likely than an Earth twin to produce a civilisation; with defensible habitability assumptions SSA gives 0.52 to 0.55, the same as site. Second, claims 7, 8 and 9 and one named camp quote numbers that the project's own studies withdrew or that their referees rejected ("ceiling", "4.6 to 942", the 0.35/0.39/0.20/0.06 step posterior, "K dwarfs keep us in the first 1%"). Third, the headline 0.5 is the biological prior restated: in every camp I ran, the posterior is the prior plus 0.03 to 0.10, and defensible biologists span 0.24 to 0.82. The most important single change is to fix the long-lived-host treatment (persistence, abiogenesis and ultraviolet over 20 to 45 Gyr windows, and the cap prior) and then present claim 2 as conditional on a stated biological prior, with the prior printed beside it.

Objections

• major / model3.py class_chain (L107-122) and BASE L_K=(20,45), p_cap_short=0.5, L_cap_long=(15,300), f_persist; the SSA column of table B; claims 2, 7 and 11
  • Argument: In the model a planet of a 0.6-0.75 Msun K dwarf is far more productive than an Earth twin. Under the site posterior the per-planet K/G yield has median 8.7, interquartile 2.8-44, 90th percentile 395 (34 at the median when the cap is long), and G-class hosts supply a median 12% (mean 19%) of all civilisations. SSA then penalises every world with hard steps or slow abiogenesis, which is the whole difference between the site and SSA columns. Four assumptions produce that yield, all leaning the same way. (a) Abiogenesis runs at a constant rate for the whole window: in the gated branch a G planet gets 0.25 of a window and a K planet s-0.75, a factor of about 20, although the settings usually invoked for prebiotic chemistry (impact-generated reducing atmospheres, a young active star, vigorous early volcanism) are front-loaded. (b) f_persist is anchored on Tyrrell's 3 Gyr survival test but is charged once whether the window is 5.6 or 45 Gyr, while every step rate scales with the window; under a constant hazard the effective window is T/ln(1/f_persist), about 3 Gyr at the median f_persist. (c) L_K is the full main-sequence lifetime with no ultraviolet penalty, although the class straddles the K5 limit of Rimmer et al. 2018 (gaps 5.3) and the mdwarf_early referee showed the K-dwarf result needs a 0.6 Msun cut with the whole main sequence habitable (four-step rank 1.05% becomes 6-29% with a 0.7 cut or a 60% window). (d) The cap prior puts 50% on 15-300 Gyr. The geophysics I could verify leans short: stagnant-lid planets keep outgassing for 1-5 Gyr (Foley and Smye 2018, arXiv:1712.03614); they 'die young', melting shuts down after a few Gyr, while plate-tectonic planets may stay volcanic for the stellar lifetime (Kite, Manga and Gaidos 2009, arXiv:0809.2305); degassing lifetime depends on radiogenic inventory and age (Unterborn et al. 2022, arXiv:2204.04243); Earth's own plate tectonics is projected to stop 1.5-5 Gyr from now (Cheng 2018 and O'Neill et al. 2016, seen only through news summaries). Nothing I know of supports 100-300 Gyr without tidal heating; 232Th has a 14 Gyr half-life (from memory).
  • Proposed change: (i) p_cap_short = 0.8 (range 0.65-0.9) and L_cap_long = (15, 50) Gyr. (ii) Make persistence depend on window length: a constant-hazard version (effective window from the incomplete-gamma ratio) as one bracket and the present flat factor as the other. (iii) L_K multiplied by U(0.6, 0.9) for habitable-zone residence, and an ultraviolet factor on lambda for the K class, log-uniform 0.1-1 (0.01-1 for M). (iv) A named camp in which abiogenesis is possible only in the first reference window on any host. (v) Until then, label the SSA column 'SSA, if long-lived hosts are as productive as assumed' and show the corrected column beside it.
  • Expected effect: SSA P(no other ever in the Galaxy) 0.33 -> 0.52 (mild set: cap prior 0.8, long cap 15-50, window fraction 0.6-0.9, K ultraviolet 0.1-1) or 0.55 (full set); site 0.53 -> 0.54-0.55. The site-SSA gap falls from 0.20 to under 0.03. The 'planets stay habitable beyond 15 Gyr' row changes from 0.33 -> 0.14 to 0.52 -> 0.36 (mild) or 0.55 -> 0.53 (full).
  • Tested: Parametrised copy m3x.py (reproduces base: site 0.531, ssa 0.332 at n=600k). SSA none-ever, one change at a time: hazard persistence 0.46; L_K x U(0.5,0.7) 0.46; K ultraviolet factor 0.01-1 0.50; abiogenesis in first window only 0.44; p_cap_short 0.7/0.85/1.0 gives 0.37/0.39/0.40; cap 5.6-10 Gyr 0.48; long cap 15-50 alone 0.335; r_K 0.2-0.5 0.36. Combined mild 0.520 (site 0.545); combined full 0.550 (site 0.539). Long-lived row: base 0.143; with hazard 0.43; with early-only abiogenesis 0.39; mild 0.355; full 0.525. Outputs in b1.out, b4.out, b5.out.
• major / Draft claims 7, 8 and 9; run3.py L25 camp 'Convergence-study steps'; BASE tags '[STUDY convergence, timing]' on m_weights and '[STUDY mdwarf_early: retention mean 0.34]' on p_M_class
  • Argument: The draft quotes first-pass study numbers that the revised reports withdrew or that the referees rejected. Timing (revised report): 'ceiling' was withdrawn, since B(m) is the all-hard-steps value and the factor rises to 19.5-26 if the later step is not hard; '4.6 to 942' was withdrawn as a division-by-near-zero artefact and replaced by 13.3 and 12.8 with dates integrated out. Convergence (revised report): the posterior P(0,1,2,3+) = 0.35/0.39/0.20/0.06 and 'about 1 real hard step among 4' are withdrawn, yet run3.py still runs exactly those weights as the 'Convergence-study steps' camp. The study now says recurrence limits 'cannot test' hard-step models, 0 to 4 candidates remain open, and 1-2 is the analyst's judgement, so 'convergence data do not support many hard steps' is the overclaim its referee struck. 'Clustered p = 0.12' is the Fisher test at the 543 Ma cut for one of three groupings; the same grouping gives 0.040 at 1 Ga, the other two give 0.008, and the trend test gives odds ratio 1.44 (0.51-5.0), p = 0.47. M dwarfs (report still first pass, referee report on file): 'rank 42 to 71%' is for all stars with lifetimes capped at 10 Gyr, not 'Sun-like hosts only'; 'for four or more, K dwarfs keep us in the first 1%' holds only for a 0.6 Msun cut with the whole main sequence habitable (1.05% becomes 2.8% and 6.3% for cuts at 0.65 and 0.7, and 5.9-29% with a 60% window); '0 of 14' is selected on outcome (12 host stars, none in the habitable zone, 0-3 of 17 with ambiguous cases); the referee says retention 0.34 is not an estimate (0.09-0.68).
  • Proposed change: Rewrite claims 7, 8 and 9 from the revised reports (wording in claim_verdicts). Rename the camp 'Few shallow steps (first-pass convergence posterior, since withdrawn)' or replace its weights with ones consistent with the revised study, for example {0:0.15, 1:0.35, 2:0.30, 3:0.13, 4:0.07}. Change the m_weights tag to 'modeller's judgement; convergence study leaves 0-4 open'. Change the p_M_class tag to 'judgement; study leaves habitable-zone retention unconstrained'.
  • Expected effect: No change to the base table. The renamed camp moves from none-ever 0.40 (withdrawn weights, depth 2-6) to 0.545 with study-consistent weights at base depth, so the draft loses its one 'study-backed' camp below 0.5.
  • Tested: Read the three revised reports and referee reports. Ran study-consistent weights: site none-ever 0.545, SSA 0.365, posterior gated share 0.11.
• major / Claims 2 and 11: 'about 0.5 (0.33 to 0.53 by anthropic convention)' and 'anthropics is the largest lever'
  • Argument: The 0.5 is the biological prior plus a small update, in every camp. Prior -> site posterior for 'no other civilisation ever in the Galaxy': Mills-leaning 0.21 -> 0.24; careful-biologist composite 0.41 -> 0.51; base 0.44 -> 0.53; Carter-leaning 0.74 -> 0.82; the driver's own camps run 0.08 (biology optimist) to 0.95 (strict ignorance). In the 400-draw joint scan the site value has a 5-95% range of 0.26-0.74, and the two largest regression coefficients are the mean number of hard steps (+0.10 per standard deviation) and w_fast (-0.07), ahead of p_nonevident (-0.07) and the expansion prior (-0.03). The range quoted in claim 2 covers only the anthropic convention, and after the K-dwarf correction that range (site to SSA) collapses to 0.52-0.55. So SIA is a lever; SSA against site is not; the biological prior is the largest lever that a reader can actually argue about.
  • Proposed change: State claim 2 conditionally, with priors printed: three or four rows by biological camp (Mills-leaning, base, Carter-leaning), each with prior and posterior. Reword claim 11 to name two levers: self-indication and the biological prior. Put the 2x2 table already in integrate3_results.json (biology_2x2) into the findings.
  • Expected effect: Headline range for 'nobody else ever in the Galaxy' widens from 0.33-0.53 to 0.24-0.82 (0.08-0.95 at the extremes of the driver's camps). The three-way split 'roughly equal' holds only for the base prior: Mills-leaning gives none 0.24, stuck 0.32, unnoticed plus hidden 0.42; Carter-leaning gives 0.82, 0.11, 0.07.
  • Tested: Mills-leaning (w_fast 0.6, steps {0:0.6, 1:0.25, 2:0.15}, depth 2-6, f_site 0.03-1, f_persist 0.3-1 dropped in gated branch, f_gate 0.1-1, f_c 0.03-1): site 0.236, SSA 0.116. Carter-leaning (w_fast 0.25, steps {1:0.15, 2:0.3, 3:0.3, 4:0.25}, depth 2-12, f_c 0.001-1): site 0.821, SSA 0.731. Composite: site 0.512, SSA 0.319. Scan regression R2 = 0.94 (site).
• major / BASE w_fast=0.4 with log10 lambda ~ U(0,3); slow_decades=(3,100); tau_over_T=(0.02,0.16); the 'empty universe' row and claim 3
  • Argument: w_fast = 0.4 is equivalent to a log-uniform rate prior with a floor near 10^-4.5 per window. The abiogenesis brief's baseline floor is 10^-30 (prior 0.09, posterior 0.33) and the gaps file instructs 'P(fast) 0.35 (0.14-0.86)' after timing. The model's value after the timeline alone is 0.67, and 0.55 after all observations, so the base case sits on the chemistry-optimist side of the project's own brief. I still think 0.4 is defensible (round 1 asked for it), but it is a camp, not neutral ground. The 'empty universe' row is set by the slow-camp hyperprior and by nothing else: no observation distinguishes lambda = 10^-25 from 10^-90. The size of the timeline's push also depends on accepting the 4.2 Ga molecular-clock LUCA date, which the timing study shows is squeezed against its calibration maximum.
  • Proposed change: Keep 0.4 as base but label it, and show w_fast = 0.2 as the 'abiogenesis brief baseline' row and 0.6 as the chemistry-optimist row. Report 'empty universe' only as a range with its driver named, or merge it with 'rare' in the headline table. Give claim 3 with both dating choices.
  • Expected effect: Site none-ever: 0.655 at w_fast 0.2 (empty universe 0.29, sterile Galaxy 0.38), 0.436 at 0.6 (empty 0.12). Slow camp 3-30 decades: empty 0.08, none-ever 0.48; 10-100 decades: empty 0.27, none-ever 0.59. Early-life delay from LUCA only: none-ever 0.46, P(fast) 0.70; accepted fossils only: 0.57 and 0.47. Timeline-only ablation: 0.44 -> 0.29 (LUCA), 0.34 (base), 0.38 (fossils).
  • Tested: All values from runs at n = 400k-600k (b2.out and the tau runs); base P(fast) after timeline only 0.675, after all observations 0.555.
• moderate / model3.py L205-211: second_genesis false-positive 0.05 and hwo 'none = (1-p)^25 * 0.97'; claim 10
  • Argument: Claim 10's numbers rest on a 3% chance that a 25-planet survey yields a false biosignature and a 5% chance that a 'confirmed independent origin' is not one. The project's biosignatures brief records five high-profile claims since 1976 and no accepted detection, puts P(shared ancestry | life found on Mars) at 0.5 (0.2-0.8), and gives a Bayes factor of only about 2 for Mars life of unknown ancestry. Abiotic oxygen is a known false positive for exactly the signature the model uses. With a slow-abiogenesis prior mass of about 0.45, the posterior after a detection is controlled by the false-positive floor. The model's own forecast should also be stated: 19% for at least one biosignature in 25 planets (16% true), and 54% that an independent origin would be confirmed in the Solar System, which is a strong and untested prediction of the fast camp.
  • Proposed change: Make the false-positive rates explicit parameters with ranges (survey-level 0.03-0.2 for HWO; 0.05-0.3 for a second genesis, the top end for Mars). Define the observation in the claim as 'community-accepted and demonstrably independent'. Print the forecast probability beside each future-observation row.
  • Expected effect: After an HWO detection: none-ever 0.27 (3%), 0.35 (10%), 0.41 (20%); empty universe 0.04, 0.085, 0.12. After a second genesis: none-ever 0.32 (5%), 0.36 (15%), 0.40 (30%); empty universe 0.03, 0.06, 0.10. With a lower detectability (q = 0.25, gated worlds oxygenated 30% of the time) the forecast falls to 11% and a detection gives 0.32.
  • Tested: Runs in b2.out. Forecasts from evidence ratios: hwo_some 0.188, hwo_none 0.812, second_genesis 0.54.
• moderate / model3.py L161-168: L_late = m_hard * x^(m_hard-1), L_early = 1-(1-q)^(m_hard+1); L_late = 1 for the gated branch; no hybrid branch; claim 8's last sentence
  • Argument: The model uses only two dates (first life, our arrival) and one corner of the timing study. Its mean timeline likelihood by number of post-abiogenesis steps is 0.52 (none, gated), 0.41 (one), 0.67 (two), 0.88 (three), 1.01 (four): four ungated hard steps are favoured 2.5 to 1 over one and 1.9 to 1 over gating. The timeline alone therefore raises the mean number of hard steps from 1.15 to 1.47 and lowers the gated share from 0.30 to 0.265, and with all observations P(three or more steps) is 0.35 against a prior of 0.18. The timing study's comparison on the dated intermediate transitions points the other way: gated hard steps beat ungated by 2.3-3.5 and pacing beats ungated by 3.8-5.0 (upper bounds, gates chosen after the fact), while the Carter coincidence returns 1.0-2.4 or 3-17 depending on the gate-time prior. Claim 8 quotes 'pacing vs gated hard steps 0.6 to 3.3', but the model has no gated-hard-steps branch, so that number is not represented. The hybrid (one to three hard steps plus 1-2 Gyr of gated delay, Watson 2025) was requested in round 1 and is still missing.
  • Proposed change: Add a hybrid branch in which the hard steps play out in a window shortened by a 1-2 Gyr delay (x = 0.60-0.72). Give the gated branch a late-arrival likelihood from an explicit gate-time distribution. Until then, show a sensitivity row multiplying the gated branch by 0.33 to 4 and say that the direction is undetermined.
  • Expected effect: Site none-ever 0.55 (gated x 0.33), 0.51 (x 2), 0.48 (x 4); posterior gated share 0.09, 0.37, 0.54; mean hard steps 2.0 to 1.0. With x = 0.60-0.72: none-ever 0.50, mean hard steps 1.52, SSA 0.31.
  • Tested: Per-step likelihood means and posterior step distribution from a 1M-sample base run; sensitivity runs in b1.out and b2.out.
• moderate / Table B rows 'no other civilisation in the observable universe' and 'rare'; the missing split of 'alone' requested in round 1
  • Argument: The split between 'empty universe' (0.20) and 'rare' (0.30) is the slow-camp hyperprior (0.08-0.29 across the variants above) and has no observational handle. The split that does have one is absent from the draft although the model computes it. Of the 0.53, 0.25 is a sterile Galaxy (fewer than one other inhabited planet), 0.05 is in between, and 0.24 is a Galaxy with at least a thousand inhabited planets and no other technology ever (0.195 with at least a million). These two worlds make opposite predictions for a biosignature survey, and the second is the Ward-Brownlee world the convergence study bears on.
  • Proposed change: Replace or supplement the empty/rare rows with 'sterile Galaxy' and 'inhabited Galaxy, no other technology', and say which future observation separates them.
  • Expected effect: No number changes. The headline becomes three-way within 'alone': sterile 0.25, microbial 0.24, intermediate 0.05 (SSA: 0.16, 0.14, 0.03). Under the prior the microbial share is 0.12, so it is the cell the observations double.
  • Tested: Computed from base samples (life_worlds against N_ever < 1), n = 1M.
• moderate / BASE f_site=(0.01,1), f_persist=(0.03,1), f_gate=(0.03,1), f_c=(0.01,1): the implied prior on P(technological civilisation | planet), never stated in the draft
  • Argument: The rebuild left the implied per-planet prior almost where version 2 had it. Per rocky habitable-zone planet: median 10^-9.2, 5-95% 10^-48.5 to 10^-2.9, P(above 1%) = 1.1% (version 2: 10^-9, 10^-49 to 10^-2.5, 2%). Given an inhabited planet that is a habitable site: median 10^-4.1, 5-95% 10^-12.7 to 10^-1.5, P(above 10%) = 1.5%. Even in the gated branch, which is supposed to be the Mills et al. world, P(above 10%) is 3.6% (median 10^-2.5), because f_gate, f_persist and f_c are multiplied as independent log-uniforms and f_persist and f_gate both charge for the planet staying clement long enough. A reader should see these numbers. A careful biologist's composite (f_site 0.03-1, f_persist 0.1-1 and dropped in the gated branch, f_gate 0.05-1, the rest as base) gives: inhabited-planet median 10^-3.7, 5-95% 10^-12.4 to 10^-0.9, P(above 1%) 25%, P(above 10%) 6%; per rocky habitable-zone planet median 10^-8.5, 5-95% 10^-48 to 10^-2.1, P(above 1%) 4%, fast-camp median 10^-4.3. The distribution is bimodal and about 46 decades wide at 90%, so its median means little; the camp-conditional medians are what to quote.
  • Proposed change: Print the implied prior (per rocky habitable-zone planet and per inhabited planet, overall and by camp) in the findings. Drop f_persist in the gated branch, where it duplicates f_gate. Narrow f_site to 0.03-1.
  • Expected effect: Small on the table: site none-ever 0.53 -> 0.51, unnoticed 0.19 -> 0.21, SSA 0.33 -> 0.32. Dropping f_persist in the gated branch alone changes nothing to three decimals. The effect is on what the reader is told the model assumes.
  • Tested: Quantiles from 400k-1M sample runs of base, composite, Mills-leaning, Carter-leaning and the driver's optimist camp.
• minor / BASE f_c=(0.01,1.0)
  • Argument: The revised convergence study gives an upper limit on the hazard of technology arising in a large-brained lineage and no lower limit, and says values 'well below 0.1' are as compatible as 0.5. A hard floor at 0.01 is therefore a judgement. The factor is also charged in full when the last hard step is itself the cognitive or language transition (round 1 item 11, not addressed).
  • Proposed change: Show f_c = (1e-4, 1) as a sensitivity row; set f_c to 0.3-1 in samples with three or four hard steps, where the last step can only be the cognitive one.
  • Expected effect: f_c down to 1e-4: site none-ever 0.53 -> 0.58, SSA 0.33 -> 0.37, unnoticed 0.19 -> 0.16.
  • Tested: Run at n = 600k (b3.out). The double-count correction was not run.
• minor / BASE p_M_class=0.30 and h_M=(0.01,1)
  • Argument: A per-planet retention fraction (itself unconstrained, 0.09-0.68 according to the referee) is used as the probability that the whole class is habitable, and then multiplied by a per-planet factor again. If a third of M-dwarf habitable-zone planets keep atmospheres, the class is habitable with certainty.
  • Proposed change: Either p_M_class near 1 with h_M log-uniform 0.003-0.3, or keep the binary and call it what it is (a class-level barrier such as ultraviolet-limited prebiotic chemistry or tidal locking).
  • Expected effect: None on the table: site 0.525-0.534 and SSA 0.323-0.336 across p_M_class 0.1, 0.7 and 1.0 with h_M 0.003-0.3. The claim that M-dwarf verdicts do not move the answer survives.
  • Tested: Three runs in b4.out.
• minor / model3.py L168 (exponent m_hard+1 applied when abiogenesis is not itself hard) and L114/L120 (hard if below 0.3, with a 1/hard! jump)
  • Argument: L_early = 1-(1-q)^(m_hard+1) is the first-of-(m+1) order statistic and is right only when abiogenesis is one of the hard steps. For lambda between about 1 and 30 it overstates the early-life term (0.83 against an exact 0.66 at lambda = 10, tau = 0.09, one later step). The 0.3 threshold makes the ordering factor jump by up to a factor of the step count.
  • Proposed change: Use the exact joint likelihood (abiogenesis at rate lambda before tau, then m hard steps, last at x), which is a one-dimensional quadrature; replace the threshold with the hypoexponential form or leave it and say so.
  • Expected effect: Exact joint: site none-ever 0.531 -> 0.539, P(fast) 0.55 -> 0.53, SSA unchanged. Threshold 0.1 or 0.5: 0.523 or 0.535.
  • Tested: Implemented the exact joint with 48-node Gauss-Legendre quadrature in m3x.py (timing_exact) and a threshold switch; runs in b1.out and b3.out.

