Appendix B·v3.0
SHING CIVITAS Research Series: Source and Evidence Audit
What was recomputed, which interpretations were withdrawn, and which records are missing. Corrections listed here travel with every citation.
Version 3.0, 2026-11-12. Scope: research claims only. No website, CMS or deployment content was reviewed or changed.
1. What was and was not examined
Examined. The editorial studio v1 (first drafts and charts), and four data packages: P10–P12, P14, P17–P18, P19. The SHA-256 hashes of the uploaded files match the manifest. Frozen protocols and reports, aggregate CSV/JSON files and per-world result rows were read. All primary results quoted in the articles were recomputed from per-world rows with independent scripts (python3 -I), including sign tests (one- or two-sided as frozen), seed-paired bootstrap intervals, and Holm adjustment.
Not examined.
- The official website (https://shingcivitas.com/) could not be fetched during preparation (the request failed and was not retried). Nothing in the articles depends on the website; its claims have not been checked against the articles.
- The simulation engine, dependency lock files, per-tick event streams, raw prompts/responses, model version snapshots and git history, which are not distributed.
- Experiments P1–P9, P13, P15 and the full P16 package (cited only as background through the P10–P18 protocols and reports).
- Any independent third-party replication (none exists).
2. Verified findings (recomputed from per-world rows, matching frozen outputs)
| Study | Verified quantities |
|---|---|
| P10 | H1 +1.425 [1.050, 1.800], 31/7/2, Holm 5.23×10⁻⁷; H2 +6.125 [2.900, 10.050]; H3 +5.875 [−0.525, 14.200], Holm 0.608; H4 −2.425 [−8.250, 2.425], Holm 0.272; old_noshare − old_base starvation 0.000 [−0.175, 0.175], care +41.925 |
| P11 | H1 +0.1544 [0.1242, 0.1879]; H2 +0.0441 [0.0283, 0.0615]; H3 +15.325 [8.925, 22.225]; H4 −0.475 [−0.700, −0.275]; Holm p for H3/H4 5.49×10⁻⁴ |
| P12 | H1 +0.2195 [0.1730, 0.2685]; H2 +1.125 [0.750, 1.525]; H3 +56.400 [40.000, 74.950]; H4 −0.600 [−0.875, −0.350], Holm 6.1×10⁻⁵; legal_both − inherited_both starvation −0.075 [−0.400, 0.250], transfers +62.050 [42.500, 83.375] |
| P12 forks | 28/38 vs 17/38 alive at 48 ticks; +0.2895 [0.1316, 0.4474]; alive-ticks +8.184 [4.026, 12.816]; fatigue +12.593 [12.082, 12.967]; mean transfer 3.086 (30 positive, 8 zero) |
| P14 | +1: −0.511 [−0.604, −0.415], 37 worlds, −18.95%; +4: −1.292 [−1.509, −1.081], 37 worlds, −25.97%; +12: −1.003 [−1.990, −0.099], 36 worlds, −0.63%, 17/19, Holm 0.868; gate failed |
| P17 | H1 +0.023 to +0.073; tool +144.425 (F2a_cap2), +234.650, +215.950, +387.650 (Holm 0.0415); reserve +465 to +3,135 (4/4); house 0/4 |
| P18 | mind9 K − A +29.475 [−18.375, 77.825], Holm 0.337; mind9 P − A +57.325 [8.975, 104.0], Holm 0.182; old K − A +59.175 [8.600, 106.425], Holm 0.0044; old P − A +9.700, Holm 0.763; mind9 K − KV ≈ +201, KV − A ≈ −172, T − A ≈ +206 (secondary) |
| P19 | W table (Table 3.1 of Article 3); constrained llm_mem − m9 −26.696, − old −38.195, Holm 1.3×10⁻⁸; llm_mem − llm_nomem +1.813 not supported; abundant tests all not supported; probe +0.1583 [0.096, 0.221], Holm 6.33×10⁻⁵; deaths per world (Table 3.3); 8,034 requests (8,029 OK, 5 failed); USD 2.833641; 120/120 replays identical; 591 infeasible steps |
3. Statistical-convention findings
- Sidedness differs by study. P10–P12, P14 and the P19 W-tests use two-sided exact sign tests; the frozen P17 and P18 primary families and the P19 probe use one-sided (greater) tests. An early recomputation using two-sided tests for P17/P18 disagreed with the frozen values; the frozen one-sided values were reproduced exactly. Articles state the sidedness per study.
