# Response to the three adversarial reviews — WP35 (2026-11-01)

The three reviews (`review_institutional_economist.md`, `review_historical_anthropologist.md`,
`review_computational_social_scientist.md`) were produced by independent reviewer personas with access to the full
repository. All three recommended **major revision**. None endorsed the paper, and none was asked to. This document lists
each substantive point, what was done, and what was not done and why. Every change is logged as an amendment (A-002 to
A-007) in `research/decisions.md`; analyses were rerun; the manuscript was recompiled (80 pages).

## A. Defects acknowledged and fixed (reruns performed)

| # | Reviewer | Point | Action |
|---|---|---|---|
| 1 | Anthropologist, Economist | Five robustness variants announced in the text (range bounds, inferred codes dropped, within-era/region standardization, ranks) had **never been computed** at the crisis level: the code skipped missing columns silently; the text cited a table that did not contain them. | Variants built at the crisis level (`build_indices.py`), estimated, tabulated (Table 10): all positive, 0.13–0.27. Rule-1 violation logged (A-002). |
| 2 | Economist | The Seshat **placebo was a tautology**: `P_placebo = z(D) − z(X_prod_infra)` fitted with `X_prod_infra` as a control is an exact reparametrization of the scale-only baseline (identical coefficient 0.2572…, SE, log-likelihood). | Placebo now fitted without the production-infrastructure control, with `P` fitted on identical controls for comparison: placebo 0.19 (SE 0.17) vs P 0.29 (SE 0.08). Abstract and text changed from "performs comparably" to "not distinguishable from P" (A-006). |
| 3 | Economist, CSS | **Silently failed CV folds**: the first forward-chaining split (train ≤ 1945) failed for every model (a region absent in training); 14 Seshat model-folds failed similarly; the text reported "three splits" and "fifty folds". | Test rows with fixed-effect levels unseen in training are now mapped to the training mode (`common.map_unseen_levels`); all 3 forward splits and all 50 Seshat folds run; counts corrected (A-004). |
| 4 | Economist | Seshat k-fold not blocked by polity (16 polities with several crisis records share identical predictors). | Folds blocked by polity in all Seshat cross-validation (A-004). |
| 5 | Economist | `epidemic_v` used as a covariate is itself a CrisisDB consequence item. | Dropped from the control set; pre-registration deviation logged (A-006). |
| 6 | Economist | Baselines fitted on different samples (n = 87–127) while the text said "same observations". | All covariate specifications and baselines now fitted on a common sample (n = 93); A1 (n = 127) kept and labelled. |
| 7 | Economist | Row "A3: Y1 ~ P + D + K" was mislabelled (it conditions on the complexity component); modern "C3: P given K" is algebraically B5. | Relabelled; C3 described as the reparametrization of B5 that it is. |
| 8 | Economist | P imposes β_D = −β_K; the unrestricted model is the right object. | Wald test of β_D + β_K = 0 added in both panels: not rejected in Seshat (p = 0.22, because β_K is imprecise), rejected in the modern panel (p < 0.001). Reported in results and abstract. |
| 9 | Economist, CSS | **Negative control and H4 misread** against the written rules: the negative control passes by the letter (0.44 < 0.48); the modern H4 rule was met. | Text now states the literal verdicts (negative control passes by < 1 SE; H4 "not rejected" in the modern panel, spline 12/15, +0.003 nats) with the interpretive caveats (A-007). |
| 10 | Economist | Holm correction asserted, never computed. | Computed (`holm_primary_tests.csv`): all three adjusted p < 0.001. Reviewer's point that H1-Seshat is an "unfailable" primary is accepted and noted in the limitations. |
| 11 | Economist | Measurement overlap: settlement hierarchy (D) and administrative levels (K) both measure hierarchy depth. | Overlap checks added (D without settlement hierarchy, K without administrative levels, both): β_P = 0.25 in all three. Presence-anywhere size bias of K acknowledged in limitations; not fixable with Polaris. |
| 12 | CSS | `t_since` imputed with the panel maximum for countries with no observed prior event. | Replaced by years since panel entry plus a left-censoring indicator in every modern specification (A-004). |
| 13 | CSS | Event count "2,692 instability events" are overlapping windows. | Abstract and text now give 2,692 event windows, 901 distinct event-years in the estimation sample, 1,676 in the panel. |
| 14 | CSS | Simulation accounting: bits ignored relay hops; only the LCA administrator was charged; the capacity ratio was defined against a load that ignored withdrawn producers; federated module administrators each received the system-wide bandwidth; distribution shocks were a no-op in uniform cells; network-shock robustness of flat regimes was a floor effect. | Codeword bits charged on every hop; every administrator on the relay path charged; **equal total budget** for central and federated designs (ratio × expected relayed load of the central design, allocated to administrators in proportion to a warm-up measurement of load); shock statistics restricted to non-uniform cells; shock effects also reported by bandwidth. All 6,632 runs rerun; simulation section rewritten (A-005). Qualitative conclusions changed in one respect: central hierarchies are now expensive in bits (83 per success vs 25 for records) and federation is the only hierarchical design that survives a tight budget. |
| 15 | CSS | Theorem 1 check wrongly described: an optimal prefix code is never longer than the fixed-length code; only the (suboptimal) Shannon code used in the simulation can be. | Remark added after Theorem 1 in the paper; simulation text corrected. |
| 16 | CSS | Simulation "confirms" success effects that are parameters (eps = 0.25 vs 0.02 vs 0). | Section 8 rewritten: success rankings are explicitly described as parameter-driven (with the eps ablation showing the reversal at eps = 0.05); the measured content is the cost side. Abstract changed from "confirm them in silico" to "instantiate them in silico on the cost side". |
| 17 | Anthropologist | Severity share is not a scale (14 weakly correlated items); Hoyer et al. 2024 cited but not applied. | Item-level analysis added (Table 12): mean inter-item correlation 0.08, first eigenvalue 19%; P and D load on elite-political items (decline, assassination, deposition, successful revolution, civil war) and negatively on century-long conflict; collapse is present in 2 of 126 records. Text added. |
| 18 | Anthropologist | K saturates in the crisis sample, so P ≈ rescaled D. | Reported: crisis-sample K mean 0.71 (0.58 overall), 40% of records with K > 0.8, corr(P, D) = 0.63 and corr(P, K) = 0.02 within the sample (Table 5, last row). |
| 19 | Anthropologist | Cox result (higher K → shorter polities) dismissed in two sentences. | Periodization control (polities per NGA) added: it raises the hazard (HR 1.37) but K stays at HR 1.71 (p = 0.009). Text rewritten to report the result as running against the hypothesis, with candidate explanations left open; added to the abstract. |
| 20 | Anthropologist | Macro-region scheme pools culturally distinct areas (Sahel with North Africa; Angkor with Polynesia). | Cultural-sphere hold-out scheme added (`configs/regions.yaml`); H5 rerun: pooled coefficients 0.24–0.37 in every held-out sphere. The original geographic scheme is kept as pre-registered. |
| 21 | Anthropologist | "Pre-crisis predictors" overstated (3.6% of facts dated). | Abstract and data section reworded; dated share reported; label kept only for consistency with the pre-registration. |
| 22 | Anthropologist | Pre-registration says "13 items" but lists 14. | Typo acknowledged (A-003); the list governs. |
| 23 | Anthropologist, CSS | Equinox-2020 replication called "blocked" in the paper vs "not executed" in the audit; Appendix A claimed "no post-hoc amendments". | Both corrected: "not executed"; Appendix A now references amendments A-001 to A-007. |
| 24 | CSS | Diagnostic file `seshat_unparsed_values.csv` was empty. | Cause: a unit test called the parser on a one-row frame and overwrote the diagnostic; parser now takes `write_diag=False` in tests; file regenerated (132 rows). |
| 25 | CSS | pymc/arviz pinned but unused; `replication/` empty. | Removed from `pyproject.toml`; `replication/README.md` written. |