Claims that survive

• The site column's P(no other civilisation ever in the Galaxy) of about 0.53 is robust to every biological correction I ran that keeps the base camp weights: exact timing likelihood 0.54, hardness threshold 0.1-0.5 gives 0.52-0.535, dropping f_persist in the gated branch 0.53, careful-biologist composite 0.51, all K-dwarf fixes 0.54-0.55, M-dwarf class probability 0.1-1.0 gives 0.525-0.534.
• The observations move the biological answer very little in any camp (prior to posterior shift of 0.03-0.10), so 'any confident single answer is a statement of priors' holds from this lens.
• Earth's timeline pushes toward a populated Galaxy in every variant (timeline alone: 0.44 to 0.29-0.38), and removing it raises none-ever to 0.635. The direction is robust; the size depends on the LUCA date.
• M-dwarf atmosphere verdicts do not change the table in either the site or the SSA column (under 0.01).
• A biosignature or second-genesis detection is the most diagnostic biological observation available: even with a 20-30% false-positive allowance it cuts none-ever from 0.53 to about 0.40 and the sterile-Galaxy share from 0.25 to 0.11-0.15.
• An HWO null is weakly informative, as it should be: none-ever 0.53 to 0.59, P(fast) 0.55 to 0.47, with a forecast probability of 0.81.
• The first-and-last-event timing formulae are numerically adequate: the exact joint form changes no class by more than 0.01.
• The 1/m! ordering factor, the rate-based abiogenesis branch, f_site in the HWO likelihood and the oxygen step are all implemented as round 1 asked.
• Under SIA the biological priors do not matter (P(fast) 0.99, gated 0.66, mean hard steps 0.12), consistent with the timing study's self-indication result.

Verdict on each headline claim

• Claim 1: reword
  • Reason: 'Less than a factor of 2' mixes two terms that pull in opposite directions and is not robust. With the timeline switched off, the silence discriminates among the non-excluded classes by at most 1.5 and favours emptier ones (mean likelihoods: empty 1.00, rare 0.93, stuck 0.84, unnoticed 0.87, hidden 0.99, en route 0.67). Earth's early life then reverses the order (empty 0.35, hidden 0.68). The spread is 1.5 with accepted fossils only and 3.2 with the LUCA date only.
  • Corrected wording: The silence rules out one thing only: expansion that claims systems and would be evident here (27% of prior mass, likelihood 0.004). Among everything else the silence itself discriminates by at most 1.5, in favour of emptier classes; the larger spread in the table (0.35 to 0.68) comes from Earth's early life, which favours fuller classes by 1.5 (accepted fossils) to 3.2 (the 4.2 Ga LUCA date). Any confident single answer is a statement of priors.
• Claim 2: reword
  • Reason: The 0.5 is the biological prior (0.44) plus 0.09; defensible biological priors give 0.24 to 0.82; the 0.33 SSA figure is an artefact of K-dwarf productivity and becomes 0.52-0.55 when corrected. 'Roughly equal' holds only for the base prior.
  • Corrected wording: Three live answers; their shares are set mainly by the biological prior. With the base prior: (a) nobody else has ever arisen in the Galaxy, 0.53 (prior 0.44), half of it a sterile Galaxy (0.25) and half a Galaxy with many inhabited planets and no other technology (0.24); (b) others arose but sustained expansion almost never happens, 0.25; (c) expansion happened and is not evident from here, 0.2. Across defensible biological priors (a) runs from 0.24 to 0.82, always within 0.10 of its prior. SSA agrees with these figures (0.52-0.55) unless K-dwarf planets are several times more productive than Earth-like ones, in which case it gives 0.33; SIA gives 0 for (a) and 0.6 for (c). Deliberate hiding stays under 0.05 unless its prior is raised by hand.
• Claim 3: reword
  • Reason: Numbers reproduce (0.441 to 0.344 to 0.514 to 0.532). The size of the first step depends on the early-life date, and the timeline also raises the number of hard steps, which the claim omits.
  • Corrected wording: Ablation. Earth's timeline alone pushes toward a populated Galaxy (P(none other ever) 0.44 -> 0.34; 0.29 if the 4.2 Ga LUCA date is taken at face value, 0.38 with accepted fossils only), because early life favours fast abiogenesis; the same likelihood raises the expected number of hard steps from 1.15 to 1.47. The absence of evident settlement here pushes back (-> 0.51); the galaxy surveys add little (-> 0.53); radio SETI adds nothing.
• Claim 4: reword
  • Reason: Outside my lens, but checked against the revised headstart report: it now gives 0.70 with a systematic range of 0.64-0.78 and says the 0.69-0.76 interval 'is not the uncertainty'; the per-planet limit is 3.4e-9 to 7.9e-9 once the planet count is marginalised. From my lens, the 2.5-2.9 billion are habitable-zone rocky planets by orbit; the model's own f_site and f_persist make the habitable ones 10 to 1,000 times fewer.
  • Corrected wording: Keep the Gaia sentence. Then: about 70% (systematic range 64-78%) of Sun-like stars at the Sun's radius are older than the Sun, mean head start about 4 Gyr, so roughly 2.5 to 2.9 billion rocky planets in habitable-zone orbits had time to settle the Galaxy at any speed from 1e-4 c to 0.1 c. If settlement is of the evident kind, the probability per such planet of producing a Galaxy-settling civilisation is below about 1e-9 for that count and below 3e-9 to 8e-9 with the count's uncertainty included.
• Claim 7: reword
  • Reason: Three statements misreport the study or ignore its referee: 42-71% is for all stars with a 10 Gyr cap, not Sun-like hosts only; 'K dwarfs keep us in the first 1%' needs a 0.6 Msun cut with the full main sequence habitable; '0 of 14' is selected on outcome. The model sentence (0.33 to 0.14) holds only under the model's K-dwarf assumptions: it becomes 0.52 to 0.36 (mild fixes) or 0.55 to 0.53 (full fixes).
  • Corrected wording: 'We are early' depends on how long planets stay habitable, not on M dwarfs alone (original data). With the measured star-formation history and every planet habitable for at most 10 Gyr, humanity is not early (rank 42 to 71% for one to ten hard steps). JWST: none of 14 rocky M-dwarf planets with a decisive verdict (12 host stars, none in the habitable zone) keeps a thick atmosphere, which leaves habitable-zone retention unconstrained and removes the earliness puzzle only if there are one or two hard steps. For four or more, whether we are early turns on late K dwarfs: rank 1% if stars down to 0.6 solar masses are habitable for their whole 30 to 45 Gyr, 6 to 29% with a 0.7 solar-mass cut or a 60% habitable window. Published geophysics puts the outgassing and climate-regulation lifetime of rocky planets at a few to about 10 Gyr unless plate tectonics persists. In the model, the SSA column equals the site column (0.52 to 0.55) unless K-dwarf planets are several times more productive than Earth-like ones; under that assumption it is 0.33, and 0.14 if planets also stay habitable beyond 15 Gyr.
• Claim 8: reword
  • Reason: 'Ceiling' and '4.6 to 942' were withdrawn in the study's revision; 'about one bit' is 0.4-1.3 bits for one marginal under conditioning on our existence; the pacing comparison needs its counterpart (ungated hard steps fit worse) and the convention stated. The model does not contain a gated-hard-steps branch, so the last sentence describes the study, not the model.
  • Corrected wording: Earth's timeline carries likelihood ratios of order 10 and about 0.4 to 1.3 bits (original reanalysis). Kipping's 13:1 for fast abiogenesis reproduces (13.3 and 12.8 with date uncertainty and all 38 repository LUCA datings integrated out); it becomes 9, 7 and 6 if 2, 3 or 4 hard steps follow, 20 to 26 if the later step is not hard, and 1.5 to 6 with accepted fossils only. Snyder-Beattie et al.'s tables reproduce only with thresholds e^-k where 10^-k is printed (an apparent slip; authors not contacted), which lowers their three-step weight below 10^-12 from over 90% to 44-58%. On the dated transitions, ungated hard steps fit worse than gated hard steps or pacing (2.3 to 5.0, upper bounds), and the latter two cannot be separated (0.6 to 3.3). All of this conditions on our own existence; under self-indication the same data put f_bio at 0.2 to 0.8.
• Claim 9: reword
  • Reason: The revised study says recurrence data 'cannot test' hard-step models and leaves 0-4 candidates open; 'do not support many hard steps' is the overclaim its referee struck. 'Clustered p = 0.12' is one of several values (0.47 for the trend, 0.040 at the 1 Ga cut, 0.008 for two other groupings). The brain result omits that no non-human has a measured cortical neuron count within a factor of 2.
  • Corrected wording: Convergence data cannot test hard-step models, but they weaken the usual argument for them (original reanalysis). Vermeij's 'unique innovations are old' result (odds 3.2-fold per tenfold age) falls to 1.4 (0.5 to 5.0, p = 0.47) when the co-inherited traits of three stem lineages are counted once. Three of 13 proposed hard steps have recurred (plastids, complex multicellularity, life on land); abiogenesis, oxygenic photosynthesis, eukaryogenesis and human-level cognition remain single and untestable, so zero to four hard steps stay open. Relative brain size within a factor 2 of ours arose in 3 to 4 mammal lineages (best well-sampled dolphin 0.73 of human), though no other species has a measured cortical neuron count within a factor 2, and nothing constrains the step from brains to technology.
• Claim 10: reword
  • Reason: The quoted 0.3 and 0.03 depend on 3-5% false-positive floors; with 10-30% (the project's own biosignatures brief gives 0 of 5 past claims accepted and 0.5 shared ancestry for Mars) the results are 0.35-0.41 and 0.06-0.12. The model's forecast (19% for HWO) and the timescale should be stated. The artefact and SETI parts are outside my lens.
  • Corrected wording: What would change the answer. An accepted, demonstrably independent second origin of life, or a confirmed biosignature among 25 HWO planets (forecast probability about 0.2; 2040s at the earliest): P(none other ever in the Galaxy) 0.53 -> 0.27 to 0.41 and the sterile-Galaxy share 0.25 -> 0.04 to 0.15, depending on whether the chance of a false or non-independent detection is 3% or 20-30%. A null from 25 planets: 0.53 -> 0.59. A thorough Solar System artefact survey finding nothing: 'unnoticed' 0.19 -> 0.10. More radio SETI: no change. M-dwarf verdicts: no change.
• Claim 11: reword
  • Reason: True for SIA. Not true for SSA against site once the K-dwarf treatment is fixed (0.52-0.55 against 0.54), and the biological prior moves the headline further than SSA does in the draft (0.24-0.82 against 0.33-0.53).
  • Corrected wording: Two levers dominate and neither can be argued away: self-indication and the biological prior. Under SIA, emptiness is ruled out and the answer becomes 'they exist; expansion fails or is not evident', with the split set by priors. Without SIA, 'nobody else ever in the Galaxy' runs from about 0.25 to 0.8 across defensible biological priors. The difference between the site and SSA columns is not an anthropic effect but an astrobiological one: it measures how productive long-lived K-dwarf planets are assumed to be.

Code errors

• models/model3.py / L146-152: extra = log_fgate + log_fpers + log_fc passed unchanged to class_chain for the G, K and M windows
  • Problem: f_persist (survival of climate and biosphere, anchored on a 3-4 Gyr test) and f_gate are applied once regardless of window length, while abiogenesis and every hard-step probability scale with the window (s up to 8 for K, up to 180 for M). Long-lived hosts get the benefit of time with none of the attrition.
  • Fix: Make persistence window-dependent, for example an effective window s_eff = (P(n, h*s)/P(n, h))^(1/n) with h = ln(1/f_persist) and P the regularised incomplete gamma function, used both in class_chain and for the table lookup; keep the flat version as the other bracket.
  • Impact: SSA none-ever 0.33 -> 0.46 (0.44 even with every planet long-lived); site 0.53 -> 0.545.
• models/model3.py / L112-113: s_life = max(s - 0.75, 0.05) for gated worlds, p_life = 1 - exp(-lam * s_life)
  • Problem: Abiogenesis opportunity grows without limit with the stellar window: a gated G planet has 0.25 of a window, a gated K planet 3-7 windows, so slow-abiogenesis K planets are about 20 times more productive (median K/G yield 7.5 in the slow/gated cell, 16.5 in slow/hard). No allowance for abiogenesis being confined to young planets or for the ultraviolet limit below K5.
  • Fix: Add a cap on the abiogenesis window (camp: first reference window only) and a per-class factor on lambda (K: log-uniform 0.1-1; M: 0.01-1).
  • Impact: Early-only abiogenesis: SSA 0.33 -> 0.44. K ultraviolet factor 0.01-1: SSA 0.33 -> 0.50.
• models/model3.py / L168: L_early = np.where(gated, q, 1 - (1 - q) ** (m_hard + 1))
  • Problem: The exponent m_hard+1 treats abiogenesis as one of the hard steps even when lambda is 1-30 per window. Example: lambda = 10, tau = 0.09, one later hard step gives 0.83 where the exact value is 0.66.
  • Fix: Exact joint likelihood: integral over [0, tau] of lamexp(-lamt)m(x-t)^(m-1) divided by the integral over [0, 1] of lamexp(-lamt)*(1-t)^m, by quadrature (implemented as timing_exact in m3x.py).
  • Impact: Site none-ever 0.531 -> 0.539; P(fast) 0.55 -> 0.53.
• models/model3.py / L205-211: 0.05 in second_genesis and 0.97 in hwo_none / hwo_some
  • Problem: False-positive rates are hard-coded and low relative to the project's biosignatures brief; they set the posterior after a detection.
  • Fix: Expose as parameters (sg_fp, hwo_fp) with ranges 0.05-0.3 and 0.03-0.2 and report the range.
  • Impact: HWO detection: none-ever 0.27 -> 0.35-0.41; second genesis: 0.32 -> 0.36-0.40.
• models/run3.py / L25: 'Convergence-study steps': m_weights {0:0.35, 1:0.39, 2:0.20, 3:0.06}
  • Problem: These are the first-pass posterior that the convergence study withdrew after its referee report; the camp name attributes them to the study.
  • Fix: Rename the camp or use weights consistent with the revised study (0-4 open, judgement 1-2).
  • Impact: Camp result 0.40 is not study-backed; study-consistent weights at base depth give 0.545.
• models/model3.py / L38 and L41 comments: p_M_class '[STUDY mdwarf_early: retention mean 0.34]', L_K '[STUDY mdwarf_early]'
  • Problem: A per-planet retention fraction that the referee calls unconstrained (0.09-0.68) is used as a class-level probability; L_K is the main-sequence lifetime, not a habitable window.
  • Fix: Relabel both as judgement; multiply L_K by a habitable-window fraction of 0.6-0.9.
  • Impact: p_M_class: none. L_K fraction 0.5-0.7: SSA 0.33 -> 0.46.

Missing explanations

• A hybrid evolution branch: one to three hard steps plus 1-2 Gyr of environmentally gated delay (Watson 2025). Asked for in round 1; the timing study's 'gated hard steps' model has no counterpart in the integrated model.
• Front-loaded abiogenesis: origin possible only on young planets (impact-generated reducing atmospheres, young-star ultraviolet, early volcanism, exposed land with wet-dry cycling). It gives fast timing on Earth, no bonus for long-lived hosts, and looks like low lambda Galaxy-wide if suitable sites are rare (Kipping 2025's own caveat).
• A Gaian bottleneck as its own branch (Chopra and Lineweaver 2016): life starts often and usually dies within about 1 Gyr. It predicts many dead or Archean-type worlds and few oxygenated ones, which HWO can distinguish from slow abiogenesis; in version 3 it is one flat multiplier.
• Persistence as a hazard that grows with window length, including the secular decline of outgassing and climate regulation as mantles cool, which acts on K- and M-dwarf planets whatever the star does.
• Biospheres that cannot support technology: ocean worlds with no land or fire, oxygen below the roughly 18% combustion threshold (Balbi and Frank 2023), ice-covered oceans. f_c has no term tied to them, and they may correlate with host class.
• Shared ancestry through lithopanspermia inside a planetary system, which turns 'life found on Mars' into a Bayes factor near 2 rather than 17 or more.

Critic’s own estimate of each answer

• Empty universe: 0.15
• Rare: 0.33
• En route: 0.04
• Stuck: 0.22
• Unnoticed: 0.17
• Hidden: 0.02
• Other: 0.07
• What “other” covers: Outcomes outside the seven classes: a Galaxy with intelligent but non-technological lineages only (ocean or low-oxygen biospheres); biospheres that arise often and die early, so that 'others' existed only as microbes; a hybrid hard-step and gated history with a different yield from either pure branch; and residual model error. My biological core is the careful-biologist composite with the K-dwarf fixes (site and SSA both 0.54-0.55 for none-ever), shaded toward 0.48 because I give the Mills-leaning reading somewhat more weight than the Carter-leaning one, and with the empty-universe share reduced because that row is a hyperprior artefact.

Presentation advice

Lead the biology with the 2x2 table that already exists in integrate3_results.json (fast or slow abiogenesis by gated or hard-stepped evolution): the cells disagree completely (fast and gated: none-ever 0.00; slow and hard-stepped: 0.87), and the reader should see that the cell weights are judgement.

Print the prior beside every posterior and say in one sentence that the observations move 'nobody else ever in the Galaxy' by 0.03 to 0.10 in every biological camp.

State the implied biological prior in plain numbers so it can be challenged. As coded: P(technological civilisation | rocky habitable-zone planet) has median 10^-9, 5-95% range 10^-48 to 10^-3 and a 1% chance of exceeding 1%; given an inhabited habitable planet, median 10^-4, range 10^-13 to 10^-1.5. The prior I would defend: P(life | habitable planet) bimodal, 0.4 (0.2 to 0.6) near 1 and the rest spread over 3 to 100 decades; P(technology | inhabited planet that stays habitable) median about 10^-3.7 with a 90% range of 10^-12 to 0.1 and a 6% chance of exceeding 10%; combined per rocky habitable-zone planet, fast-camp median 10^-4, overall 4% chance of exceeding 1%. It is about 46 decades wide. Swapping it in moves the site column from 0.53 to 0.51, so the table is robust to the centre of the biological prior and sensitive only to its camps (0.24 Mills-leaning, 0.82 Carter-leaning).

Split 'alone' into a sterile Galaxy (0.25) and an inhabited Galaxy with no other technology (0.24). Drop 'empty universe' as a headline number or give it as 0.08 to 0.29 with its driver named.

Mark the SSA column as conditional on the productivity of long-lived hosts until the model is fixed, and show the corrected value beside it.

Give every future-observation row its forecast probability and its false-positive assumption.

Do not attach a study tag to a prior the study does not support: the step-count weights and p_M_class are judgement.

Were the round 1 fixes made?

Fixed: rate-based abiogenesis with a timing likelihood derived from the same rate (a small exponent slip remains, effect under 0.01); fewer and shallower hard steps (prior mean 1.3, depth 2-8 decades) with a named Carter camp; the 1/m! ordering factor; f_site and the oxygen step in the HWO likelihood (forecast now 19%, a null is weakly informative); x_now set to 0.70-0.85; a host-star likelihood, though only inside the SSA weights; existence conditioning for the empty-universe row; the 2x2 table, which is computed but absent from the draft.

Partly fixed: the double-counted habitability factors (f_hz and f_rare merged into f_site, but f_persist and f_gate still multiply in the gated branch; removing f_persist there changes nothing, so this is cosmetic); the sterile against microbial split (computed as side statistics, not shown as a split of 'alone'); the M-dwarf ultraviolet penalty (folded loosely into h_M, none on lambda, none for K dwarfs).

Not fixed: the hybrid branch of hard steps plus gated delay; a late-arrival likelihood for the gated branch tied to a gate-time distribution; f_c charged on top of a cognitive last step; lithopanspermia in the second-genesis row.

Over-corrected or newly wrong: the 'Convergence-study steps' camp adopts the step posterior that round 1 took from the first-pass study and that the study has since withdrawn. The new K-dwarf class answers the host-star request but gives long-lived hosts a per-planet yield about 9 times Earth's (interquartile 3 to 44), which now drives the whole SSA column. The implied per-planet prior is almost unchanged from version 2 (median 10^-9.2 against 10^-9; P(above 1%) 1.1% against 2%); the headline moved from 0.75 to 0.53 mainly through the expansion prior, not through biology.