- Bootstrap size. The care-versus-rest fork intervals for alive-ticks and fatigue differ slightly from a 10,000-resample recomputation because the frozen analysis used 5,000 resamples; frozen values are quoted.
- Holm is rank-based. A placeholder Holm adjustment tried during checking (multiplying by seven) was wrong; the frozen P19 probe Holm p (6.33×10⁻⁵) is correct.
- Pooled counts (P11 stage counts, P12 funnel, P14 labels, P19 focal instances) are descriptive and never sample sizes.
4. Corrections and withdrawn interpretations (must travel with any citation)
| Item | Status |
|---|---|
| P11 funnel/“legally autonomous” care estimates and repeat-care explanations (including the figures 34%/32%, 14–19% reuse and 85%) | Withdrawn after P12/P15; only numeric contrasts retained |
| P12 legal contract | Applies to the care pathway only; executor still clips transfer quantity using hidden inventory; no extra survival vs inherited contract |
| P16-B “zero-yield gathering wears out tools” | Zero actual effect (patch never fired); real cause: one durability unit per gathering action |
| P17/P18 estimands | T − A, K − A, K − KV, KV − A, P − A and R − A are different questions and must not be interchanged |
| P18 mind9 P − A | Interval excludes zero but sign test fails: reported “not supported” |
| P18 Stage D | Not run (gate failed); the learning design document is an unimplemented plan |
| P19 mortality | First summary's “2.5 starving residents per world” and “about 2.3 starvation deaths per world” withdrawn; 74/72 (constrained) vs 85/78 (both worlds) have different denominators |
| P14 | prediction_gate_passed = false; no action phase; temporal-accounting defect (226 of 3,142 receipts in 13 worlds) unrepaired; prior ablation and temporal audit are post hoc |
| F1_pop192 | Confounds density with group size; “constrained” is defined by energy < 45 |
5. Provisional or unverified claims
- P17 mutual-aid and exchange removal (secondary): reported in the frozen outputs as without robust positive value; not independently recomputed here.
- Mechanism claims: P18 Stage C diagnostics (craft never positively valued; 72–90% of tool-less decisions lack 3 wood) and P19 time-allocation explanation are exploratory.
- P19 cost figures are estimates at archived peak prices, not invoices.
- Authorship/tooling statement. The protocols name an AI coding assistant as drafting author. The owner must confirm the exact disclosure wording.
- Freeze dates. Protocols were frozen in a private local repository; commit hashes and dates cannot be checked from the packages.
- File counts. The studio drafts state 37 (P10–P12), 46 (P17–P18), 25 (P19) and 22 (P14) files; the manifest and package contents do not agree with these counts, and the studio's
ARTICLE_DATA_MAP.mdhashes are stale. Current hashes are those in the website data-download manifest.
6. Missing primary records
| Missing | Affects |
|---|---|
| Simulation engine source and dependency lock files | Re-running any simulation |
| Per-tick event streams and git history | Verifying freeze dates, twin-hash checks, replay claims |
| Raw prompts/responses, model snapshot identifiers, provider invoices | P19 re-querying, cost figures, any provider-specific claim |
| Distributed analysis modules imported by scripts | Re-running analysis pipelines |
| P16-A/B complete records beyond the aggregate rows included | P16 background statements |
| Independent replication or external audit | All claims |
| Public preregistration | Freeze claims |
7. Reference verification status
Updated 2026-11-12: All unique external bibliographic works cited by the introduction and four articles have now been checked against publisher/journal/DOI/conference/library records. Corrected two material metadata problems: (1) Hamilton (1964) consists of two separate papers, parts I and II, pages 1–16 and 17–52; (2) NBER Working Paper 31122 names John J. Horton, Apostolos Filippas, and Benjamin S. Manning, not Horton alone (initial issue 2023, revision Feb 2026). Every reference now includes a stable DOI, publisher, proceedings, or library URL. This is bibliographic verification, not a content-level replication or endorsement of SHING CIVITAS findings. The register is in Appendix C · Reference register. DeepSeek model identity is grounded in archived P19 protocol (deepseek-flash) and the provider's October 17, 2026 alias update; no immutable model snapshot was captured.
8. Language and claims audit
- No result is generalized to real humans or policy. Words such as “emerge” are used only in the sense of the introductory article.
- In v3 the P19 article states the DeepSeek model and API identifier explicitly; the owner's decision to publish and the remaining contractual risk are recorded in the publication checklist.
- Marketing language, if present in studio drafts (for example “convincing plan keeps a resident alive”), was not carried over.