## B. Points accepted but not fixable with the available data (stated in the paper)
- Historical comparability of presence codes across 13,000 years; "presence anywhere" codes mechanically increasing in
  territory; coder circularity between Polaris institutional codes and CrisisDB codes (Anthropologist, Economist).
- EA033 is itself a scalar-stress response, so H1-E replicates Carneiro (1967) rather than testing compression; EA031's
  open top category (Anthropologist). Stated in limitations.
- The modern K items correlate 0.75 with electoral democracy and are coded ex post (Economist). Stated; the democracy
  control and placebo remain the only available mitigations.
- Oster's δ = 1 is implausible when a control is a near-component of the regressor; RV lacks a benchmark (Economist).
  The sensitivity paragraph now says so and no longer uses the word "reverses" in the abstract.
- The D+K baseline nests the P model, so P cannot beat it in expectation (Economist). Stated in limitations; the Wald test
  of the restriction is now the primary object of that comparison.
- The hierarchy saturation is confirmed, not discovered (CSS). Stated.

## C. Points not accepted, with reasons
- "The 80% rule could be gamed or is too harsh" (Economist): the rule was fixed before any result; we report fold-level
  results so that any other rule can be applied, and we note that the verdict is the same under a 50% rule against the
  scale-only baseline in Seshat (42%).
- "Transductive PCA/standardization before CV is leakage" (CSS): it is outcome-blind and absorbed by the GLM slope and
  intercept; the fold-internal check (`h2_seshat_cv_noleakage.csv`, A-001) shows the verdict unchanged. Both reviewers
  who raised it agreed it is immaterial for the logit; it would matter for variants that re-fit PC1 on the test fold,
  which none do.
- "Replace the geographic macro-regions by cultural spheres" (Anthropologist): added as an additional scheme, not as a
  replacement, because the geographic scheme was pre-registered.

## D. What changed in the verdict
Nothing in the direction of the main conclusion: H1 holds as association; H2 fails the pre-registered out-of-sample rule
in both panels; H4 is rejected in Seshat and now reported as "not rejected, small" in the modern panel; H5 holds. Two
things became stronger against the hypothesis after the corrections: the restriction β_D + β_K = 0 is rejected in the
modern panel, and the Cox result (higher capacity, shorter polities) survives a periodization control. One thing became
weaker against it: the Seshat placebo, correctly specified, is weaker than P (0.19 vs 0.29) though not distinguishable
from it. The simulation conclusions changed in substance for the hierarchies (central designs are expensive and fragile at
equal budgets; federation is the surviving hierarchical design) and in framing for the success effects (parameters, not
findings).