Critic: the studies as quoted in the draft findings

Overall verdict

The eight studies are sound and reproducible: I reran five pipelines in a scratch copy (timing, convergence, seti_meta, tf_kiii, mdwarf_early) and every output table matched the published one exactly, and I confirmed the flyby, headstart and fundamental-plane headline arithmetic independently. The draft findings are not faithful to them. Claims 4, 5, 7, 8 and 9 mostly quote first-pass numbers and phrases that each study's own referee struck and the revised report withdrew: "at any waste-heat temperature", "first limits", "11.1 +- 0.5", "72.5% (69 to 76%)", "ceiling", "4.6 to 942", "four or more", and a mislabelled "Sun-like hosts only" rank. Claim 5 as titled is false by its own study's revised text. Inside the model the studies carry almost no weight: deleting the two optical galaxy studies changes no posterior at three decimals, the head-start table changes P(none other ever in the Galaxy) by at most 0.004, and the radio and M-dwarf inputs by at most 0.02. The galaxy-survey step that does move the result comes mostly from a published near-CMB limit (Lacki 2016) applied to the wrong temperature branch. The one study-tagged input that matters is the hard-step prior, which the studies explicitly say they cannot constrain, and a posterior the convergence study withdrew is still run as a named camp. The single most important change is to rewrite claims 4 to 9 from the revised report.md texts with their conditions and prior-art credits, and to describe the studies as bounding and scoping the inputs rather than driving the answer.

Objections

• fatal / Draft claim 5 (title, 'about 2%', 'among about 1e5', 'first limits') and the model3.py line 73 tag 'any temperature [STUDY tf_kiii + fp_kiii]'
  • Argument: The claim as worded is contradicted by the revised study reports it cites. tf_kiii report.md says the first-pass headline '1.8%, 0.90%, any temperature' does not survive: the Tully-Fisher limit holds at full strength only for waste heat cooler than about 250 K (3.1 to 3.2% at 300 K, 5 to 6% above 400 K). The revised value is 2.6% (34 unresolved candidates among 1,766 effective galaxies; 1.9 to 3.5% across variants; 4.5% in the cleanly selected ALFALFA subsample), not 'about 2%'. Before vetting, the limit equals Zackrisson et al. 2015 (verified abstract arXiv:1508.02406: 1,359 discs, conservative 3%), so 'first limits' is wrong for Tully-Fisher, and Annis 1999 used the fundamental plane first. fp_kiii: 0.29% is correct, but the effective sample at 75% dimming is 14,853 galaxies (16 observed against 9,663 effective), not 1e5. The limit is for uniform grey dimming of all starlight; the study's own revision says that enclosing only dwarfs and subgiants removes 20 to 36% of an old galaxy's light, where the limit is 7 to 17% or absent. Five objects with 75% or greater deficit remain unexplained and the residual tail is asymmetric (11 dim against 1 bright beyond 0.6 dex, p = 0.003). A hostile expert needs only the studies' own referee reports to refute the claim.
  • Proposed change: Retitle to 'No star-harvested galaxies in optical scaling relations either, within stated limits' and use the corrected wording in claim_verdicts. Relabel line 73 as 'cool waste heat (below about 250 K) or grey dimming; early types and discs'. State the unexplained candidates (34 Tully-Fisher, 5 fundamental-plane) as not excluded rather than writing 'no'.
  • Expected effect: No change to any model number: deleting cold_curve entirely gives site none-ever 0.528 against 0.529 and ssa 0.330 against 0.330. The change is to wording and credibility only.
  • Tested: Recomputed limits from the study tables: 34/1,766 gives 2.56%; 5/683 gives 1.54%; ALFALFA gold 21/672 gives 4.5%; fundamental plane 16 against 9,663 gives 0.285%. Tully-Fisher pipeline rerun: limits.csv identical. Model runs at 400,000 samples: no cold curve, Tully-Fisher only, and doubled fundamental-plane weight all give site 0.528 to 0.529 and ssa 0.330.
• major / Draft claims 4, 7, 8, 9: first-pass numbers that the studies' referees struck
  • Argument: Claim 4: '11.1 +- 0.5' is the first-pass error; revised is 11.1 +- 1.0 per Myr (variants 9.9 to 11.9; 12.5 to 13.7 with brown dwarfs). '72.5% (69 to 76%)' is one sample's statistical-only interval; revised is 'about 70%' with systematic range 64 to 78% (an independent Gaia sample gives 64%). 'Below about 1e-9' holds only at fixed fiducial planet counts; the revised report gives 3.4e-9 to 7.9e-9 once planet-count uncertainty is included. The flyby revision also states where the argument fails (3% of systems settleable, hops under 0.15 pc, velocity change 5 km/s or less: crossing 0.7 to 3.6 Gyr), which the title 'do not explain the silence' omits. Claim 7: 'Sun-like hosts only ... rank 42 to 71%' is the row for a 10 Gyr habitable-lifetime cap with all stars; the Sun-like-only row (hosts above 0.8 Msun, unlimited lifetime) is 66%, 21%, 9.8% for 1, 4, 6 hard steps. 'One or two hard steps' should be 'about one' (critical M-dwarf weight 0.053, 5.3e-5, 5.3e-8 for 1, 2, 3 steps). 'For four or more' is explicitly withdrawn: the rank at four steps is 1.03%, 2.8%, 6.6%, 21% for host cuts 0.6, 0.65, 0.7, 0.8 Msun. Claim 8: 'ceiling' is wrong (the value is 19.5 to 26 if a later step is not hard); '4.6 to 942' was removed as a divide-by-near-zero artefact (revised range 3.3 to 30 with LUCA dates, 1.5 to 5.7 with fossils only). Claim 9: 'p = 0.12' is the friendliest of five numbers (Fisher p 0.008 under two other groupings, 0.040 at a 1 Ga cut, permutation 0.47), and the title 'do not support many hard steps' contradicts the revised finding that recurrence statistics cannot test hard-step models in either direction.
  • Proposed change: Regenerate claims 4 to 9 from each study's revised report.md headline and key-numbers table, not from result.json first-pass fields or report_firstpass.md. Use the corrected wordings in claim_verdicts. Add a build check that every number in the findings appears in a revised report.
  • Expected effect: No model number changes. Stated uncertainties roughly double (rate), the per-planet limit loosens by 3 to 8 times, and three qualitative statements reverse or become conditional.
  • Tested: Read all eight report.md, referee_report.md and result.json files in full. Recomputed from raw GCNS: encounter rate 11.1 per Myr (11.76 for all objects within 50 pc), exposure 20,949 systems in 2.33 Gyr, limits 1.43e-4 and 7.3e-4. Head-start planets 2.49e9 to 2.87e9 and 1.05e-9 to 1.20e-9 confirmed. mdwarf_early rerun: all 23 output files identical, including rank_grid_n_Lmax.csv and rank_noM_cut_window.csv.
• major / model3.py line 72 warm_curve and draft claim 3 ('the galaxy surveys add little (-> 0.53)')
  • Argument: warm_curve = ((0.01, 3e3), (0.1, 1e5), (0.85, 1e6)) is Lacki 2016's Planck 100 GHz 'blackbox' curve (under 1 in 800 at 1% absorption, 1 in 30,000 at 10%, 1 in 810,000 full; evidence/gaps_and_corrections.md line 113). The arXiv abstract (1604.07844) says a blackbox has a temperature just above the cosmic microwave background. The model applies it to the 70% branch where waste heat emerges at 100 to 600 K. There the real constraint is Griffith et al. 2015 (0 of about 1e5 above 85%, 50 natural-looking above 50%), and the project's own evidence/seti.md line 117 says warm searches do not exclude use of under about 25% of starlight. So the model credits infrared surveys with excluding 1 to 25% covering fractions that no survey reaches at those temperatures. Nearly the whole survey step in the ablation comes from this, and none from the two original studies.
  • Proposed change: Set warm_curve to Griffith only, for example ((0.5, 4.7e3), (0.85, 1e5), (1.0, 1e5)), and move the Lacki curve to a near-CMB sub-branch of the cold case. Reword claim 3 to say the surveys add 0.00 to 0.01 and that all of it is from published infrared and millimetre limits.
  • Expected effect: Site P(none other ever in Galaxy) 0.529 falls to 0.516-0.524; empty universe 0.195 falls to 0.182-0.190; en route 0.042 rises to 0.044-0.048; ssa none-ever 0.330 falls to 0.319-0.326. The survey step shrinks from +0.015 to between +0.002 and +0.010.
  • Tested: 400,000-sample runs. No surveys at all: site 0.514, ssa 0.317. Griffith-only warm curve with 50 candidates kept: 0.516 and 0.319. Generous Griffith curve down to 25%: 0.518 and 0.321. Griffith warm plus Lacki in the cold branch: 0.522-0.524 and 0.324-0.326. Monte Carlo standard deviation is about 0.001.
• major / model3.py line 51 m_weights tagged '[STUDY convergence, timing]' and run3.py line 25 camp 'Convergence-study steps'
  • Argument: The hard-step prior is the only study-tagged input that moves the answer, and neither study supports it. The convergence revision says the compatible range is 0 to 4 (0 to 5) and that the data cannot test hard-step models in either direction; the timing study finds the number of steps barely constrained. The base weights (0.30, 0.30, 0.22, 0.12, 0.06) are modeller judgement. The camp 'Convergence-study steps' uses P(0..3) = (0.35, 0.39, 0.20, 0.06), which is the first-pass posterior the study's referee ordered removed and the revision withdrew. Round 1 asked for that camp; it is now stale and misattributed.
  • Proposed change: Retag m_weights as judgement. Rename or delete the camp (for example 'Few hard steps (illustrative)'). Add a camp 'Flat over the study-compatible range 0 to 4'. Report the spread across these in claim 2.
  • Expected effect: P(none other ever in Galaxy), site: base 0.53; withdrawn posterior 0.40; one or two steps only 0.43; flat 0 to 4 gives 0.64; flat 0 to 5 gives 0.71. Ssa: 0.33; 0.24; 0.31; 0.44; 0.51. The 'rare' class ranges from 0.20 to 0.42 (site).
  • Tested: 400,000-sample runs of each m_weights variant; the withdrawn-posterior run reproduces the JSON camp value 0.402 (0.400 here).
• major / '(original data)', 'original', 'first' labels on claims 4, 5, 7
  • Argument: Several headline results reproduce published work, and the revised reports say so. The exposure argument of claim 4 was published by Zuckerman (arXiv:2603.07333, verified: absence of probes implies no civilisation passed within about 100 light-years over the past few billion years). The encounter rate matches Garcia-Sanchez et al. 2001 (11.7 +- 1.3) and Fernandez-Puig et al. (arXiv:2605.16496, verified: 10.6 +- 4.5). The settlement-front result reproduces Carroll-Nellenback et al. 2019. The head-start fraction is described in its own report as a reproduction of the known 'about three quarters' (Lineweaver). The Tully-Fisher test repeats Zackrisson et al. 2015 at seven times the sample; the fundamental-plane idea is Annis 1999. 'K dwarfs suffice' and 'not early with a 10 Gyr cap' confirm Hanson et al. 2021, and the self-datum update is Cook 2022.
  • Proposed change: Label each item as new measurement, reanalysis, or reproduction, and name the prior paper in the claim. What is new: rendezvous-capped settlement with measured Gaia velocities and its failure regimes; the first selection-corrected fundamental-plane limit at 1e5 scale; the table of critical M-dwarf weight against number of hard steps; the Snyder-Beattie table check.
  • Expected effect: No numerical change; removes the most likely priority complaints.
  • Tested: Fetched and read the three arXiv abstract pages listed. The Garcia-Sanchez, Annis, Carroll-Nellenback, Hanson and Cook attributions are as quoted in the study reports, not independently fetched.
• moderate / model3.py lines 62-63 and 176-177 (headstart_phi) and claim 4's link to the model
  • Argument: Three problems. (a) phi is the fraction of Sun-like stars older than the Sun by more than T, but it multiplies N_ever, the number of civilisations that have already arisen. The model's own birth-date tables give the fraction of arisen civilisations born more than 0.3 to 1 Gyr ago as 0.94 to 0.98 for G hosts and 0.67 to 0.96 for long-lived hosts, so phi of 0.63 to 0.69 discounts twice. (b) Because log10_v is at least -4, T_fill lies between 0.3 and 1 Gyr, so six of the eight table nodes are never used, and the value used is the first-pass 0.72 rather than the adopted 0.70 with its 0.64 to 0.78 range. (c) The table is exactly zero beyond 9.5 Gyr, which creates a pocket with no local penalty. The study is therefore decorative in the model, and the claim-4 limit of 1e-9 is not what the model implies: its planet prior gives a 95% per-planet limit of about 3e-9 without the habitable-site factor and 1.6e-7 with it.
  • Proposed change: Replace phi by the fraction of arisen civilisations older than T_fill computed from cosmo_tables, or set it to 1 and say so. Taper the table smoothly instead of ending at zero. In claim 4 quote the limit the model implies next to the study's fiducial one.
  • Expected effect: Base: phi = 1 moves site none-ever 0.529 to 0.531 and ssa unnoticed 0.280 to 0.287; the low end of the revised range changes nothing. Lowering the speed floor to 1.8e-5 c changes nothing (site 0.526-0.528, ssa 0.328-0.329), but at 1e-5 c the ssa column collapses to an effective sample size of 2 with 'stuck' at 0.997.
  • Tested: 400,000-sample runs: phi = 1; phi at low revised range; log10_v floor at -4.5, -4.75, -5.0 with two seeds each (the ssa collapse is identical in both seeds, site column stays 0.524-0.525).
• moderate / tau_over_T = (0.02, 0.16) [STUDY timing] and claim 8's 'about one bit'
  • Argument: The range is a fair envelope of the study's LUCA-based values (0.038 to 0.057) but its upper edge excludes the fossils-only values (0.17 to 0.19), and the draft does not show that this dating choice is one of the larger levers among study inputs. 'About one bit' (0.4 to 1.3, binned, under the conditioning-on-success convention only) understates what the timeline does inside the model, where removing it moves site none-ever from 0.53 to 0.64 and ssa from 0.33 to 0.47. The Snyder-Beattie finding is worded in the draft as established ('reproduce only with'); the revised study calls it an apparent slip, the authors have not been contacted, and the same pipeline misses Kipping's published Jeffreys-prior odds by 6% (1.545 and 1.490 against 1.638 and 1.572) without explanation.
  • Proposed change: Extend tau_over_T to (0.02, 0.20) or show LUCA-only and fossils-only as camps. In claim 8 give the bit range and the convention. Word the Snyder-Beattie point as 'we could reproduce their Table 2 only with ...; not yet confirmed with the authors', and resolve or disclose the 6% Kipping discrepancy before publication.
  • Expected effect: LUCA-only (0.03, 0.06): site none-ever 0.47, ssa 0.27. Fossils-only (0.16, 0.20): site 0.59, ssa 0.40. Base 0.53 and 0.33 sits between.
  • Tested: 400,000-sample runs of both ranges. Timing pipeline rerun: all 5,198 numeric values in results.json identical, including the rms 0.02 against 0.57 dex comparison (this reruns the analyst's code; I did not derive the Snyder-Beattie table independently).
• moderate / p_M_class = 0.30 '[STUDY mdwarf_early: retention mean 0.34]', L_K, r_K, and the claim 10 M-dwarf row
  • Argument: The study says habitable-zone retention is unconstrained (0.09 to 0.68 across fits; my rerun reproduces 0.088, 0.25, 0.68) and that 0.34 is a reference mixture, not an estimate. The model turns that per-planet atmosphere-retention number into a probability that the whole M-dwarf class can host biospheres. The study's main positive finding, that the answer without M dwarfs turns on the K-dwarf mass cut and habitable window, enters only through fixed ranges. The claim that M-dwarf verdicts change nothing is robust in the model, but it sits beside a study that finds our own datum favours no-M-dwarf models by a factor of 4 to 9 in a no-deadline setting; the report should say why the model shows less (it marginalises over lifetime caps and a per-planet factor).
  • Proposed change: Retag p_M_class as judgement bounded by the study range. Add the K-dwarf cut and 60% window as shown sensitivities. Add one sentence reconciling the study's factor of 4 to 9 with the model's small shift.
  • Expected effect: p_M_class 0.09, 0.68, 1.0: site none-ever 0.532, 0.524, 0.521-0.524; ssa 0.334, 0.322, 0.314-0.323. K-dwarf share cut at 0.65-0.7 Msun: ssa 0.361. 60% window: ssa 0.340. All within about 0.03 of base.
  • Tested: 400,000-sample runs of the six variants; mdwarf_early pipeline rerun identical.
• moderate / Publication readiness of the three studies marked partial (flyby, tf_kiii, convergence) and open referee points in the others Others: fp_kiii has five unexplained objects and one cluster with an unexplained 15% deficit; seti_meta's Kardashev-II limits rest on legacy surveys at nominal database thresholds, with 81% of the 1-10 GHz weight from model-counted background stars in about ten pointings; headstart was only partly reproduced by its referee.
  • Proposed change: Before any of these appears in a report: strip process text and regenerate result.json; flyby, lead with the model-free lower bound (8.6 +- 0.8 per Myr or more) beside 11.1; tf_kiii, describe the 34 as not excluded and quote the cleanly selected 4.5% as the conservative figure; convergence, verify the 18 rows against sources or drop the extended table; fp_kiii, list the five objects; seti_meta, lead with modern-survey-only figures.
  • Expected effect: Conservative figures loosen the quoted limits by 1.3 to 2 times (flyby rate, Tully-Fisher) and the Kardashev-II limit from about 100 to 2,000-9,000. No model number changes.
  • Tested: Read each report and referee report; five pipelines rerun with identical outputs (timing, convergence, seti_meta, tf_kiii, mdwarf_early). The seti_meta rerun stopped in the final report-building step only because my scratch copy lacked three downloaded input tables kept in the study's data folder; all 23 result tables had already been written and matched.
• minor / Draft section B table and claim 1 ('27% of prior mass')
  • Argument: The prior mass of the contradicted class is 0.309 in integrate3_results.json, not 0.27; the printed prior column sums to 0.96. Mean likelihoods by class also differ slightly from the JSON: en route 0.413 (draft 0.44), stuck 0.571 (0.58), unnoticed 0.603 (0.61), hidden 0.689 (0.68). These look like values from an earlier run.
  • Proposed change: Regenerate the table from the JSON: prior excluded 0.31; mean likelihoods 0.35, 0.50, 0.41, 0.57, 0.60, 0.69, 0.004.
  • Expected effect: Claim 1 becomes '31% of prior mass'. The 'less than a factor of 2' statement still holds (0.35 to 0.69 is a ratio of 1.96).
  • Tested: Read integrate3_results.json; reran model3 at 1,000,000 samples and reproduced the JSON exactly.
• minor / Draft claim 6 ('Kardashev-II beacons to roughly 100 to 2,000')
  • Argument: This is the closest to faithful of the six study claims, but the range is not the revised one. The revised value is 113 with a range of 55 to 18,000 at 1-10 GHz; 2,000 is the modern-surveys-only figure and 9,000 results with one survey's own star counts. The much tighter 21 cm figure (4.3) is just three divided by the assumed detection efficiency. 'Every star could host one' is supported: under the study's generous leakage bound, a civilisation at every star gives 0.72 expected detections.
  • Proposed change: Use the corrected wording in claim_verdicts, and say the limits are conditional on always-on isotropic transmitters.
  • Expected effect: No model change. Widening c_det to include Kardashev-II beacons moves site none-ever 0.529 to 0.551 and ssa 0.330 to 0.357; including radar-only civilisations or dropping the radio term changes nothing (0.528-0.529).
  • Tested: seti_meta rerun: key_numbers.json (3,966 values) and all result tables identical. Model runs for the three c_det variants.

Claims that survive

• All revised study numbers reproduce. Five pipelines rerun in scratch gave outputs identical to the published ones: timing (5,198 numeric values), convergence (odds ratio 3.20; collapsed 2.94, 1.75, 1.44), seti_meta (23 result tables), tf_kiii (limits.csv), mdwarf_early (23 output files).
• Encounter rate 11.1 per Myr within 1 pc and about 21,000 passages in 2.33 Gyr: confirmed from the raw Gaia nearby-star catalogue independently of the study code. The uncertainty should be +- 1.0.
• Fundamental-plane limit of 0.29% at 75% uniform dimming (16 observed against 9,663 effective) and Tully-Fisher 2.6% (34 of 1,766): arithmetic confirmed. They stand as scoped limits.
• Radio SETI adds nothing to the model: removing the term or widening the detection-chance range to radar-only civilisations leaves P(none other ever in the Galaxy) at 0.528-0.529 (site).
• Galaxy surveys add little: this survives and gets stronger. With each survey applied in its own temperature range the step is 0.514 to 0.516-0.524, not 0.529.
• M-dwarf verdicts barely move the model: setting the class probability anywhere from 0.09 to 1.0 keeps site none-ever within 0.521-0.532 and ssa within 0.314-0.334.
• Time and distance are not the explanation within the model for front speeds down to 1.8e-5 c: results are unchanged (site 0.526-0.529).
• The model results file reproduces exactly on rerun, and the ablation and future-evidence numbers quoted in claims 3, 7 and 10 match it (0.44, 0.34, 0.51, 0.53; 0.33 to 0.14; 0.53 to 0.27-0.32; unnoticed 0.19 to 0.105).
• Humanity is not early if habitable lifetimes are capped near 10 Gyr (rank 42 to 71% for 1 to 10 hard steps, all stars), and the earliness question turns on habitable lifetime and step count rather than on M dwarfs alone.

