# WP35 — Limitations (written after Phases 4–5, before the paper)

Each item names the deliverable it qualifies. "EMPIRICAL" refers to Seshat/EA/V-Dem estimates; "SIMULATED" to the agent-based results; neither is causal.

# A. Data (EMPIRICAL)

1. Crisis-selected outcome. CrisisDB covers 150 of 626 polities; H2-A tests severity conditional on a recorded crisis.
2. Undated facts. 93.7% of Polaris social-complexity facts have no date; "pre-crisis" capacity is mostly whole-polity capacity. A dated-only test was not estimable.
3. Presence codes. Seshat K items record whether an institution existed, not its reach, quality or cost.
4. Scale-only D. No usable indicators of division of labour or dispersion in Polaris; occupational complexity is coded for 75 polities only and was excluded.
5. Collinearity. corr(D, K) = 0.84; P keeps 17% of the joint variance and is a residual; measurement error in either index leaks into P.
6. Coder circularity. Polaris codes and CrisisDB codes are produced by the same project from overlapping historiography; a polity known to have collapsed may be coded as less administratively capable.
7. Polity boundaries. Seshat polities are periodization decisions; "polity end" (Y2) is not an instability event.
8. Modern K and democracy. The six V-Dem capacity items correlate 0.75 with electoral democracy; the capacity interpretation competes with regime type. Expert coders rate periods knowing later events.
9. Modern D. Maddison population starts in 1820 for most countries; land area only from 1886; no dispersion measure.
10. Ethnographic Atlas. Cross-section with focal years −2000 to 1965; no outcome; EA031 is a coarse ordinal measure; midpoints of open categories are bounds.
11. Crosswalks. 39 V-Dem countries (mostly historical or micro-states) lack Maddison population and drop out.

# B. Design and inference (EMPIRICAL)

12. No identification. Associations only (research/identification.md). The modern negative control shows P associated with past events as strongly as with future events, so P marks chronically unstable countries rather than forecasting onsets; the Oster-adjusted coefficient crosses zero and the Cinelli–Hazlett robustness value is 0.07.
13. Power. Seshat binary Y1b is underpowered (MDE ≈ 0.6 SD); only the fractional severity outcome has adequate power.
14. Few clusters. 42 regions / 10 macro-regions; cluster-robust inference with few clusters is anti-conservative; alternative clusterings are reported.
15. Baselines decide the verdict. By the pre-registered rule, P must beat every baseline out of sample in ≥ 80% of folds: it fails against the D-only baseline in Seshat (leave-one-region-out) and against the K-only baseline in the modern panel (both schemes). The mismatch index therefore adds no predictive content beyond its own components.
16. Placebo. A Seshat placebo index (scale minus production infrastructure) performs about as well as P; the modern placebo (scale minus polyarchy) is weaker than P but still significant.
17. Nonlinearity. The modern hinge/spline improvement is 0.001 nats per observation and meets the 80% rule by one fold; it is reported as marginal, not as a threshold.
18. Equinox-2020 replication was pre-registered but not executed (time); it is listed as a blocked task in the final audit.

# C. Theory (PROPOSITION / CONJECTURE)

19. Theorems 1–5 concern an idealized communication system; nothing in them is tested on historical data. C1–C3 are conjectures; C1 (proxy monotonicity) is undermined by the collinearity and placebo results above.
20. Capacity is exogenous in the model; historically it is built in response to demand.

# D. Simulations (SIMULATED)

21. Institutions are imposed, not emergent; results describe the cost structure of imposed regimes.
22. Message and bit accounting is stylized (one-bit acknowledgements, fixed admin overhead per task); absolute numbers are not comparable to any historical quantity. Only within-simulation comparisons are meaningful.
23. Decoding success does not depend on the plan drawn, so distribution shocks affect cost but not success by construction; this is a modelling choice, not a finding.
24. Hierarchy saturation depends on the administrative-capacity ratio, a free parameter swept from 0.25 to 4; the "phase" boundary near ratio 1 follows from Theorem 5 and is confirmed, not discovered.
25. Shannon codes (not Huffman) are used so that the measured stale-code penalty is comparable to Theorem 3; with weakly concentrated plan distributions a Shannon code can be longer than raw coding (one cell), and a learning cost of 200 bits per agent is repaid only with a large, concentrated plan space and a long horizon.

# E. Process

26. Executed autonomously by a language-model agent in one session; the PI has not yet reviewed the code or the text.
27. Reference verification relied on Crossref metadata; chapter contents of Johnson (1982) and the published tables of Sng (2014) were not read.