Verdict on each headline claim

• Claim 1: reword
  • Reason: The prior mass of the contradicted class is 0.31 in the results file, not 0.27. The rest holds.
  • Corrected wording: The silence rules out one thing and one thing only. The observations remove 'expansion that claims systems and would be evident here' (31% of prior mass, likelihood 0.004) and discriminate among everything else by less than a factor of 2. Any confident single answer to the Fermi question is a statement of priors.
• Claim 2: reword
  • Reason: Numbers match the results file, but the split depends on the hard-step prior, which the studies do not constrain. Across step priors compatible with the convergence study, P(none other ever in the Galaxy) runs 0.40 to 0.71 (site) and 0.24 to 0.51 (ssa).
  • Corrected wording: Three live answers, roughly equal. (a) Nobody else has ever arisen in the Galaxy: about 0.5 (0.33 to 0.53 by anthropic convention, 0 under SIA; 0.24 to 0.71 across hard-step priors the data cannot distinguish). (b) Others arose but sustained expansion almost never happens: about 0.25 to 0.33. (c) Expansion happened and is not evident from here: about 0.2 to 0.3 (0.6 under SIA). Deliberate hiding stays under 0.05 unless its prior is raised by hand.
• Claim 3: reword
  • Reason: The ablation numbers match, but the survey step is overstated by a mis-assigned published curve, and the original optical studies contribute nothing.
  • Corrected wording: Ablation. Earth's timeline alone pushes toward a populated Galaxy (P(none other ever) 0.44 -> 0.34, because early life favours fast abiogenesis); the absence of evident settlement here pushes back (-> 0.51); galaxy surveys add 0.01 or less (-> 0.52), all of it from published infrared and millimetre limits and none from our two optical studies; radio SETI adds nothing.
• Claim 4: reword
  • Reason: Uses first-pass uncertainties, omits the conditions under which the argument fails, and presents as original an argument and several numbers that are published.
  • Corrected wording: Time and distance do not by themselves explain the silence, provided settlers can use ordinary stellar encounters (new calculations on Gaia data; the argument follows Zuckerman 2026 and Carroll-Nellenback et al. 2019). Stars pass within 1 pc of the Sun about 11 +- 1 times per Myr, in line with published values of 10.6 to 11.7, so about 21,000 (11,000 to 59,000) systems have passed within 1 pc since the Great Oxidation. About 70% (64 to 78%) of Sun-like stars at the Sun's radius are older than the Sun, by about 4 Gyr on average, which reproduces Lineweaver's result. For a fiducial 4 billion habitable-zone rocky planets around such stars, 2.5 to 2.9 billion had time to settle the Galaxy at front speeds from 1e-4 c to 0.1 c. If settlement is persistent and of the evident kind, fewer than a few in a billion of those planets produced a Galaxy-settling civilisation (1e-9 at the fiducial planet count, 3e-9 to 8e-9 with planet-count uncertainty). The argument fails if only a few percent of systems can be settled, if hops are limited to under 0.15 pc or 5 km/s, or if settlements do not persist: the crossing time then rises to 0.7 to 3.6 Gyr.
• Claim 5: reword
  • Reason: 'At any waste-heat temperature' and 'first limits' are withdrawn or contradicted by the revised studies; 'about 2%' and 'among about 1e5' are first-pass or loose; unexplained candidates remain.
  • Corrected wording: No star-harvested galaxies in optical scaling relations either, within stated limits (new data analysis, extending Annis 1999 and Zackrisson et al. 2015). Fundamental plane, 114,000 early-type galaxies: fewer than 0.3% (0.24 to 0.45%; 1.4% under the most conservative noise model) have 75% or more of all their starlight uniformly dimmed. The limit does not cover enclosing dwarf stars only, which removes just 20 to 36% of an old galaxy's light. Tully-Fisher, 9,400 discs: fewer than 2.6% (1.9 to 3.5%; 4.5% in the cleanly selected subsample) are dimmed by 75% or more, provided the waste heat is cooler than about 250 K. Thirty-four dim discs and five dim early-type galaxies remain unexplained. These limits narrow the cold-waste-heat gap left by infrared searches; they change no number in our model, because the absence of evident settlement in our own Galaxy is a much stronger constraint.
• Claim 6: reword
  • Reason: Closest to faithful. The Kardashev-II range should be the revised one and the conditional nature stated.
  • Corrected wording: Radio SETI constrains behaviour, not abundance (original synthesis). If transmitters are always on and isotropic, the whole record allows fewer than about 4e5 (2e5 to 3e6) beacons of 1e17 W in the Galaxy at 1 to 10 GHz, and fewer than about 100 Kardashev-II beacons (55 to 18,000 depending on which older surveys are trusted; 2,000 to 9,000 from modern surveys alone). It says nothing about Earth-2025-like civilisations: even on a generous reading, one around every star would have produced fewer than one expected detection.
• Claim 7: reword
  • Reason: Mislabels the 10 Gyr cap row as 'Sun-like hosts only', and uses the withdrawn 'four or more' and 'one or two hard steps' statements. The model numbers at the end are correct.
  • Corrected wording: 'We are early' depends on how long planets stay habitable and on the number of hard steps, not on M dwarfs alone (new calculations; the K-dwarf and 10 Gyr results confirm Hanson et al. 2021). If habitable lifetimes are capped near 10 Gyr, humanity is not early even counting all stars (rank 42 to 71% for 1 to 10 hard steps). JWST: 0 of 14 rocky M-dwarf planets with a decisive verdict (12 host stars) has a thick atmosphere, but none is in the habitable zone, where retention is unconstrained (0.09 to 0.68). Removing M dwarfs removes the earliness puzzle only for about one hard step: M-dwarf planets would have to be down-weighted below 0.05, 5e-5 and 5e-8 for one, two and three steps. Without M dwarfs the answer turns on the K-dwarf cut: at four hard steps the rank is 1% if hosts down to 0.6 solar masses count and 21% if only hosts above 0.8 do. In the model, learning that planets stay habitable beyond 15 Gyr moves the SSA column's P(none other ever in the Galaxy) from 0.33 to 0.14; learning that they do not moves it to 0.41.
• Claim 8: reword
  • Reason: 'Ceiling' is wrong and '4.6 to 942' was removed by the study as an artefact. The Snyder-Beattie point is unconfirmed with the authors.
  • Corrected wording: Earth's timeline is worth about one bit (0.4 to 1.3; original reanalysis). Kipping's 13:1 for fast abiogenesis (10 to 19 across date choices) is the value for one hard step; it is 9, 7 and 6 if two, three and four steps are all hard, and ranges from about 3 to 30 across published LUCA dates and assumptions about later steps (1.5 to 6 with fossils only). We could reproduce Snyder-Beattie et al.'s Table 2 only with thresholds e^-k where 10^-k is printed; if that is a slip, their weight on extreme rarity falls from over 90% to 44 to 58%, with their qualitative conclusion unchanged. This has not been confirmed with the authors. Pacing versus gated hard steps: Bayes factor 0.6 to 3.3.
• Claim 9: reword
  • Reason: The title contradicts the revised finding (cannot test either way), and p = 0.12 is one of five values. The encephalisation figure has a range.
  • Corrected wording: Convergence data cannot test the number of hard steps (original reanalysis). Vermeij's 'unique innovations are old' result (odds ratio 3.2, 1.2 to 9.8) weakens to 1.4 (0.5 to 5.0; p between 0.04 and 0.47 by test and age cut) when innovations inherited from LUCA, LECA and the animal body plan are each counted once, but stays at p = 0.008 under two other groupings. Only 3 of 13 proposed hard steps have recurred, and recurrence counts are compatible with anything from 0 to 4 hard steps. Big brains evolved repeatedly (the top non-human encephalisation is 0.73 to 0.86 of human; 3 to 4 lineages are within a factor 2), but no data constrain the step from brains to technology.
• Claim 10: keep
  • Reason: All numbers match the results file (second origin 0.32, biosignature 0.27; empty universe 0.03 to 0.04; unnoticed 0.105; radio null 0.54; M-dwarf alive or dead 0.315 or 0.340 in SSA). The M-dwarf and radio statements held under every variant I ran.
• Claim 11: keep
  • Reason: Outside this lens. One related fragility: the SSA column also loses its effective sample (to 2) if the expansion-speed floor is lowered to 1e-5 c, so the anthropic columns should carry their effective sample sizes.

Code errors

• models/model3.py / line 72, warm_curve
  • Problem: The curve is Lacki 2016's near-CMB Planck limit but is applied to the 100-600 K branch (p_warm = 0.7), where only Griffith et al. 2015 applies and nothing is constrained below about 25% covering fraction.
  • Fix: warm_curve = ((0.5, 4.7e3), (0.85, 1e5), (1.0, 1e5)); move the Lacki curve to a near-CMB sub-branch of the cold case with its own probability.
  • Impact: Site none-ever 0.529 -> 0.516-0.524; empty universe 0.195 -> 0.182-0.190; ssa none-ever 0.330 -> 0.319-0.326.
• models/model3.py / line 73, cold_curve and its comment
  • Problem: Tagged 'any temperature', which the Tully-Fisher study withdrew (valid below about 250 K). About 90% of the 0.75 and 0.9 nodes come from the fundamental-plane study, which covers uniform dimming of early-type galaxies only. The 0.75 node (1200) equals the fundamental-plane value plus the first-pass Tully-Fisher value (1,034 + 167); the revised sum is about 1,150.
  • Fix: Relabel; set the 0.75 node to 1150; consider a Tully-Fisher-only curve ((0.75, 115), (0.9, 200)) for harvesting that spares giant stars.
  • Impact: None at three decimals (site 0.528-0.529, ssa 0.330 with the curve deleted, halved or doubled).
• models/model3.py / lines 62-63 and 176-177, headstart_phi
  • Problem: A fraction of stars older than the Sun is multiplied onto a count of civilisations that have already arisen (double discount of about 1.5 times); only the 0.3-1 Gyr nodes are reachable; the value is the first-pass 0.72 with no uncertainty; the table ends at exactly zero at 9.5 Gyr.
  • Fix: Use the fraction of arisen civilisations older than T_fill from cosmo_tables, or phi = 1; taper smoothly to a small positive floor.
  • Impact: Base changes by 0.002 to 0.007. With the speed floor at 1e-5 c the ssa column collapses to an effective sample size of 2.
• models/run3.py / line 25, camp 'Convergence-study steps'
  • Problem: Uses P(0..3 steps) = (0.35, 0.39, 0.20, 0.06), the first-pass posterior the convergence study withdrew on its referee's instruction.
  • Fix: Rename as an illustrative few-steps camp and add a flat 0-4 camp reflecting the study's stated compatible range.
  • Impact: The camp gives site none-ever 0.40; flat 0-4 gives 0.64 (ssa 0.24 and 0.44).
• models/model3.py / lines 38 and 51, study tags on p_M_class and m_weights
  • Problem: Both tags present modeller judgement as study output. The M-dwarf study calls 0.34 a reference mixture, not an estimate (range 0.09-0.68, and it is per-planet atmosphere retention, not class habitability). Neither the convergence nor the timing study constrains the step count.
  • Fix: Retag both as judgement, citing the study ranges as bounds.
  • Impact: Labelling only; p_M_class matters by at most 0.02, m_weights by up to 0.18.
• report/FINDINGS_DRAFT.md / section B table and the mean-likelihood line
  • Problem: Prior for the contradicted class printed as 0.27 (results file 0.309; column sums to 0.96). Mean likelihoods en route 0.44, stuck 0.58, unnoticed 0.61, hidden 0.68 differ from the file (0.413, 0.571, 0.603, 0.689).
  • Fix: Regenerate from models/integrate3_results.json.
  • Impact: Claim 1 becomes 31% of prior mass.

Missing explanations

• Slow or blocked settlement inside the Galaxy. The flyby study itself finds regimes where the settlement front takes 0.7 to 3.6 Gyr or stalls (3% of systems settleable, hops under 0.15 pc, velocity changes of 5 km/s or less). The model has no class for 'settlement under way in the Galaxy but not yet here': such universes are filed under 'stuck' when the head-start factor is zero.
• Harvesting that spares giant stars, or waste heat between about 30 and 100 K. Neither optical study nor the infrared and millimetre surveys constrain these, yet the model's cold branch treats covering fractions of 50 to 93% as constrained at any temperature.
• Low covering fractions (1 to 25%) with warm waste heat. The project's own evidence file says these hide within natural dust emission; the model currently treats them as excluded at the 1-in-3,000 to 1-in-100,000 level.
• The unexplained objects themselves (34 dim discs, five early-type galaxies with 75% or greater deficit, one cluster with a 15% deficit, and an 11 to 1 dim-to-bright tail asymmetry). They are almost certainly measurement problems, but the report should say they were not vetted to exclusion.

Critic’s own estimate of each answer

• Empty universe: 0.15
• Rare: 0.26
• En route: 0.04
• Stuck: 0.28
• Unnoticed: 0.18
• Hidden: 0.02
• Other: 0.07
• What “other” covers: Outcomes the model does not represent: settlement under way within the Galaxy but slow or percolating and not yet here, and model misspecification more generally.

Presentation advice

Ranking of the eight studies by new, defensible insight:

1. timing: exact reproduction; the Snyder-Beattie table check (pending confirmation with the authors), the one-formula account of how Kipping's odds fall with more hard steps, and the two-convention split are new and feed the model's most sensitive data input.
2. mdwarf_early: the table of critical M-dwarf weight against step count and the K-dwarf-cut dependence are new and honestly conditional.
3. fp_kiii: the only new measurement at scale (114,000 galaxies), defensible if scoped to uniform dimming.
4. flyby: the rendezvous-limited settlement calculation with measured velocities and its failure regimes are new; the rate and the exposure argument are not.
5. seti_meta: a careful unified synthesis whose conclusion is known in outline.
6. convergence: mainly a negative result, sensitive to grouping.
7. tf_kiii: a larger repeat of Zackrisson et al. 2015 with the same unvetted limit.
8. headstart: a reproduction of Lineweaver, and inert in the model.

Lead the report with timing and mdwarf_early, because they bear on the two levers that move the answer, and use fp_kiii as the showcase new measurement with its scope stated. Present flyby plus headstart as a conditional argument with prior-art credit.

State plainly that the original studies bound and scope the inputs rather than drive the result: deleting any of them changes P(none other ever in the Galaxy) by 0.02 or less, while the hard-step prior and the fossil-versus-LUCA dating move it by 0.06 to 0.18.

Most embarrassing finding a hostile expert would raise, per study:

• flyby: the headline argument was published by Zuckerman in 2026, and the catalogue's direct count falls short at the headline radius (9 against 13.1 expected).
• headstart: first-pass precision of +-3% when an independent Gaia sample gives 64%.
• tf_kiii: the first-pass 'any temperature' claim was falsified by its own infrared check, and the unvetted limit equals the 2015 result.
• fp_kiii: a grey-dimming limit says little about realistic star harvesting, and the dim tail is asymmetric (11 to 1).
• seti_meta: headline limits lean on older surveys at nominal thresholds and on modelled background stars in about ten pointings.
• mdwarf_early: none of the 14 decisive planets is in the habitable zone, yet the model stamps a single retention number.
• timing: it alleges a units slip in a published table without contacting the authors while missing Kipping's own published odds by 6%.
• convergence: 18 of 39 extended rows came from memory, and the withdrawn posterior still appears as a named model camp.

Use the revised report.md files as the only source for quoted numbers, and show effective sample sizes beside every anthropic column.

Were the round 1 fixes made?

Fixed: the single cold-survey number (2,000 galaxies) was replaced by a covering-fraction curve built from the two optical studies' actual limits, and the three zero-count equivalents are correct to within 5%. Fixed but now inert: the head-start factor comes from the study, but it is misapplied to a count of arisen civilisations and changes results by at most 0.007. Over-corrected: round 1 asked for the convergence study's step posterior as a camp; version 3 added it, but the study has since withdrawn that posterior, so the camp name now misattributes a retracted result. Not fixed: study tags still present modeller judgement as study output (hard-step weights, M-dwarf class probability), and the cold curve carries the withdrawn 'any temperature' label. Not fixed and newly visible: the infrared survey curve is assigned to the wrong temperature branch, which accounts for most of the survey step in the ablation. Unchanged and fine: the M-dwarf and radio rows show almost no effect, and that held under every variant I ran. The misleading flyby tag flagged in round 1 is gone; the flyby study now enters only through the expansion-speed floor, which its results support down to about 1.6e-4 c.

Critic: a believer in expanding aliens (Robin Hanson’s “grabby aliens”)

Overall verdict

Version 3 delivered most of what round 1 asked for, and the cosmological machinery is sound: the tables reproduce the mdwarf_early ranks (n=4, 30 Gyr K window: 1.01% against the study's 1.03%), N_reach agrees with Kipping 2026 (6.9e5 at 0.1c, abstract fetched), and under Hanson's own premises the SSA column lifts P(expanders hold at least 1% of the universe) from 0.03 to 0.35. The draft headline claims are not defensible as worded from this lens, for three reasons. First, the SSA column is a "why a G star" calculation, not an earliness one: giving K dwarfs the same 5.6 Gyr window as G stars returns SSA P(none other ever in the Galaxy) to 0.53 (site 0.53), a date-only SSA gives 0.49, and switching the expander deadline off moves no SSA class by more than 0.012. So the "0.33 to 0.53 by anthropic convention" range is one astrophysical assumption about late-K planets. Second, "expanders en route" at 0.04 is the prior (0.033) on a two-decade slice of x; P(expanders hold at least 1%) is 0.18 / 0.25 / 0.42 (site / ssa / sia) and is not in the draft, and 42% of it has intergalactic speeds below 0.01c, where the smooth kernel is unphysical. Third, the "stuck" answer (0.25 to 0.33) is the tail of the expansion prior (median p_exp 5e-11) and falls to 0.01-0.05 under the project's own expansion brief. The single most important change is to split and relabel the SSA column (date-only next to date-plus-host, with a prior on K-dwarf habitability and window) and to report P(expanders hold at least 1%) by regime and speed in every table. Scripts are in the critic’s working folder (model3x.py is an instrumented copy that reproduces the base run exactly). model3.py was edited at 04:50 UTC during my run (Wexp added to the kept samples); results were identical before and after.

Objections

• major / model3.py lines 235-236 (share_G = W[0] / D_all; ssa = w_site * share_G); section A description of ssa; the 0.33 figures in claims 2, 7 and 11
  • Argument: The weight is the share of all civilisations ever that arise on G-class hosts. It is a legitimate full-datum SSA likelihood for the model's host structure, but it is not Hanson's earliness argument and the draft presents it as one. Four tests show the whole site-to-SSA shift is the host class, not the date. (i) Conditioning on the date only gives P(none other ever in the Galaxy) 0.485, not 0.334. (ii) Keeping every term but not conditioning on host gives 0.515 (site 0.532). (iii) Giving K dwarfs the G window (L_K = 5.6), or removing K dwarfs, gives SSA 0.534 and 0.517. (iv) Switching the expander deadline off changes no SSA class by more than 0.012. The reason is structural: at n=4 hard steps, old K-dwarf planets supply 43 times as many civilisations today as G planets (209 times at n=6, independent of L_K), so even a deadline falling today leaves G hosts at 14% (n=4) or 4% (n=6) of observers. The best possible deadline recovers 0.068 (n=4) or 0.020 (n=6) of a G-only world's likelihood; under the date-only datum it over-recovers (3.0). The applied penalty at n=4, L_K=30 is 0.18%, 5.5 times harsher than the 1% rank quoted in claim 7. All of this rests on one unlisted assumption: planets of 0.6-0.75 Msun stars are as habitable per planet as G-star planets, with certainty, for 20-45 Gyr, while every planet of a 0.75-1.3 Msun star gets exactly 5.6 Gyr.
  • Proposed change: Report two SSA columns: date-only, weight = L_early * L_local * L_survey * L_seti * sum_c(W_c b_c(T0)) / D_all, and date-plus-host (current). Add a prior that K-dwarf planets are habitable at all (the gaps brief already cites Rimmer et al. 2018 on UV below K5) and a per-planet weight as for M dwarfs. Give the G class a spread of windows instead of a point. Rewrite section A: "ssa: the likelihood includes the fact that our star is Sun-like".
  • Expected effect: SSA P(none other ever in the Galaxy) 0.33 -> 0.49 (date-only) or 0.52-0.53 (K windows not longer than G). Claim 7's 0.33 -> 0.14 becomes 0.49 -> 0.35 date-only. Posterior P(M-dwarf planets habitable) 0.20 -> 0.29 date-only, 0.10 if K is dead. The site column moves by under 0.04 in all of these.
  • Tested: e_ssa.py, e_more.py, e_sens.py, e_trunc.py, 1M samples each. Base SSA classes (empty/rare/en route/stuck/unnoticed/hidden/excluded): model 0.120/0.190/0.046/0.323/0.280/0.026/0.016; date-only 0.187/0.262/0.050/0.264/0.198/0.018/0.020 (ESS 327k); host not conditioned 0.188/0.288/0.053/0.256/0.179/0.016/0.021; deadline off 0.123/0.195/0.041/0.323/0.288/0.026/0.004. Table ranks from cosmo_tables.npz with r_K 0.66, L_K 30: n=1,2,3,4,6 -> 38%, 11%, 3.3%, 1.01%, 0.11% (study: 38, 12, 3.6, 1.03, 0.079); G share 22%, 5.0%, 0.98%, 0.18%, 0.006%.
• major / model3.py BASE log10_v=(-4.0,-0.05) used for both T_fill and kv = RHO_GAL * v3 * p_exp * ig; q_ig=0.6 drawn independently of v**
  • Argument: One speed, log-uniform from 30 km/s to 0.9c, serves as both the in-galaxy front speed and the intergalactic speed. Prior P(v >= 0.1c) is 0.24. In the posterior, 42% of P(x >= 0.01) has v below 0.01c (0.076 of 0.179 site; 0.114 of 0.249 ssa), and 45% of the claiming en-route mass (0.0123 of 0.0273); only 0.0073 has v >= 0.1c. N_reach(v) falls below one galaxy at about 0.001c, the Milky Way's escape speed is 0.0018c, and crossing the 4.6 Mpc mean spacing at 0.003c takes 5 Gyr. So x >= 0.01 at these speeds is the smooth-density formula counting home-galaxy expanders as holding 1% of the universe. The project's own expansion brief recommends separate priors: in-galaxy log-uniform 3e-4 to 0.3c, intergalactic uniform 0.1 to 0.99c. Posterior P(v > 0.3c | claiming, x >= 0.01) is 0.13, equal to the prior, so the model contains none of Hanson's speed inference; the draft should not import it.
  • Proposed change: Draw v_ig separately (uniform 0.1-0.99c per the brief, or log-uniform 0.01-0.99c); allow intergalactic expansion only for v_ig >= 0.003c; set q_ig to 0.8-0.9 as round 1 asked, or show the 0.2-0.9 range explicitly. Print the speed prior beside every statement that depends on speed.
  • Expected effect: v_ig uniform 0.1-0.99c: site rare 0.304 -> 0.251, stuck 0.250 -> 0.205, unnoticed 0.188 -> 0.253, P(x >= 1%) 0.18 -> 0.25 (ssa 0.25 -> 0.32). Adding q_ig = 0.9: rare 0.193, stuck 0.128, unnoticed 0.325, en route 0.077, P(x >= 1%) 0.41 site, 0.53 ssa, 0.79 sia. The guard alone (base speed prior, intergalactic only above 0.003c): P(x >= 1%) 0.179 -> 0.127, en route 0.041 -> 0.029; above 0.01c: 0.102 and 0.022. "Three roughly equal answers" does not survive the brief's speed prior with q_ig 0.9: none-ever 0.55, stuck 0.13, unnoticed plus hidden 0.35.
  • Tested: e_variants.py, e_final.py, e_fsf2.py, e_meet.py. Speed quantiles (5/25/50/75/95%) in claiming worlds with x >= 0.01: 0.001, 0.002, 0.014, 0.113, 0.575 c (site). q_ig alone: 0.9 gives en route 0.064 site / 0.073 ssa, P(x >= 1%) 0.28 / 0.39; 0.2 gives 0.013 / 0.015 and 0.057 / 0.078. The scan's SSA en-route correlates 0.73 with q_ig; it never varies the speed prior.
• major / Draft claim 7 and the FUTURE row "planets_long_lived"; p_cap_short=0.5 (not in section D); m_weights
  • Argument: The quoted number reproduces (ssa 0.334 -> 0.143), but the mechanism is not the one a reader will assume. The model answers "then there were no hard steps": mean hard steps 1.08 -> 0.51, P(no hard steps after abiogenesis) 0.38 -> 0.56, stuck plus unnoticed 0.60 -> 0.77, while claiming expanders holding 1% go only 0.037 -> 0.043. With hard steps held high, the same observation leaves P(none other ever in the Galaxy) near 1 and raises expanders instead. So the size and even the reading of the response are set by the step prior and by the host conditioning in the SSA weight. Three wordings are also wrong against the study. (a) "Sun-like hosts only ... rank 42 to 71%" is the study's all-stars-capped-at-10-Gyr case; Sun-like only (above 0.8 Msun, star-limited) is 66%, 21%, 9.8%, 2.2% for n = 1, 4, 6, 10. (b) "for four or more, K dwarfs keep us in the first 1%" holds only at the 0.6 Msun cut with full windows: n=4 gives 1.03%, 2.8%, 6.6% for cuts at 0.6, 0.65, 0.7; the model tables give 5.1%, 1.0%, 0.2% for L_K = 20, 30, 45. (c) "A cap near 10 to 15 Gyr dissolves the puzzle" is true of the rank (16% at n=4, 15 Gyr) but not of the weight the model applies (G share 0.029); with planets shown short-lived the SSA value is still 0.41 against site 0.54. Finally, the SSA posterior itself puts P(planets die within 15 Gyr) at 0.72 (0.97 in the Carter camp): a strong prediction resting on a 50% atom that section D does not list.
  • Proposed change: State the response as conditional and give both channels in one table (hard steps fall; expanders truncate). Add p_cap_short and the K-dwarf assumption to section D and give the cap a physical basis. Correct the three wordings.
  • Expected effect: Base numbers unchanged. The last sentence of claim 7 becomes a range: P(none other ever) falls 0.33 -> 0.14 (date-plus-host SSA), 0.49 -> 0.35 (date-only), 0.90 -> 0.67 (Carter steps), 0.94 -> 0.82 (expansionist plus Carter). P(claiming expanders hold 1%) rises by 1.2x at base and 4x in hard-step camps (0.024 -> 0.105; 0.076 -> 0.316). p_cap_short 0.1 / 0.5 / 0.9 gives SSA none-ever 0.20 / 0.33 / 0.40.
  • Tested: e_future.py, e_sens.py, e_hanson.py, e_dig.py. Carter steps plus long-lived, ssa: 0.216/0.368/0.087/0.180/0.084/0.007/0.059, ESS 21,948. Expansionist plus Carter plus long-lived, ssa: 0.265/0.135/0.251/0.021/0.216/0.019/0.093, ESS 18,327. Hanson premises (5-7 hard steps, no cap, M dwarfs alive, every expander claims and crosses galaxies, v 0.05-0.99c, 2M samples): P(x >= 0.01) 0.032 site, 0.346 ssa; median occupied fraction 0.35; P(none other ever in the Galaxy) 1.00; empty universe still 0.56 because the cap prior's low end (15-20 Gyr) remains a refuge. Site column: long-lived 0.522, short-lived 0.542.
• major / W_exp=(1.0,30.0); draft claim 2(b) "about 0.25 to 0.33"; section D bullet 2; no prospects-for-humanity output anywhere in the draft
  • Argument: The "stuck" class is the tail of the expansion prior. Its median p_exp is 5e-11 (site), 8e-12 (ssa), 4e-16 (sia); 76-81% of it has p_exp below 1e-6 and 55% of it is worlds without intergalactic expansion. The prior gives P(p_exp < 1e-6) = 0.24 and P(p_exp < 1e-12) = 0.09; the data move the first to 0.36 / 0.39 / 0.42. Under the prior the project's own expansion brief recommends (log-uniform 1e-4 to 1; the "Expansion is easy" camp) stuck is 0.045 site and 0.049 ssa; across the stored camps it runs 0.045-0.384 site and 0.049-0.491 ssa. The draft states 0.25-0.33 as a finding. The same parameter is humanity's own chance of expanding, and the draft never says what the model implies: posterior median 4e-4 (site) or 1e-4 (ssa), mean 0.09, and a 36-42% probability that fewer than one technological civilisation in a million ever expands. That is a late Great Filter and round 1 asked for it explicitly.
  • Proposed change: Relabel the class "a filter ahead: fewer than one technological civilisation in a million ever expands" and show it under two expansion priors. Add a prospects claim with three numbers per convention: P(p_exp < 1e-6); P(no other civilisation has ever arisen in the volume humanity could reach); P(no expander has yet originated in it).
  • Expected effect: Claim 2(b) becomes "0.05 to 0.33 depending on how much prior weight sits below p_exp = 1e-6 (0.01 to 0.49 across camps)". New outputs at base: P(no other civilisation ever in our reach volume) 0.27 at 0.1c (5.5e6 galaxies) and 0.23 at 0.5c (6.9e8 galaxies) site, 0.16 / 0.14 ssa, 0 sia. P(no expander origin yet in it) 0.66 / 0.61 site, 0.54 / 0.50 ssa, 0.25 / 0.22 sia. Mean p_exp prior 0.12 -> posterior 0.09.
  • Tested: e_more.py, e_variants.py, e_final.py. p_exp quantiles (5/25/50/75/95%) site: 2.6e-19, 1.7e-9, 3.8e-4, 0.07, 0.58; prior median 0.0068. Stuck decomposition site: not intergalactic 0.55, intergalactic with p_exp < 1e-6 0.39, intergalactic with v < 0.01c 0.06, other 0.00. Expansionist prior: stuck 0.011, mean p_exp 0.22, P(first in reach volume) 0.44 site / 0.28 ssa.
• major / model3.py lines 239-249 (settled = mu_arr >= 1; c_route = rest & (x >= 0.01)) and CLASS_LABELS; table B rows 3, 5 and 7
  • Argument: (i) The en-route row is the prior: 0.033 -> 0.041 site, 0.046 ssa. It is a two-decade slice of a prior on x that spans more than forty decades, cut again at mu_arr < 1. (ii) P(x >= 0.01), which the label names, is 0.179 / 0.249 / 0.418; it sits in the results file as p_expanders_hold_1pct and is absent from the draft. 85% of it is in the low-footprint and stealth regimes, and 86% of that has x >= 1: universe-wide expanders, filed under "the Galaxy has been settled". (iii) "Settled in a way we would see (contradicted)" holds SSA posterior 0.016 at base, 0.059 with Carter steps and long-lived planets, 0.093 with expansionist priors as well, and 0.108 with large future star formation. These are worlds with claiming expanders already holding 63% or more, with us in the remainder; Hanson's own estimate is 40-50%. 91% of that mass has a truncated denominator, so it is exactly the deadline worlds. A class with up to 11% posterior cannot be labelled contradicted. (iv) The class boundary is a hard threshold while the likelihood is Poisson.
  • Proposed change: Define classes by regime and occupancy. "Claiming expanders hold at least 1% of the universe; we are in the unclaimed remainder" = claim & x >= 0.01 at any mu_arr. "Low-footprint or hidden expanders hold at least 1% of the universe, including here" = non-claim & x >= 0.01. Keep empty, rare and stuck. Print P(x >= 0.01) by regime and by speed band in every camp table.
  • Expected effect: No likelihood changes. En route 0.041 -> 0.027 site / 0.037 ssa / 0.018 sia with honest content; the low-footprint universe-wide class is 0.152 / 0.212 / 0.400; the "contradicted" row falls to 0.002-0.012. New headline number: P(expanders of any kind hold at least 1%) 0.18 to 0.42 at base and 0.25 to 0.9 across expansion priors.
  • Tested: base.py, e_dig.py. P(x >= 0.01) by regime (claiming, low-footprint, stealth): site 0.027, 0.139, 0.013; ssa 0.037, 0.194, 0.018; sia 0.018, 0.368, 0.033. P(x >= 1) by regime, site: 0.001, 0.120, 0.011. Prior P(0.01 <= x < 1) = 0.045, prior P(x >= 0.01) = 0.317.
• moderate / run3.py CAMPS: "Expansion is easy" changes only W_exp; no camp changes speed, q_ig or combines them
  • Argument: There is no expansionist camp with intergalactic reach, so the draft cannot say what Hanson, Olson or Armstrong and Sandberg priors imply inside this model. I ran one. The result is not "grabby aliens en route". With the model's 25% low-footprint atom it is "low-footprint expanders are everywhere" (unnoticed 0.47 site, 0.63 ssa). If settlement would be obvious it is Hart and Tipler: nobody else in the Galaxy at 0.84 site, 0.69 ssa. En route reaches 0.25, and claiming expanders holding 1% reaches 0.32, only with hard steps plus long-lived planets plus SSA. That is a result worth stating.
  • Proposed change: Add three camps to the table and report P(expanders hold at least 1%) for every camp: expansionist (W_exp 1-4, v_ig uniform 0.1-0.99c, q_ig 0.9); the same with settlement obvious (p_nonevident 0.05, p_stealth 0.005); the same with Carter steps, with and without long-lived planets.
  • Expected effect: Expansionist: site 0.302/0.115/0.054/0.011/0.473/0.042/0.003, none-ever 0.575, P(x >= 1%) 0.534; ssa 0.188/0.068/0.039/0.011/0.630/0.058/0.007, none-ever 0.368, P(x >= 1%) 0.672; date-only ssa 0.285/0.099/0.068/0.014/0.475/0.043/0.016, P(x >= 1%) 0.558; sia unnoticed 0.91, hidden 0.09, P(x >= 1%) 0.90 (ESS 1,146). Settlement obvious: site 0.517/0.193/0.087/0.016/0.163/0.018/0.006, none-ever 0.837, P(x >= 1%) 0.259; ssa 0.413/0.146/0.086/0.024/0.281/0.030/0.021, none-ever 0.687, P(x >= 1%) 0.383. Plus Carter steps: site 0.642/0.161/0.069/0.004/0.112/0.010/0.003, P(x >= 1%) 0.190; ssa P(x >= 1%) 0.237. Plus long-lived planets, ssa: en route 0.251, P(x >= 1%) 0.557.
  • Tested: e_variants.py, e_final.py, e_future.py, 1M samples each. SIA effective sample size is 195 in the settlement-obvious variant and about 30 with Carter steps, so those SIA rows are not usable.
• moderate / cosmo_tables.py sfr(): Madau and Dickinson extrapolated to the future with the same form (frac_sf_done = 0.90); the normalisation of b by all-time star formation
  • Argument: Future star formation is fixed at 0.11 of the past. The mdwarf_early study lists published values from 0.02 (Sobral et al.) to 10 (the Behroozi and Peeples upper bound). Under SSA this is a direct multiplier on how much a deadline is worth, independent of hard steps and of K dwarfs, and it is the one lever that raises the en-route class without touching the site column. The tables also normalise births to all-time star formation while Nstars is today's count, so N_ever is 10% low at base and would be ten times low if a larger future were adopted without renormalising.
  • Proposed change: Put a prior on future over past star formation (three points 0.1, 1, 10 or log-uniform 0.02-10) and build tables per value; normalise so that W refers to stars formed to date.
  • Expected effect: Future equal to past: SSA claiming expanders holding 1% 0.037 -> 0.055, en route 0.046 -> 0.060, "contradicted" 0.016 -> 0.027. Future ten times past: 0.170, 0.141 and 0.108, none-ever 0.334 -> 0.286. Site column identical to three decimals once N_ever is renormalised.
  • Tested: e_future_sf.py and e_fsf2.py: rebuilt the tables with a floor sfr(T0) * exp(-(t - T0)/tau), tau = 54 and 540 Gyr (frac_sf_done 0.499 and 0.091), and shifted log10_Nstars by +0.257 and +0.996 to hold N_ever to date fixed. Without that shift the site none-ever drifts to 0.542 and 0.572, which is the normalisation artefact.
• moderate / Draft omission: no meeting-time output, although round 1 asked for one; label "have not reached us"
  • Argument: In LambdaCDM the conformal time left is 16.57 Gly, so the kernel saturates: x can still grow by a factor of 17 to 41 in worlds dominated by G hosts. Other-galaxy expanders therefore ever reach us with probability above one half only if x today is above roughly 0.02-0.04. In the claiming en-route worlds, 20% (site) to 22% (ssa) never arrive with better than even odds, and the median time to an even chance of arrival is 15 Gyr (site) or 13 Gyr (ssa), with a 5th percentile of 2.0 or 1.6 Gyr and P(within 1 Gyr) of 0.7% or 1.3%. Hanson et al. quote 0.2-2 Gyr, from a power-law cosmology and speeds above c/3. Half of the difference is the slow-speed contamination in the speed objection; the rest is the horizon. This is a real result of the merged model and it is not reported.
  • Proposed change: Add the meeting-time distribution and P(ever reached) to the outputs, and one sentence on the horizon: only home-galaxy expanders can pre-empt late civilisations in regions no outside expander ever reaches (the D_all code already does this correctly).
  • Expected effect: No change to class numbers. New outputs: P(other-galaxy expanders ever reach the Milky Way) 0.175 site / 0.240 ssa / 0.414 sia, most of it worlds where they are already here and not evident; P(ever reached | en route) 0.79 / 0.76 / 0.82.
  • Tested: e_meet.py (2M samples; 64,313 claiming samples with x >= 0.01 and mu_arr < 3; arrival time defined by x(t) - x(T0) = ln 2) and e_more.py. Kinf/K0 from the tables: 36 (n=0), 14 (n=1), 21 (n=2), 31 (n=4) on the G row.
• minor / model3.py lines 191-198 (mu_vis via the kernel evaluated at eta0 + R(1-v)/v, clipped at ETA_GRID[-1]; detection rule N_eff >= 100)
  • Argument: Evaluating the kernel at a later time is exact only while eta0 + R(1-v)/v stays below the maximum conformal time, that is for v above 0.075c. Below that the clip under-counts visible domains: by 0.4 dex at 0.03-0.076c, 1.6 dex at 0.01-0.03c, 3.8 dex at 0.001-0.01c and 6.8 dex below. Above 0.075c the later-time kernel also counts civilisations born between now and t_vis, which over-counts by up to a factor 3 (95th percentile) in worlds dominated by K and M hosts. The geometry is right where it matters for Hanson's argument: the ratio of visible-but-not-arrived domains to x is 0.41 at 0.3-0.9c, 2.5 at 0.15-0.3c and 10 at 0.076-0.15c, matching round 1.
  • Proposed change: Integrate over past births only: shell = (4 pi / 3) * [min(vd + R(1-v), d)^3 - (vd)^3] with d = eta0 - eta(t'). Replace the N_eff >= 100 switch with a per-domain detection probability 1 - exp(-N_eff * phi_d), phi_d the fraction of the survey volume one domain covers.
  • Expected effect: No class moves by more than 0.003 with either fix. The cosmic part of the survey term has no leverage on the result, which supports claim 3 ("surveys add little").
  • Tested: t_muvis.py, 400k samples, base against fix_muvis=1 (geometry) and fix_muvis=2 (geometry plus per-domain detection): site en route 0.042 / 0.042 / 0.041; ssa 0.048 / 0.048 / 0.047.
• minor / Section D bullet 5 (N_reach 4e5 to 1.5e6 at 0.1c); cosmo_tables.py lines 66-73 (n = 0 fixed 4.5 Gyr delay, n >= 1 no delay)
  • Argument: The N_reach range is right to within a factor of about 2 and the soft spot can be closed. Kipping 2026 (arXiv:2606.04044, abstract fetched: half the universe filled at one spawn per million galaxies at 0.1c for a 4.5 Gyr start) implies 6.9e5; the tables give 4.1e5 for n = 0 and 8.4e5, 6.0e5, 4.9e5 for n = 2, 3, 4. The top value, 1.48e6 at n = 1, is an artefact: with one hard step civilisations can arise at tau -> 0, immediately after planet formation at z > 2, and the cubic distance weighting amplifies them. The n = 0 row has a 4.5 Gyr delay and the n = 1 row none. Worlds with gated evolution and slow abiogenesis use the n = 1 row although class_chain requires abiogenesis in the first quarter of the window followed by a fixed delay.
  • Proposed change: Give n >= 1 a minimum delay d of 1-3 Gyr: density n (tau - d)^(n-1) / (L - d)^n. Add a dedicated row for gated worlds with slow abiogenesis. In section D replace the comparison with grabby.py by the comparison with Kipping.
  • Expected effect: x changes by at most 3.6x in n = 1 worlds. Scaling x by 3 or by 1/3 across the board moves no class by more than 0.007 and P(x >= 1%) by 0.007 (0.172 / 0.179 / 0.186 site).
  • Tested: e_variants.py (x_scale 1/3 and 3); table inspection of kern/F at T0: V_eff 1,425 Gly^3 (n=0), 5,134 (n=1), 2,919 (n=2), 1,696 (n=4), 1,153 (n=8) for L = 5.6 Gyr.
• minor / model3.py line 178 (mu_loc = N_ever * phi * p_exp) against lines 233-234 (denominator exposure p_exp * N(t)); run3.py ABLATION rows under SSA
  • Argument: N_ever already counts only civilisations born to date; multiplying by phi (the share of stars older than the Sun by T_fill, 0.69-0.72 for fast fill) applies an age cut twice. From the tables the right factor is F(T0 - T_fill)/F(T0), about 0.98 at 0.3 Gyr. The SSA numerator then uses exposure phi * p_exp * N while the denominator uses p_exp * N(t), which inflates worlds with a home-grown wave about to arrive. Separately, with use_local = False the numerator has no survival factor at all while the denominator stays truncated, so the stored SSA rows for "Prior", "+ Earth's timeline only" and "+ galaxy surveys" read excluded = 1.000 with effective sample size 2,800-5,200.
  • Proposed change: Use the same delayed exposure in both places, taken from the tables. Put S(T0) inside the SSA weight instead of relying on L_local, or do not store SSA columns for ablation rows without the local term.
  • Expected effect: With phi = 1: site classes move by at most 0.004; SSA "contradicted" 0.016 -> 0.012. Claim 3 uses the site column and is unaffected, but the ablation SSA rows in integrate3_results.json must not be quoted.
  • Tested: e_sens.py (headstart_phi all 1.0); integrate3_results.json ablation rows read directly.

Claims that survive

• Cosmology and kernel are right: t0 13.79 Gyr, eta0 47.1 Gly, future conformal reach 16.57 Gly, exact Poisson (KJMA) occupancy, linear in the birth law. N_reach at 0.1c of 4.1e5 (n=0) to 1.5e6 (n=1) brackets Kipping 2026's 6.9e5; scaling x by 3 either way moves no class by more than 0.007. Section D's fifth soft spot is closed.
• The birth-date tables reproduce the mdwarf_early ranks: with r_K 0.66 and a 30 Gyr K window, n = 1, 2, 3, 4, 6 give 38%, 11%, 3.3%, 1.01%, 0.11% against the study's 38, 12, 3.6, 1.03, 0.079.
• The deadline machinery works when given Hanson's premises: P(expanders hold at least 1%) goes from 0.03 (site) to 0.35 (ssa), the median occupied fraction is 0.35 (Hanson: 40-50%), and P(none other ever in the Galaxy) is 1.00. With Carter steps, expansionist priors and long-lived planets, claiming expanders holding 1% reach 0.32.
• Evident, system-claiming expansion is the one thing the observations remove (mean likelihood 0.004), and the cosmological term extends it: ignoring other galaxies' expanders would overstate stuck by 0.07 (0.25 against 0.32) and rare by 0.03.
• Galaxy surveys add little and the visibility of en-route domains adds nothing: correcting the light-cone geometry and the detection rule moves no class by more than 0.003. Fast expanders are hard to see (0.41 visible-but-not-arrived domains per unit x above 0.3c against 10 near 0.1c).
• M-dwarf verdicts barely move the answer, even through the SSA column: none-ever 0.334 -> 0.315 (habitable) or 0.340 (dead). The planet-lifetime cap absorbs it.
• The direction of claim 7's hypothetical is robust: P(none other ever in the Galaxy) falls in every variant I ran (0.33 -> 0.14; 0.49 -> 0.35 date-only; 0.90 -> 0.67 Carter; 0.94 -> 0.82 expansionist plus Carter). Only the size and the mechanism are prior-dependent.
• The site column is robust to everything on the cosmological side: future star formation, K-dwarf windows, the cap prior and the M-dwarf prior each move P(none other ever) by under 0.04.
• SIA behaves as Olson and Ord predict: it drives the expansion probability into the prior's tail (stuck median p_exp 4e-16) and leaves claiming expanders at 0.018. The draft's statement that the SIA split is set by priors is correct.

Verdict on each headline claim

• Claim 1: reword
  • Reason: True of the class table under the site convention, and stronger than stated: what is removed is claiming expansion from any of about 1e6 galaxies at 0.1c or 1e8 at 0.5c, not only from the Milky Way. But "discriminate among everything else by less than a factor of 2" is false under SSA for the parameters that matter: the host and date datum moves P(planets die within 15 Gyr) from 0.50 to 0.72 (0.97 in the Carter camp) and spans a factor of 550 between a G-only world and an n=4, 30 Gyr K-dwarf world.
  • Corrected wording: The silence rules out one thing: expansion that claims systems, whether it starts in our Galaxy or in any of roughly 4e5-1.5e6 galaxies at 0.1c (about 1e8 at 0.5c). Among the remaining classes the observations discriminate by less than a factor of 2 under the site convention; under SSA our host star and date discriminate strongly among hard-step and planet-lifetime assumptions, though not among the classes.
• Claim 2: reword
  • Reason: The three answers are not findings of equal standing. (a) The 0.33 end is the K-dwarf window assumption (0.52-0.53 without it, 0.49 date-only). (b) is the tail of the expansion prior (median p_exp 5e-11; 0.045 under the brief's prior; 0.01-0.49 across camps). (c) scales with p_nonevident. A fourth answer is missing: P(expanders hold at least 1% of the universe) is 0.18 / 0.25 / 0.42, and claiming expanders en route range from 0.03 to 0.32 across defensible priors. With the brief's intergalactic speed prior and q_ig 0.9 the split is 0.55 / 0.13 / 0.35, not roughly equal.
  • Corrected wording: Four live answers. (a) Nobody else has ever arisen in the Galaxy: 0.5 (0.33 if late-K-dwarf planets are long-lived and our star is treated as a random draw; 0 under SIA). (b) Others arose but fewer than one in a million ever expands (a filter ahead of us): 0.05-0.33, set by the prior tail on expansion. (c) Expansion happened and is not evident from here: 0.2-0.3 (0.6 under SIA), proportional to the prior on low-footprint settlement. (d) Expanders that claim systems hold at least 1% of the universe and have not reached us: 0.03-0.04 on base priors, up to 0.3 with several hard steps, long-lived planets and SSA. Expanders of any kind hold at least 1% of the universe with probability 0.18 / 0.25 / 0.42.
• Claim 3: keep
  • Reason: Site numbers reproduce (0.44, 0.34, 0.51, 0.53). Add the cosmic row, which is a larger effect than the surveys: ignoring other galaxies' expanders moves stuck 0.25 -> 0.32 and unnoticed 0.19 -> 0.14. Do not publish the SSA columns of the ablation rows without the local term; they are artefacts (excluded = 1.000).
  • Corrected wording: Add: "Allowing expanders from other galaxies moves 0.07 out of 'stuck' and 0.03 out of 'rare' into 'unnoticed' and 'en route'; it matters more than the galaxy surveys."
• Claim 4: reword
  • Reason: Correct and understated. If evident expanders also cross between galaxies the bound tightens by N_reach: with 2.5-2.9e9 planets per galaxy and 4e5-1.5e6 galaxies at 0.1c the 95% limit is about 1e-15 per planet (about 1e-17 at 0.5c), consistent with Kipping 2026's one spawn per million galaxies. Minor: the model's local term multiplies N_ever by the head-start fraction, an age cut applied twice (effect under 0.004).
  • Corrected wording: Append: "If such civilisations also cross between galaxies, the limit is about 1e-15 per planet at 0.1c and about 1e-17 at 0.5c, because any of 4e5-1.5e6 (or about 1e8) galaxies could have sent them."
• Claim 5: keep
  • Reason: Not re-derived from this lens. In the model these limits enter through the cold curve (130-1,200 effective galaxies) and move no class; the cosmic visibility term built on them has no leverage (under 0.003). Worth adding the occupancy reading: temperature-independent limits bind only for slow expanders, since above about 0.3c fewer than 0.5 unarrived domains are visible per unit x.
• Claim 6: keep
  • Reason: Outside this lens; not examined beyond confirming that the SETI term has no effect on any expander quantity (term mean 0.99-1.00 in every class).
• Claim 7: reword
  • Reason: The 0.33 -> 0.14 figure reproduces, but three wordings contradict the study and the mechanism is unstated. "Sun-like hosts only, rank 42-71%" is the all-stars 10 Gyr cap case; Sun-like only is 66% to 2.2% for n = 1 to 10. "First 1% for four or more" holds only at the 0.6 Msun cut (1.0%, 2.8%, 6.6% for cuts 0.6, 0.65, 0.7). A cap dissolves the rank puzzle but not the weight the model applies (short-lived planets still leave SSA 0.41 against site 0.54). The response runs through fewer hard steps (1.08 -> 0.51), not through expanders (0.037 -> 0.043), and is 0.49 -> 0.35 under a date-only SSA and 0.90 -> 0.67 with Carter steps.
  • Corrected wording: "We are early" depends on planets and on hard steps, not on M dwarfs alone. If every habitable lifetime is capped near 10 Gyr, humanity is not early (rank 42-71%). With star-limited lifetimes and hosts down to 0.6-0.7 Msun, four hard steps put us in the first 1-7% and six in the first 0.1-1.5%. JWST's 0 of 14 removes the puzzle only near one hard step. In the model, learning that planets stay habitable beyond 15 Gyr lowers the SSA P(none other ever in the Galaxy) from 0.33 to 0.14 (0.49 to 0.35 if only our date, not our host star, is conditioned on), mainly by making hard steps improbable (mean 1.1 -> 0.5); if several hard steps are held fixed, the same observation instead raises the probability that claiming expanders hold at least 1% of the universe fourfold (0.02 -> 0.11, or 0.08 -> 0.32 with expansionist priors).
• Claim 8: reword
  • Reason: Not re-derived. From this lens the claim needs one qualifier: under SSA the host-star and date datum is worth far more than one bit (a factor 550, about 9 bits, between a G-only world and an n=4 world with 30 Gyr K windows), and it lowers mean hard steps from 1.71 to 1.08. "One bit" is true of the within-planet timeline only.
  • Corrected wording: Earth's own timeline (early life, late intelligence) is worth about one bit. Under SSA, the star we orbit and the date we appear are worth much more, but only through assumptions about long-lived K-dwarf planets.
• Claim 9: keep
  • Reason: Outside this lens. It matters here because the number of hard steps decides whether long-lived planets imply expanders or a populated Galaxy (objection on claim 7); the draft should make that link.
• Claim 10: reword
  • Reason: The listed rows reproduce. Two corrections: M-dwarf verdicts do not change the answer even through the earliness argument in this model (SSA 0.334 -> 0.315 / 0.340). The largest pending levers from this lens are missing: whether late-K-dwarf planets are habitable for 20-45 Gyr (SSA 0.33 against 0.53), future star formation (SSA en route plus imminent-arrival worlds 0.06 -> 0.25 at the Behroozi and Peeples bound), and a detected expanding domain.
  • Corrected wording: Add: "Evidence on how long late-K-dwarf planets stay habitable moves the SSA P(none other ever in the Galaxy) between 0.33 and 0.53. M-dwarf verdicts move it by 0.02. A much larger future of star formation than the standard extrapolation would raise claiming expanders en route from 0.05 to 0.14-0.25 under SSA."
• Claim 11: reword
  • Reason: The SIA half stands. The site-to-SSA half is misattributed: the 0.53 -> 0.33 gap disappears if K dwarfs get the G window (0.53) or if only our date is conditioned on (0.49). It is an astrophysical assumption expressed through an anthropic convention, and it can be argued about with data.
  • Corrected wording: Anthropic convention is the largest lever. Under SIA emptiness is ruled out and the answer becomes "they exist; expansion fails or is not evident", with the split set by priors. The difference between the site and SSA columns is not anthropics alone: it comes from treating our Sun-like host as a random draw in a universe where late-K-dwarf planets stay habitable four to eight times longer, and it vanishes if they do not.

Code errors

• models/model3.py / lines 191-196 (d_eta, t_vis, Kv, mu_vis)
  • Problem: The count of visible-but-not-arrived domains evaluates the kernel at eta0 + R(1-v)/v, clipped at the last grid point. Conformal time is bounded (16.57 Gly ahead), so the clip binds for v < 0.075c and under-counts by 0.4 dex (0.03-0.076c) to 6.8 dex (below 0.001c). Above 0.075c the later-time kernel also counts civilisations born after today, over-counting by up to 3x (95th percentile) in worlds dominated by K and M hosts.
  • Fix: Integrate over births up to T0 only: mu_vis = RHO_GAL * p_exp * ig * integral of births(t') * (4 pi / 3) * [min(vd + R(1-v), d)^3 - (v*d)^3] dt', d = eta0 - eta(t').
  • Impact: No class moves by more than 0.003 (t_muvis.py). Real error, negligible consequence.
• models/model3.py / line 178 (mu_loc = N_ever * phi * p_exp) and lines 233-234 (nlogS local term)
  • Problem: N_ever already excludes civilisations not yet born; multiplying by phi (0.69-0.72) applies the age cut twice. The SSA numerator then uses exposure phi * p_exp * N while the denominator uses p_exp * N(t) with no delay.
  • Fix: Use F(T0 - T_fill)/F(T0) from the tables in both numerator and denominator.
  • Impact: With phi = 1: site classes change by at most 0.004; SSA "excluded" 0.016 -> 0.012.
• models/run3.py / ABLATION rows with use_local=False, SSA columns (model3.py lines 225-236)
  • Problem: The SSA numerator carries the survival factor only through L_local. With the local term off the numerator is untruncated and the denominator truncated, so the SSA weight grows without bound in heavily settled worlds.
  • Fix: Put S(T0) inside the SSA weight, or skip the truncation when use_local is False, or do not store SSA columns for those rows.
  • Impact: Stored SSA rows "Prior", "+ Earth's timeline only" and "+ galaxy surveys" read excluded = 1.000, mean p_exp 0.43, effective sample size 2,800-5,200. Draft claim 3 uses the site column, so it is unaffected.
• models/model3.py / line 156 (lidx = nearest L_GRID value) with lines 157-158
  • Problem: W uses the exact window (W_K scales as L^n) while F0 and K0 come from the nearest grid window (scaling as L_grid^-n), so the K and M contributions to N_ever and x are off by (L/L_grid)^n.
  • Fix: Multiply F0, K0 and b(T0) by (L_grid/L)^n when both exceed T0, or interpolate in log L.
  • Impact: Error factor for n >= 3: 5-95% range 0.59-1.7, maximum 2.8 (K) and 5.6 (M). Not tested end to end; small against a prior spanning tens of decades, and D_all is unaffected because F(infinity) = 1.
• models/cosmo_tables.py / lines 66-73 (birth law)
  • Problem: n = 0 has a fixed 4.5 Gyr delay; n >= 1 has no minimum delay, so civilisations can arise immediately after planet formation. N_reach jumps from 4.1e5 (n=0) to 1.48e6 (n=1). Gated worlds with slow abiogenesis use the n = 1 row although class_chain requires abiogenesis in the first quarter of the window followed by a fixed delay.
  • Fix: Density n (tau - d)^(n-1) / (L - d)^n with d of 1-3 Gyr for n >= 1; a dedicated row for gated worlds with slow abiogenesis.
  • Impact: x differs by up to 3.6x in n = 1 worlds; scaling x by 3 or 1/3 moves no class by more than 0.007.
• models/cosmo_tables.py / build(): division by sfr_total over 0.05-3000 Gyr; sfr() future extrapolation
  • Problem: Births are normalised to all-time star formation (90% done) while Nstars is today's count, so N_ever is 10% low and N_alltime omits future stars. The future is a single extrapolation (future/past 0.11) although the project's own study cites 0.02-10.
  • Fix: Normalise by star formation to date; make future/past a sampled parameter with tables per value.
  • Impact: Negligible at base. With a larger future the site column drifts (none-ever 0.532 -> 0.572 at 10x) unless renormalised; SSA en route 0.046 -> 0.060 (1x) and 0.141 (10x).
• models/model3.py / lines 174-175, 179 (v, ig, kv)
  • Problem: The intergalactic flag is drawn independently of speed, and the same speed (down to 1e-4 c, 30 km/s) is used for crossing between galaxies. Below about 0.001c N_reach is under one galaxy and the smooth-density kernel is meaningless.
  • Fix: Separate v_ig prior; ig requires v_ig >= 0.003c.
  • Impact: Guard alone: P(x >= 1%) 0.179 -> 0.127 site, 0.249 -> 0.168 ssa; en route 0.041 -> 0.029. With the brief's uniform 0.1-0.99c: P(x >= 1%) 0.25 site, 0.32 ssa.

Missing explanations

• Home-grown deadline: we are among the first in our own Galaxy and someone, possibly us, claims it before most long-lived-star civilisations arise. The D_all code contains it (local truncation term) but it has no class and no sentence. SSA posterior with truncation by home-galaxy expanders only: 0.015 at base, 0.028 with long-lived planets, 0.080 with Carter steps and long-lived planets.
• The grabby deadline explains our date but not our star. At four hard steps, old K-dwarf planets produce 43 times as many civilisations today as G planets (209 times at six), so a deadline falling today still leaves Sun-like hosts at 14% (4%) of observers. This is Kipping's red-sky argument extended to K dwarfs and it is a genuine finding of the merged model; I have not seen it stated this way in Hanson et al. or Cook (from memory).
• Late-K-dwarf planets are not long-lived habitats (UV too weak for prebiotic chemistry below K5, tidal locking, volatile loss). There is no parameter for it, and it alone decides whether the SSA column differs from the site column.
• Fast claiming expanders (v at least 0.1c) as a named world, separate from slow ones: 0.007 site, 0.008 ssa at base. The current en-route class mixes them with domains moving at galactic orbital speeds.
• A much larger future of star formation than the Madau and Dickinson extrapolation (Behroozi and Peeples). It makes us early even on G hosts with no hard steps and is the only lever that raises the en-route class without touching the site column.
• The Great Filter reading: a filter ahead of us of six or more orders of magnitude has posterior 0.36-0.42, and it is the content of the 'stuck' class. The draft never uses the frame.
• Correlated or synchronised timing (Annis; Cirkovic and Vukotic), asked for in round 1: still no variable, so 'we are among the first' can only arise by chance or by the deadline.
• A delay between becoming technological and starting to expand (Hanson's quiet phase). Expanders start at birth in both the local and cosmic terms; a quiet phase of 1e6-1e8 yr changes nothing cosmologically but should be stated.

Critic’s own estimate of each answer

• Empty universe: 0.27
• Rare: 0.22
• En route: 0.08
• Stuck: 0.13
• Unnoticed: 0.2
• Hidden: 0.03
• Other: 0.07
• What “other” covers: Outside the taxonomy: model misspecification (host structure, discrete galaxies), correlated timing, simulation or unknown physics. My numbers are a blend of the base run and a variant with W_exp 1-10, intergalactic speed uniform 0.1-0.99c, q_ig 0.9 and low-footprint prior 0.10 (site 0.395/0.195/0.081/0.029/0.241/0.054; date-only ssa 0.370/0.166/0.107/0.035/0.241/0.054), counting the 'contradicted' SSA mass as en route.

Presentation advice

Lead the expansion section with one number per convention that the draft currently omits: P(expanders hold at least 1% of the universe) = 0.18 / 0.25 / 0.42, then split it by regime (claiming 0.03 / 0.04 / 0.02; low-footprint or hidden 0.15 / 0.21 / 0.40) and by speed (v at least 0.1c: 0.05 / 0.07 / 0.10). Say plainly which rows are priors restated: en route (prior 0.033, posterior 0.041), stuck (the tail of the expansion prior below 1e-6), unnoticed (proportional to the 0.25 low-footprint prior). Show the SSA column twice, date-only and date-plus-host, and name the K-dwarf window assumption next to it. Add a short "prospects for humanity" box: median model-implied chance of ever becoming a sustained expander 4e-4 (site) or 1e-4 (ssa), mean 0.09; P(fewer than one civilisation in a million ever expands) 0.36-0.42; P(no other civilisation ever in the volume we could reach) 0.23-0.27 (site), 0.14-0.16 (ssa); P(no expander has yet started in it) 0.61-0.66, 0.50-0.54, 0.22-0.25 (sia); median time to an even chance of meeting claiming expanders, where they exist, 13-15 Gyr, with one in five such worlds never meeting us. Add the expansionist camp table from the sixth objection and state its lesson: expansionist priors give Hart and Tipler (alone in the Galaxy at 0.84) or ubiquitous low-footprint expanders (0.47-0.63), and give Hanson's en-route world (0.25-0.32) only with several hard steps, long-lived planets and SSA together. Rename the "contradicted" row or redefine it so that it cannot carry posterior mass. Print the speed prior beside any statement about speed; the posterior P(v > 0.3c) among claiming en-route worlds equals the prior.

Were the round 1 fixes made?

Fixed and correct: (1) expanders from other galaxies now pre-empt us through x = RHO_GAL v^3 p_exp q_ig sum(W K0); the kernel and N_reach check out against Kipping 2026. (5) The Bayes factor is gone. (6) The feasibility, doom and will atoms are replaced by one continuous prior on p_exp. (10) Carter terms count only steps that are still hard (p < 0.3 over the window). (9, partly) An en-route class exists and hidden or unnoticed arrivals may come from another galaxy.

Fixed but mislabelled or over-corrected: (2) A birth-date and host likelihood exists and reproduces the study's ranks, but it is implemented as the G-host share, so the date contributes nothing (none-ever 0.515 with host not conditioned, against site 0.532) and the host does everything. (3) K dwarfs were added as round 1 asked, but as a certainty with a step at 0.75 Msun and a fixed 5.6 Gyr for every G-class planet; that one assumption now produces the entire site-to-SSA gap. (4) Two anthropic branches exist, but the SSA branch's deadline does nothing at base (classes move by at most 0.012 with it switched off). (6) The continuous p_exp prior reaches 1e-30, which round 1 asked for, but the draft now reports its tail (the stuck class, median 5e-11) as a finding instead of labelling it as prior, which round 1 also asked for. New hand-set atoms have replaced the old ones: p_cap_short 0.5, q_ig 0.6, p_nonevident 0.25.

Not fixed: (1, part) the speed prior: round 1 asked for uniform 0.1-0.99c or log-uniform 0.01-0.99c; version 3 uses log-uniform from 1e-4 c for intergalactic travel, and q_ig is 0.6 instead of about 0.8. (7) Humanity's prospects: mean_p_exp is in the results file but nothing is in the draft, and P(first expanders in our reach volume) is not computed. (8) The visibility module is right only above 0.075c and the speed prior is not printed beside speed statements (harmless numerically). (9) No meeting-time distribution; "empty universe" still uses the observable universe rather than the reach volume (0.20 against 0.23-0.27 for the reach volume, so the difference is small); P(expanders hold at least 1%) is computed but not reported per camp in the draft. Correlated timing is still absent.

Critic: claims of novelty and the citations

Overall verdict

The draft headline claims are not defensible as written from this lens, mainly because FINDINGS_DRAFT.md (02:52) predates the revised study reports (05:02) and still carries first-pass numbers and novelty claims that the project's own referees and revised reports withdrew. Claims 4, 5, 7, 8 and 9 each contain at least one statement the corresponding report.md now contradicts. The worst is claim 5: "the first limits that do not depend on the waste-heat temperature" is false, because Annis 1999 and Zackrisson et al. 2015 published temperature-independent optical limits and the project's own Tully-Fisher limit weakens above about 250 K. Most "original data" tags should read "reproduction with new data" or "new in degree"; the three items that look new after my searches are the fundamental-plane search at 1e5-galaxy scale, the apparent e^-k / 10^-k slip in Snyder-Beattie et al. 2021 Tables 2-3, and the clustered reanalysis of Vermeij 2006. One model input is mis-cited: warm_curve applies Lacki 2016's limit for screens just above CMB temperature (an unrefereed preprint) to waste heat at 100-600 K. None of this moves the posterior by more than about 0.02, and I tested that removing the project's optical limits or rescaling the head-start fraction moves P(no other civilisation ever in the Galaxy) by 0.001. The single most important change is to rewrite section C from the revised study reports with a prior-work credit on every claim, and to say plainly that the original studies check inputs and do not drive the answer.

Objections

• major / FINDINGS_DRAFT.md section C, claims 4, 5, 7, 8, 9 (stale against studies/*/report.md)
  • Argument: The draft quotes first-pass results that the revised reports withdrew. Claim 4: '11.1 +- 0.5' is now 11.1 +- 1.0; '72.5% (69 to 76%)' is now 'about 70%, range 0.64-0.78, does not test Lineweaver more sharply'; '1e-9' is 1.0-1.2e-9 at the fiducial planet count and 3.4e-9 to 7.9e-9 marginalised. Claim 5: 'about 2%' is now 2.6% (1.9-3.5%), and tf_kiii says the first-pass headline '(1.8%, 0.90%, any temperature) does not survive'. Claim 7: 'rank 42 to 71%' is the 10 Gyr cap, all-stars case, not 'Sun-like hosts only' (hosts above 0.8 Msun give 21% at n=4); 'one or two hard steps' should be one (critical M-dwarf weight 0.053 for one step, 5.3e-5 for two); 'four or more' should be five or more (n=4 gives 1.03% at a 0.6 Msun cut; smallest n is 5, 6, 7, never for cuts 0.6, 0.65, 0.7, 0.8). Claim 8: timing withdrew 'ceiling' ('neither a floor nor a ceiling'; Kipping's 13 is a lower limit in his own model) and withdrew '4.6 to 942' as a division-by-near-zero artefact. Claim 9: 'clustered p = 0.12' is one of six numbers (Fisher 0.125 at 543 Ma, 0.040 at 1 Ga; regression p = 0.47; two other groupings give 0.008), and 'do not support' was replaced by 'cannot test'.
  • Proposed change: Regenerate section C from the revised report.md headline paragraphs; exact replacement text for each claim is in claim_verdicts. Add a build check that every number in the findings matches a key in studies//result.json or data/report_numbers.txt.
  • Expected effect: No change to model output. Wording and quoted study numbers change in five of eleven claims; quoted uncertainties roughly double for the flyby rate and the head-start fraction; the Tully-Fisher limit rises from about 2% to 2.6%; the per-planet limit gains a 3e-9 to 8e-9 marginalised version.
  • Tested: Compared every number in claims 4-9 against the eight report.md files; the revision notes in each report name the withdrawn figures. File times: FINDINGS_DRAFT.md 02:52, reports 05:02.
• major / Claim 5: 'These are the first limits that do not depend on the waste-heat temperature' and the title 'No star-harvested galaxies at any waste-heat temperature'
  • Argument: Three errors. (1) Not first: Annis 1999 (JBIS 52, 33-36) used both the Tully-Fisher relation (57 discs) and the fundamental plane (106 ellipticals) and found no outlier dimmed by 1.5 mag; Zackrisson et al. 2015 (ApJ 810, 23) set under 3% (conservative) and 0.3% (tentative) from 1,359 discs. Both are optical and temperature-independent. The tf_kiii referee said so and the revised report withdrew the 'closes the cold loophole' claim. (2) Not 'any temperature': the Tully-Fisher limit is at full strength only below about 250 K (3.1% at 300 K, about 6% above 400 K) because warm waste heat fills in the 3.4 micron band. (3) Not 'star-harvested': fp_kiii says the limits are 'on light, not stars'; cloaking all dwarfs and subgiants removes only 20-36% of the light (Lacki 2019), where the limit is 7-17% of galaxies or absent, and it withdrew the claim that the test covers star-harvesting generally. Also 'no' should be 'no confirmed': 34 Tully-Fisher and 23 fundamental-plane outliers are unexplained, with an 11-to-1 dim/bright asymmetry beyond 0.60 dex (p = 0.003). The '1e5 galaxies' has an effective sample of 14,853 at f = 0.75; the '9,400 discs' has 1,766.
  • Proposed change: Novelty label 'new in scale, old method (Annis 1999; Zackrisson et al. 2015)'. Use the corrected claim 5 wording in claim_verdicts. Cite Curtis et al. 2026 (arXiv:2608.12458, G-HAT V, submitted to ApJ), which two months earlier proposed a stellar-to-dynamical-mass screen and set under 2.6% of galaxies hosting 25% warm swarms from 129 galaxies; fp_kiii does not cite it.
  • Expected effect: Wording only. What survives as new: the first fundamental-plane application at 1e5-galaxy scale with selection-corrected limits and a region search, and a Tully-Fisher limit 1.8 times tighter than Zackrisson's Poisson-equivalent 4.6%, entirely from vetting.
  • Tested: Fetched arXiv:1508.02406 (credits the method to Annis 1999; 3% and 0.3% confirmed), arXiv:1604.07844, arXiv:2608.12458. Search confirmed Annis used 57 discs and 106 ellipticals on both relations. My search for a later fundamental-plane technosignature search found none.
• major / model3.py line 72: warm_curve=((0.01, 3e3), (0.1, 1e5), (0.85, 1e6)) labelled '(Lacki 2016; Griffith 2015)', applied when waste heat is at 100-600 K
  • Argument: The three points are Lacki 2016's partial-shroud curve (1 in 800 at 1%, 1 in 30,000 at 10%, 1 in 810,000 full). Lacki's search is for 'blackboxes' at a temperature 'just above that of the cosmic microwave background', found as Planck 100 GHz sources. It says nothing about 100-600 K. The paper is an arXiv-only preprint with no journal reference. For 100-600 K the published limits are Griffith et al. 2015 (0 of about 1e5 galaxies above 85%; 50 consistent with over 50%) and Curtis et al. 2026 (under 2.6% at 25%); the project's own seti.md says below about 25% is 'unconstrained'. So the warm curve overstates the constraint by 10 times at full coverage and by 2-3 orders of magnitude at 1-10%. The error came from gaps_and_corrections.md section 7 ('Galaxy-scale waste-heat limit: covering-fraction curve') and from round 1's request for 'the Lacki/Griffith curve (warm)'.
  • Proposed change: Use three temperature regimes. Near-CMB (prior of order 0.15): Lacki curve ((0.01, 3e3), (0.1, 1e5), (0.85, 2.5e6)), flagged unrefereed. Mid-infrared 100-600 K: ((0.25, 115), (0.5, 4.8e3), (0.85, 1e5)) from Curtis 2026 and Griffith 2015, zero below 0.25. Far-infrared 10-100 K: the existing fundamental-plane plus Tully-Fisher curve. Fix the comments on lines 72-73 and gaps section 7.
  • Expected effect: Site P(no other ever in Galaxy) 0.531 -> 0.523 (three regimes) or 0.521 (warm curve replaced only); SSA 0.332 -> 0.325. 'Empty universe' 0.196 -> 0.188, 'stuck' 0.250 -> 0.258. The surveys' pull without the local term falls from +0.049 (0.345 -> 0.394) to +0.023 (0.345 -> 0.368), so about half of the survey term's leverage was the mis-attributed curve. Claim 3's 'galaxy surveys add little' gets truer.
  • Tested: Ran model3 at n = 1e6 with the replaced warm curve and a patched copy with three regimes at n = 6e5 (scratch: t_curves.py, model3b.py, t4.py). Numbers above are from those runs.
• major / The '(original data)' tags on claims 4, 5 and 7, and the reader's request for 'original research... novel insights'
  • Argument: The original studies have almost no leverage on the model, and the draft does not say so. Removing the project's optical limits (cold_curve) changes site P(none) from 0.5319 to 0.5309. Rescaling the head-start curve across the revised range (0.64/0.72) changes it from 0.5319 to 0.5313. Radio SETI changes it by 0.00. The answer is set by priors, Earth's timeline and the 'nobody evidently here' term. Separately, claim 4's contents are published: the rate (11.7 +- 1.3, Garcia-Sanchez et al. 2001; 10.6 +- 4.5, Fernandez-Puig et al. 2026, A&A 710, A252); the exposure-count Fermi argument (Zuckerman 2026, arXiv:2603.07333, section VII: n V pi R^2 T with V = 48 km/s gives about 2 million old solar-type stars within 100 light-years in 2 Gyr, and the absence of probes is the inference); waiting for close passages (Hansen & Zuckerman 2021); the head start (Lineweaver 2001: 74 +- 9%; Madau 2023; Legassick 2015); and the per-planet limit, which headstart calls 'standard Hart-Tipler arithmetic'. Kipping 2026 gives a cosmological version about seven orders tighter per planet.
  • Proposed change: Relabel claim 4 'reproduction with new data', claim 5 'new in scale', claim 7 'new in degree'. Add a short table per study: closest prior result, this project's number, what is new, and the change in P(none) when the study's input is removed. State in the summary that the original data confirm the inputs to Hart's argument at higher precision and leave the answer where the priors put it.
  • Expected effect: No numeric change. The report stops implying that its data studies produced its conclusions and gains a table whose last column is 0.001, 0.001, 0.00 and at most 0.02.
  • Tested: Model runs at n = 1e6 with cold_curve set to zero and with headstart_phi scaled by 0.64/0.72 (t2.py). Fetched Zuckerman 2026 full text, Fernandez-Puig 2026 abstract, Hansen & Zuckerman 2021 abstract; extracted '74 +- 9%' from the Lineweaver 2001 PDF.
• major / Claim 7: '"We are early" depends on planets, not on M dwarfs alone (original data)... The decisive unknown is how long rocky planets stay habitable (a cap near 10 to 15 Gyr dissolves the puzzle)'
  • Argument: Both halves are published. Hanson et al. 2021's abstract already says over 99% of advanced life appears after today 'unless n<3 and max planet duration <50Gyr'; mdwarf_early lists 'K dwarfs suffice' and 'not early with a 10 Gyr cap' under 'Not new... both confirm Hanson et al.'. Kipping 2025b (arXiv:2510.01215, unrefereed) frames the same two levers, a critical stellar mass and a planetary window T_win, and finds that 'reducing T_win alone doesn't adequately explain the data': a low-mass cutoff is still required (M_crit > 0.34 Msun at 2 sigma, > 0.74 Msun at 1 sigma). The study agrees: a cap 'leaves our G-type host unexplained' by a factor of 7.6-9.4. So 'dissolves the puzzle' is too strong; the cap dissolves earliness in time and leaves the host-type puzzle. Also the M-dwarf evidence is thinner than the sentence implies: none of the 14 planets is in the habitable zone and habitable-zone retention is 'unconstrained (0.09-0.68)'.
  • Proposed change: Credit Hanson et al. 2021, Kipping 2025b, Loeb et al. 2016 and Cook 2022 in the claim. State as new only the critical M-dwarf weight per number of hard steps (0.053, 5.3e-5, 5.3e-8) and the link from the atmosphere census to the earliness calculation. Replace 'dissolves the puzzle' with 'removes the earliness but leaves our G-type host unexplained by a factor of about 8'. Give p_cap_short = 0.5 a literature basis or label it a judgement; the planet_lifetime study folder is empty.
  • Expected effect: Wording. The 0.33 -> 0.14 SSA result stands as a model output (matches the results file: 0.143).
  • Tested: Fetched the abstracts of arXiv:2102.01522 and arXiv:2510.01215; checked the claim's numbers against mdwarf_early/report.md key-number rows.
• major / Claim 8: 'Snyder-Beattie et al.'s tables reproduce only with thresholds e^-k where 10^-k is printed, which halves their claimed weight on extreme rarity' and 'the generalisation of Kipping's ratio'
  • Argument: The units finding looks right and looks new, but the draft states as fact an inference about a named paper whose authors have not been contacted. I confirmed from the PMC text that Table 2's columns are printed as 10^-2, 10^-12, 10^-24, that the prior is log-uniform on expected transition time from 1e-14 to 1e10 Gyr, and that the text says 'over 90%... less than a 10^-12 chance' for three steps. A hand check supports the study: with both steps at the prior's slowest rate and a window near 5 Gyr, the minimum two-step probability is about (5e-10)^2 / 2 = 1.3e-19, so no posterior weight can lie below 10^-24, yet the printed two-step rows put 17-39% there. I found no erratum or published comment. Separately, the multi-step generalisation of Kipping's ratio is, in the study's own words, 'a corollary of hard-steps order statistics' (Carter 1983; Hanson et al. 2021), 'not new as a result'.
  • Proposed change: Write 'an apparent natural-log/base-10 labelling slip, inferred by reproducing the tables (rms 0.02 dex against 0.57 dex) and confirmed by an independent re-implementation; the authors have not been contacted and no erratum was found as of October 2026; the paper's qualitative conclusion stands'. Notify the four authors before the report circulates. Describe the multi-step values as a corollary, not a generalisation.
  • Expected effect: Wording. The numerical statement (three-step weight below 10^-12 falls from over 90% to 44-58%) is unchanged.
  • Tested: Fetched PMC7997718 (table headers, prior, text sentences); hand calculation above; web search for an erratum or comment returned none.
• moderate / Claim 9: 'Convergence data do not support many hard steps... Vermeij's result is carried by three stem lineages (clustered p = 0.12)'
  • Argument: The direction of inference is unstated and partly backwards. Vermeij 2006 found singular innovations older than repeated ones and read it as information loss: apparent uniqueness is an artefact of age. Mills et al. 2025 use exactly that against hard steps ('not because they are improbable'). The reanalysis shows the age effect is not robust once LUCA, LECA and body-plan traits are counted once (odds ratio 3.2 -> 1.4, interval 0.5-5.0). That does two opposite things: it cuts the number of independent one-off events to a few, and it weakens the statistical support for the information-loss argument against hard steps. Presenting it under 'do not support many hard steps' hides the second. The single p-value is also selective (see the first objection). The three clusters are trait clusters, not 'stem lineages'; the stem-lineage grouping is a different collapse with p = 0.26.
  • Proposed change: State both consequences and all three groupings. Retitle 'Convergence data cannot establish many hard steps and cannot rule them out'. Credit Bains & Schulze-Makuch 2016 for classifying innovations by number of origins and Mills et al. 2025 for the argument being tested.
  • Expected effect: No numeric change in the model. The 'convergence-study steps' camp should no longer be presented as data-derived: the study calls its 1-2 hard steps a judgement.
  • Tested: Fetched Vermeij 2006 (PMC1413639) and Mills et al. 2025 (PMC11827626) to confirm how each uses the age effect; one web search for a prior non-independence critique of Vermeij's test found none.
• moderate / Claims 1 and 11, and section B's prior column
  • Argument: (1) Arithmetic: the 'evident settlement' prior is 0.309 in integrate3_results.json and in my rerun, not 0.27; the draft's prior column sums to 0.96; claim 1's '27% of prior mass' should be 31%. Mean likelihood for 'en route' is 0.41, not 0.44. (2) Novelty: 'the silence excludes only evident settlement' is published in qualitative form. Carroll-Nellenback et al. 2019's abstract says their results 'break the link between Hart's famous Fact A' and human uniqueness and that Earth 'might remain unvisited in the midst of an inhabited galaxy'; Freitas 1985 (Icarus 62, 518; from memory) and Haqq-Misra & Kopparapu 2012 argue that unsearched artefacts are not excluded; 'any confident answer is a statement of priors' is the thesis of Sandberg, Drexler & Ord 2018. (3) Claim 11's 'largest lever' holds only for SIA. Between site and SSA the answer moves by 0.20 (median 0.21 across the scan), while priors move it from 0.26 to 0.74 across the 400-draw scan and from 0.08 to 0.95 across camps. The SIA result itself is in Olson & Ord 2021, Finnveden 2019 and Cook 2022.
  • Proposed change: Fix 0.27 -> 0.31 and 0.44 -> 0.41. Credit the prior work in claims 1 and 11 and present the quantification as the new part. Reword claim 11 as in claim_verdicts.
  • Expected effect: Prior column sums to 1.00. Claim 11 no longer ranks anthropics above priors except at the SIA end.
  • Tested: Reran model3 at n = 1e6 (prior 'excluded' 0.309; mean likelihoods 0.353, 0.497, 0.411, 0.572, 0.604, 0.677, 0.004). Computed scan quantiles from integrate3_scan.json: site P(none) 5-95% is 0.255-0.742; site minus SSA is 0.125-0.298.
• moderate / Claim 6 '(original synthesis)' and model3.py line 78: c_det=(1e-12, 1e-5) '[STUDY seti_meta]'
  • Argument: Novelty: 'radio SETI constrains behaviour, not abundance' is Wright, Kanodia & Lubar 2018 (the hot-tub result) and Sheikh et al. 2025; population bounds by broadcast luminosity for the Milky Way are in Lacki 2024 (ApJ 966, 183) and Lacki 2025 (PASP 137, 114103: Kardashev III radio broadcast populations rarer than about one in a million large galaxies); in-beam star counts are Wlodarczyk-Sroka et al. 2020 and Mason et al. 2025. The study's revised novelty statement is narrower than the draft's: one de-duplicated coverage function and the conditional structure of the 21-cm limit (a factor of 63). Parameter: the study's per-transmitter detection probabilities run from 8.7e-18 (randomly pointed radars) through 6.9e-6 (1e17 W beacons) to 0.026-0.5 (Kardashev-II beacons). The prior's cap at 1e-5 excludes the study's own loud-beacon class, so 'radio SETI adds nothing' is partly built in.
  • Proposed change: Label claim 6 'original synthesis of a known conclusion' and cite the five papers. Either extend c_det to the study's full range or state that the base prior excludes loud beacons and quote the sensitivity.
  • Expected effect: With c_det up to 0.03: site P(none) 0.532 -> 0.544, SSA 0.334 -> 0.348. Up to 0.5: 0.553 and 0.359. Claim 3 should read 'radio SETI adds at most 0.02'.
  • Tested: Ran model3 at n = 6e5 with c_det=(1e-12, 3e-2) and (1e-12, 0.5) (t3.py). Fetched arXiv:2405.04651, 2508.00249, 1809.07252, 2502.02614.
• minor / Sources that carry weight but are unrefereed or dated 2026
  • Argument: A reader checking citations will find these have no journal reference as of the arXiv pages I read: Lacki 2016 (arXiv only, sets the strongest galaxy curve); Kipping 2025b (submitted to AAS Journals); Kipping 2026 (submitted to Astrobiology; sets the galaxy density 0.01 per Mpc^3 behind RHO_GAL); Curtis et al. 2026 (submitted to ApJ); Olson & Ord 2021; Sandberg, Drexler & Ord 2018 (the arXiv page still says submitted to Proc. R. Soc. A); Whitmire 2025. Items I could not check against a primary source: the JWST '0 of 14' census as of October 2026 (hand-curated soft labels; the study adds 3 of 17 if ambiguous cases count and 'none in the habitable zone'); the journal reference 'ApJ 1001, 167' for Zuckerman 2026 (the arXiv comment says only 'ApJ published'); Fernandez-Puig et al.'s attribution of the 2018 excess to spurious parallaxes (not in the abstract); Annis 1999 primary text; Tyrrell 2020 (paywall redirect); Madau 2023's 77% (not in the abstract).
  • Proposed change: Mark each unrefereed source as a preprint where cited. Quote the M-dwarf census as '0 of 14 with a decisive verdict (3 of 17 if ambiguous results count), none in the habitable zone'. Distinguish the two Annis 1999 papers (JBIS 52, 19 on gamma-ray bursts; JBIS 52, 33 on scaling relations).
  • Expected effect: None on numbers.
  • Tested: Fetched arXiv abstract pages for 1604.07844, 2510.01215, 2606.04044, 2608.12458, 2106.13348, 1806.02404, 2503.18217, 2603.07333, 2605.16496; all exist and match the briefs' descriptions.
• minor / model3.py line 54: f_persist=(0.03, 1.0) '(Tyrrell 2020)'; line 38: p_M_class=0.30 '[STUDY mdwarf_early: retention mean 0.34]'; line 73 comment 'any temperature'
  • Argument: The prior's lower bound of 0.03 excludes the per-run success rate the project's gaps brief verified from Tyrrell 2020 (1.45% of runs stayed habitable for 3 Gyr), and that brief recommended log-uniform 0.01-1. The 0.34 is described by its own study as 'a bracket for propagation, not an estimate', with habitable-zone retention 'unconstrained (0.09-0.68)'. 'Any temperature' on the cold curve repeats the wording the Tully-Fisher study withdrew.
  • Proposed change: Use f_persist=(0.01, 1.0) or cite why 0.03. Change the comments to 'judgement, bracketed by mdwarf_early 0.09-0.68' and 'far-infrared regime, below about 250 K'.
  • Expected effect: f_persist lower bound 0.01: site P(none) 0.532 -> 0.542, SSA 0.334 -> 0.340.
  • Tested: Ran model3 at n = 6e5 with f_persist=(0.01, 1.0) (t3.py).
• minor / Missing citations that a specialist would look for
  • Argument: Not cited anywhere in the project as far as I found: Hansen 2022 (AJ; arXiv:2112.00852), a Gaia EDR3 catalogue of 132 unbound close-passage pairs within 100 pc proposed as targets for migration events, which is the nearest observational precursor to the flyby settlement study; Curtis et al. 2026 in fp_kiii; any literature on geophysical limits to planetary habitability behind p_cap_short and L_cap_short (for example Unterborn et al. 2022, ApJL 930, L6, on degassing lifetimes of stagnant-lid planets; content from the search summary only). For claim 10, the direction of the second-origin and biosignature updates is established (Balbi & Grimaldi 2020 and Angerhausen et al. 2025 are in the project's own brief; Bostrom 2008, from memory).
  • Proposed change: Add these to the prior-work lists of flyby, fp_kiii, the planet-lifetime prior and claim 10.
  • Expected effect: None on numbers.
  • Tested: Fetched arXiv:2112.00852 and searched for Unterborn et al. 2022; grep of evidence/.md and studies//report.md for these names.

Claims that survive

• The apparent e^-k / 10^-k slip in Snyder-Beattie et al. 2021 Tables 2-3: I confirmed the printed thresholds and the stated prior from the PMC text, and a hand calculation shows the stated prior cannot give a two-step probability below about 1.3e-19, so printed weight below 10^-24 cannot be literal. I found no erratum and no prior notice.
• The cosmological reach kernel is consistent with published work: using the project's own cosmo_tables with uniform spawning after 4.5 and 7.3 Gyr, half-filling at 0.1c needs 1 origin per 8.4e5 and 2.1e5 galaxies, against Kipping 2026's 1e6 and 240,000. This brackets both the model's N_reach (4e5 to 1.5e6) and the round-1 critic's (1.3e5 to 3e5), so that soft spot is a birth-date effect.
• The encounter rate of 11.1 per Myr within 1 pc agrees with three published values (11.7 +- 1.3, 10.4, 10.6 +- 4.5), and Bailer-Jones et al. 2018's 19.7 +- 2.2 is confirmed as the outlier.
• About 70% of Sun-like stars older than the Sun agrees with Lineweaver 2001's 74 +- 9%, which I confirmed in the paper's text.
• Claim 3's qualitative ablation result (surveys add little, radio adds almost nothing) is robust to literature-correct survey curves and to admitting loud beacons: every variant I ran moved P(none) by 0.02 or less.
• Claim 10's numbers match integrate3_results.json (second origin 0.317, biosignature 0.273, 'empty universe' 0.030 and 0.038, artefact null 'unnoticed' 0.105), as does claim 7's 0.33 -> 0.14.
• No published fundamental-plane technosignature search at scale, no published clustered reanalysis of Vermeij's table, and no de-duplicated cross-survey radio coverage function turned up in my searches; these three stand as new, with limited search depth.
• Every 2025-2026 preprint I fetched exists and says what the project reports: arXiv 2504.05993, 2510.01215, 2606.04044, 2603.07333, 2605.16496, 2608.12458, 2508.00249, 2503.18217.
• The quoted numbers for Griffith et al. 2015, Zackrisson et al. 2015, Lacki 2016, Bailer-Jones et al. 2018, Kipping 2020 and 2025, and Hanson et al. 2021 are correct against their abstracts; the error with Lacki 2016 is where the model applies it, not the numbers.

Verdict on each headline claim

• Claim 1: reword
  • Reason: The prior mass is 0.309, not 0.27. The qualitative point is published (Freitas 1985, from memory; Haqq-Misra & Kopparapu 2012; Carroll-Nellenback et al. 2019), and 'a statement of priors' is Sandberg, Drexler & Ord 2018. The quantification is the new part.
  • Corrected wording: The observations remove one class and barely discriminate among the rest. 'Expansion that claims systems and would be evident here' holds 31% of prior mass and has mean likelihood 0.004; every other class has mean likelihood between 0.35 and 0.69. That the silence constrains only evident, persistent settlement is an old point (Freitas 1985; Carroll-Nellenback et al. 2019), and that confident answers are statements of priors is Sandberg, Drexler & Ord's (2018); what is new here is the size of the effect inside one model.
• Claim 2: reword
  • Reason: 'About 0.5' rounds a 0.33-0.53 range to its top, and 'roughly equal' is true under SSA (0.33, 0.32, 0.31) but not under site (0.53, 0.25, 0.21). The quoted ranges omit the prior scan, which is much wider.
  • Corrected wording: Three live answers. (a) Nobody else has ever arisen in the Galaxy: 0.53 (Earth as a given site) or 0.33 (SSA), 0 under SIA; 0.26 to 0.74 across the prior scan. (b) Others arose but sustained expansion almost never happens: 0.25 to 0.32 (0.06 to 0.36 across the scan). (c) Expansion happened and is not evident from here: 0.19 to 0.28, 0.6 under SIA (0.07 to 0.56 across the scan). Deliberate hiding stays under 0.05 in the base model and under 0.10 in the scan.
• Claim 3: reword
  • Reason: Numbers match the results file. Add that the survey contribution comes from published infrared and microwave limits, half of it from a mis-applied curve, and that the project's optical limits contribute 0.001.
  • Corrected wording: Ablation. Earth's timeline alone pushes toward a populated Galaxy (P(none other ever) 0.44 -> 0.34, because early life favours fast abiogenesis); the absence of evident settlement here pushes back (-> 0.51); the galaxy surveys add 0.01 to 0.02, almost all from published infrared and microwave limits and 0.001 from this project's optical limits; radio SETI adds nothing under the base detectability prior and at most 0.02 if Kardashev-II-class beacons are allowed.
• Claim 4: reword
  • Reason: Stale uncertainties and fraction; every element is published (rate, exposure argument, head start, per-planet arithmetic). The study reports call these reproductions.
  • Corrected wording: Time and distance are not the bottleneck (reproduction with new data). Star systems pass within 1 pc of the Sun at 11.1 +- 1.0 per Myr (Gaia), in line with 11.7 +- 1.3 (Garcia-Sanchez et al. 2001) and 10.6 +- 4.5 (Fernandez-Puig et al. 2026). Following Zuckerman's (2026) exposure argument, about 21,000 (11,000 to 59,000) systems passed within 1 pc since the Great Oxidation and 4,000 to 6,000 in the last 0.5 to 0.7 Gyr. About 70% (64 to 78%) of Sun-like stars now near the Sun's orbit are older than the Sun, by 4.1 Gyr on average (3.4 to 4.4), consistent with Lineweaver's (2001) 74 +- 9%. So roughly 2.5 to 2.9 billion habitable-zone rocky planets (uncertain by a factor of several) had time to settle the Galaxy at any front speed from 1e-4 c to 0.1 c. If settlement would be evident, the per-planet probability of producing a Galaxy-settling civilisation is below about 1e-9 (3e-9 to 8e-9 with the planet count's uncertainty included). This is Hart-Tipler arithmetic with measured inputs.
• Claim 5: reword
  • Reason: 'First' is false (Annis 1999; Zackrisson et al. 2015); 'any temperature' is false for discs above about 250 K; 'star-harvested' was withdrawn by the study; 'about 2%' is now 2.6%; outliers remain unexplained.
  • Corrected wording: No confirmed strongly dimmed galaxy in optical scaling relations (new in scale; the method is Annis 1999 and Zackrisson et al. 2015). Among 114,000 SDSS early-type galaxies (effective sample 15,000) fewer than 0.29% (0.24 to 0.45%; 1.4% under pessimistic assumptions) can be uniformly dimmed by 75% at fixed mass. Among 9,400 HI-rich discs (effective sample 1,800) fewer than 2.6% (1.9 to 3.5%) can be, against 3 to 4.6% from Zackrisson et al.'s 1,359 discs. These limits matter where waste heat emerges in the far-infrared, out of reach of mid-infrared (Griffith et al. 2015) and microwave (Lacki 2016) searches; for discs they hold at full strength only below about 250 K. They constrain grey dimming of light, not the harvesting of low-mass stars, which removes only 20 to 36% of the light. They are systematics-limited: 34 disc and 23 early-type outliers are unexplained and were not followed up.
• Claim 6: reword
  • Reason: Numbers are close to the revised study. The conclusion is published (Wright et al. 2018; Sheikh et al. 2025) and the nearest population bounds are Lacki 2024 and 2025; the new part is the de-duplicated coverage function.
  • Corrected wording: Radio SETI constrains behaviour, not abundance (an original synthesis of a known conclusion). One de-duplicated coverage function built from 29 survey components limits always-on isotropic 1e17 W narrowband beacons anywhere in 1 to 10 GHz to fewer than about 4e5 in the Galaxy (1.7e5 to 2.6e6) and Kardashev-II beacons to about 100 (55 to 18,000; 2,000 from post-2015 searches alone). Earth-2025-level leakage and planetary radars are unconstrained: the limits exceed the number of stars, as Wright et al. (2018) and Sheikh et al. (2025) imply. Closest prior work: Lacki (2024, 2025), Grimaldi & Marcy (2018), Mason et al. (2025).
• Claim 7: reword
  • Reason: Three factual slips against the study (the 42-71% case, 'one or two', 'four or more') and the main points are in Hanson et al. 2021 and Kipping 2025b. 'Dissolves the puzzle' overstates: the host-type puzzle remains.
  • Corrected wording: 'We are early' rests on long-lived planets, as Hanson et al. (2021) and Kipping (2025b) already showed; what is new is how little an atmosphere census can change it. With habitable lifetimes capped at 10 Gyr for all stars and the measured star-formation history, humanity is not early (rank 42 to 71% for 1 to 10 hard steps). Of 14 rocky M-dwarf planets with a decisive verdict, none in the habitable zone, none keeps a thick atmosphere; that removes the earliness only if there is a single hard step (M-dwarf planets must be down-weighted below 0.05 for one step and below 5e-5 for two). Without M dwarfs, K dwarfs of 0.6 to 0.65 Msun that stay habitable for 26 to 45 Gyr keep us in the first 1% for five or more hard steps; with hosts above 0.8 Msun we are never in the first 1%. A lifetime cap removes the earliness but leaves our G-type host unexplained by a factor of about 8. In the model, learning that planets stay habitable beyond 15 Gyr moves the SSA column's P(none other ever in the Galaxy) from 0.33 to 0.14.
• Claim 8: reword
  • Reason: 'Ceiling' and '4.6 to 942' were withdrawn by the revised study; 13 is a lower limit in Kipping's own model. The units finding must be hedged and the authors told. 'About one bit' holds only under conditioning on our existence.
  • Corrected wording: Earth's timeline carries likelihood ratios of order 10, about one bit under the standard conditioning (original reanalysis). Kipping's (2025) 13:1 for fast abiogenesis reproduces and is stable across the Moody et al. LUCA datings (13.3 for the main runs, 12.8 pooling all 38 repository variants). It is model-bound: if 2, 3 or 4 hard steps follow abiogenesis the value is 9, 7, 6 (a corollary of hard-steps order statistics), if the later step is not hard it is 20 to 26, and with accepted fossils only it is 1.5 to 6. Snyder-Beattie et al.'s (2021) Tables 2-3 reproduce only if the printed thresholds 10^-k are read as e^-k, an apparent labelling slip that would lower their three-step weight below 10^-12 from over 90% to 44 to 58%; this is inferred from the tables, the authors have not been contacted, and their qualitative conclusion stands. Pacing versus gated hard steps: Bayes factor 0.6 to 3.3. Under self-indication the same data make a rate-based biological filter unlikely.
• Claim 9: reword
  • Reason: The single p-value is selective, 'stem lineages' names the wrong grouping, the study replaced 'do not support' with 'cannot test', and the reanalysis also weakens the information-loss argument that Mills et al. 2025 use against hard steps.
  • Corrected wording: Convergence data cannot establish many hard steps and cannot rule them out (original reanalysis). Vermeij's (2006) finding that one-off innovations are older than repeated ones reproduces on his tables (odds ratio 3.2 per tenfold age, 1.2 to 9.8) but falls to 1.4 (0.5 to 5.0) when the co-inherited traits of LUCA, LECA and the animal body plan are each counted once (Fisher p = 0.125 at his 543 Ma cut and 0.040 at 1 Ga; the effect survives two milder groupings). This cuts both ways: there are fewer independent one-off events, and there is less statistical support for the view that singularity is an artefact of age. Three of 13 proposed hard steps have recurred; for the oldest one-off events, recurrence limits cannot test hard-step models. Near-human relative brain size arose in 3 to 4 mammal lineages (the best-sampled dolphins reach 0.73 of the human value; no other species has a measured cortical neuron count within a factor of 2); the step from brains to technology has an upper limit only.
• Claim 10: keep
  • Reason: All numbers match integrate3_results.json. Add 'in the model' and a credit for the direction of the updates (Balbi & Grimaldi 2020; Angerhausen et al. 2025), and note that 'more radio SETI: no change' depends on the detectability prior (objection 9).
  • Corrected wording: What would change the answer, in the model. A second origin of life or one biosignature among 25 HWO planets: P(none other ever in the Galaxy) 0.53 -> about 0.3 and 'empty universe' 0.20 -> 0.03. A thorough Solar System artefact survey finding nothing: 'unnoticed' 0.19 -> 0.10. More radio SETI: no change unless loud beacons are given prior weight. M-dwarf verdicts: no change except through the earliness argument.
• Claim 11: reword
  • Reason: 'Largest lever' is true only for SIA. Site versus SSA moves the answer by 0.2; priors move it from 0.26 to 0.74 in the scan and 0.08 to 0.95 across camps. The SIA result is published (Olson & Ord 2021; Finnveden 2019; Cook 2022).
  • Corrected wording: The anthropic convention is decisive only at the SIA end, and it cannot be argued away. Between treating Earth as a given site and SSA the answer moves by about 0.2 (0.53 against 0.33), less than the spread across defensible priors (0.26 to 0.74). Under SIA, emptiness is ruled out and the answer becomes 'they exist; expansion fails or is not evident', with the split set by priors, as Olson & Ord (2021), Finnveden (2019) and Cook (2022) found.

Code errors

• models/model3.py / line 72, BASE['warm_curve'] and its comment '(Lacki 2016; Griffith 2015)'; used at line 189
  • Problem: The curve is Lacki 2016's partial-shroud limit for screens just above CMB temperature, applied to samples drawn as 'waste heat emerges at 100-600 K'. The published mid-infrared limits are 10 times weaker at full coverage and absent below about 25%.
  • Fix: Split temperature into three regimes (near-CMB, 100-600 K, far-infrared) with their own curves, as in objection 3. A patched copy is at model3b.py (adds p_cmb and cmb_curve).
  • Impact: Site P(none) 0.531 -> 0.523; SSA 0.332 -> 0.325; survey-only pull halves (+0.049 -> +0.023).
• models/model3.py / lines 235-236, share_G = W[0] / np.maximum(D_all, 1e-300)
  • Problem: When use_local is False, samples whose denominator underflows get SSA weights near 1e301, so the SSA column is meaningless (posterior 'excluded' = 1.000, evidence 2.8e301). The 'ssa' entries of the no-local ablation rows in integrate3_results.json are garbage. The base run is protected only because the local term multiplies those samples by zero.
  • Fix: Compute the SSA share in log space with a physical floor on D_all (at least one civilisation, ours), or drop SSA from ablation rows that omit the local term and say so in the table.
  • Impact: No effect on the draft's quoted numbers (claim 3 quotes site only), but the stored SSA ablation values must not be published.
• report/FINDINGS_DRAFT.md / section B table, prior column, row 'Settled in a way we would see'; claim 1; mean-likelihood line
  • Problem: Prior is printed as 0.27; integrate3_results.json and a rerun give 0.309, and the printed column sums to 0.96. Mean likelihood for 'en route' is printed 0.44; the file gives 0.41.
  • Fix: Use 0.31 and 0.41; change '27% of prior mass' to 31%.
  • Impact: Presentation only.
• models/model3.py / line 73 comment '# any temperature [STUDY tf_kiii + fp_kiii]'; line 38 comment 'retention mean 0.34'; line 78 '[STUDY seti_meta]'; line 54 '(Tyrrell 2020)'
  • Problem: Four source tags do not match their sources: the Tully-Fisher part holds only below about 250 K; 0.34 is 'a bracket, not an estimate'; the c_det range excludes the study's Kardashev-II class (0.026 to 0.5); the f_persist floor of 0.03 excludes Tyrrell's 1.45% per-run success.
  • Fix: Correct the comments, and either widen c_det and f_persist or state the exclusion.
  • Impact: f_persist floor 0.01: site P(none) +0.010. c_det up to 0.03 or 0.5: +0.012 or +0.021.

Missing explanations

• Patchy or steady-state settlement: Carroll-Nellenback et al. 2019 (and Landis 1998, from memory) describe a Galaxy with sustained settlement and finite settlement lifetimes in which Earth sits in an unvisited pocket. The model has no such class: 'stuck' needs expansion to fail almost always and 'unnoticed' needs low footprint. The flyby study's own stall corners (few settleable systems, short hops, small velocity budgets) point the same way.
• Synchronised timing: Annis 1999 (JBIS 52, 19) and Cirkovic & Vukotic 2008 propose galaxy-wide resets so that nobody is much older than us. Round 1 raised it; model3 still has no correlated-timing variable, and the head-start census is used as if emergence times were independent.
• Geophysical limits on planetary habitability are treated as a coin flip (p_cap_short = 0.5) with no literature behind it, although this is the parameter claim 7 calls decisive.
• The far-infrared gap itself: natural galaxies reprocess about a third of their starlight in the far-infrared, so partial harvesting with 20-100 K waste heat is indistinguishable from dust at any covering fraction below about 0.5. The report should say this is the regime no existing survey constrains, since it is where the 'unnoticed' class can also hide harvesting.

Critic’s own estimate of each answer

• Empty universe: 0.15
• Rare: 0.24
• En route: 0.04
• Stuck: 0.28
• Unnoticed: 0.23
• Hidden: 0.02
• Other: 0.04
• What “other” covers: Explanations outside the model's partition: patchy or steady-state settlement that has not reached the Sun's neighbourhood, synchronised late emergence, or a structural error in the chain. The rest is an equal-weight average of the site and SSA columns after the literature-correct survey curve, which moved nothing by more than 0.01; I give SIA no separate weight.

Presentation advice

Use three fixed novelty labels and apply one to every claim: 'reproduction with new data', 'new in degree', 'new'. By my searches only three items earn 'new': the fundamental-plane search at scale, the Snyder-Beattie table inconsistency, and the clustered Vermeij reanalysis. Put a prior-work table in the report with four columns: claim, closest published result and its number, this project's number, change in P(no other ever in the Galaxy) when the input is removed. The last column is the honest answer to 'what did the original research add': 0.001 for the optical galaxy limits, 0.001 for the head-start census, 0.00 to 0.02 for radio. Mark every unrefereed source as a preprint at the point of use (Lacki 2016, Kipping 2025b and 2026, Curtis et al. 2026, Olson & Ord 2021, Sandberg, Drexler & Ord 2018). State the temperature regime beside every galaxy limit: near-CMB (Lacki), 100-600 K (Griffith, Curtis), far-infrared (this project). Never write 'first' without naming Annis 1999 and Zackrisson et al. 2015 in the same sentence. For the Snyder-Beattie finding, write 'apparent', say the authors were not contacted, and contact them before release. Generate section C from the study result files so stale numbers cannot recur.

Were the round 1 fixes made?

Round 1 had no literature lens, so I checked the round-1 objections that touch citations and survey limits. Fixed: the single 'N = 2,000 at any temperature' cold limit was replaced by a covering-fraction curve built from the fundamental-plane and Tully-Fisher effective samples, zero above 0.93 (line 73); expanders from other galaxies were added, and the kernel reproduces Kipping 2026's published thresholds to within 15% when I feed it his assumptions. Fixed in the wrong place: the study reports absorbed their referees' novelty corrections (Zuckerman 2026, Fernandez-Puig 2026, Hanson 2021, Lineweaver, Legassick, Lacki 2019, Bains & Schulze-Makuch), but FINDINGS_DRAFT.md did not; 'any temperature' was withdrawn in tf_kiii yet survives in claim 5 and in the model3.py comment. Mis-implemented: round 1 asked for 'the Lacki/Griffith curve (warm)', and v3 did exactly that, which bakes in a temperature-regime error that started in gaps_and_corrections.md section 7; Lacki's curve belongs to near-CMB screens. Not fixed, but immaterial: round 1 noted the Milky Way analogue is a disc while the fundamental-plane sample is early types; the cold curve still sums both, and using Tully-Fisher alone changes site P(none) from 0.532 to 0.531. Not fixed: the synchronised-timing hypothesis (Annis; Cirkovic & Vukotic) raised in round 1 has no variable in model3.