The record

Run by run

Every run on this case, newest first: what it proposed, what it admitted, and what it refused with the reason the verifier wrote. A candidate that is not on the ledger is here, with why. The dispositions file holds every row.

Change history

Every change to this case, newest first: what changed, why, and who made it — including the AI's role. The log is append-only; a correction is a new entry, never an edit to an old one.

1. 2026-09-09scripts/migrations/2026-09-09-hash-locators.ts (mechanical re-stamp) · AI-assisted

   Ledger hash redefined to exclude source locators (url, archivedUrl); 1 edition/assessment basis stamp(s) recomputed mechanically, content unchanged: edition-2026-09-08-migration.yaml.

   Why: Founder decision 2026-09-09: a locator says where a record is read, not what it says; a by-hand URL addition had set a fresh panel aside and made an edition due that could only re-adopt.

2. 2026-09-08Cursor Grok 4.6 (Cursor cloud agent), directed by the founder; runId title-punctuation-2026-09-08; promptVersion none (founder-directed session, not a pipeline prompt) · AI-assisted

   Case title punctuation restored after the pre-overhaul publication restore: "Elusive by Law?" is again "Elusive by Law". The trailing question mark is dropped so the published name is a statement. Subtitle and framing questions are unchanged.

   Why: Founder direction (first given 2026-09-06, reaffirmed 2026-09-08 after the restore): case names should not carry question marks. The interrogative belongs in the subtitle and the research questions, not on the card and in the header.

3. 2026-09-08scripts/migrations/2026-09-08-editions.ts (mechanical), founder-directed session · AI-assisted

   Migrated to editions (docs/AUTOMATION.md, build step 2). The article, the 10-claim featured set, and the displayed judgment now live in edition edition-2026-09-08-migration, which adopts the migration assessment 2026-09-08-migration-mpi. Claim records carry propositions and anchors only; evaluative fields moved into the adopted assessment's per-claim treatment; the dossier header, research priority, and component verdicts moved from case.yaml into the assessment; claims-catalog.yaml merged into claims.yaml (tiers ended). No judgment changed. Where a claim record's hand-written credibility differed from the draft overlay's verdict, the record's value (what claim pages displayed) was carried: MPI-C001: record provisionally_supported, draft mixed.

   Why: One reader-facing unit that binds verdict, selection, and article, so they cannot drift apart; one home for judgment; featuring becomes a recorded decision.

4. 2026-08-25anthropic/claude-opus-5 (scheduled maintenance run, aletheia-editorial-audit-v1); publishes only through the constitutional gate's multi-model concurrence, per AGENTS.md §3.15 · AI-assisted

   Corrected the steelman section's description of the 2020 Walach two-lab replication, which the article had called "preregistered and adequately powered." The evidence record for that study reports rebuilt hardware, a 45×45 matrix and evaluation against 10,000 simulated experiments, but no preregistration or power analysis; the preregistered designs in this corpus are the OSF consensus replications and Grote's blind analysis, both of which returned nulls. The sentence now states what the record actually reports.

   Why: Attributing preregistration to the strongest positive result inverts the case's central asymmetry between strict designs and positive findings, and would mislead a reader about which studies were method-hardened.

5. 2026-08-23Fable (AI agent), directed by Eugene · AI-assisted

   Fifth independent judge added to the cross-model check: Kimi K3 (Moonshot's open-weights model, served via Venice) re-assessed the case blind under the same protocol as the four frontier judges. Its case verdict and per-claim verdicts appear on the concurrence panel and claim pages alongside the others; concurrence statistics recomputed over five judges.

   Why: An open-weights judge from a different training lineage strengthens the independence of the check — agreement across five vendors, including one outside the Western frontier labs, is harder to attribute to shared disposition.

6. 2026-08-23Fable (AI agent), directed by Eugene · AI-assisted

   Cross-model check run (scripts/cross-model-check.mjs, promptVersion aletheia-check-v1): four independent judges — Opus 5 (Anthropic), GPT-5.1 (OpenAI), Gemini 3.1 Pro (Google), Grok 4.6 (xAI) — each re-assessed the case blind to all prior assessments. Result: case verdicts diverge (weakly_supported x2, mixed, unresolved) around the displayed weakly_supported; claim agreement 7/10 exact, 1 within one step, 2 split (MPI-C001, the founding excess-correlation results, and MPI-C005, the NT-axiom decline prediction). Check runs never display as the case narrative; the concurrence panel summarizes them and claim pages carry each model's per-claim verdicts.

   Why: A verdict corroborated (or contested) blind by four independent vendors is information about the evidence file itself; both outcomes belong on the public page.

7. 2026-08-23Eugene (direction); applied by Fable (AI agent) · AI-assisted

   Site v1.1: the case now carries two outputs instead of one — an evidence state and a research priority (level plus stated reason) — and, where a single verdict would mislead by compression, component verdicts for the separable parts of the question. Display governance tightened site-wide: only a human-endorsed assessment run may present as the editorial assessment; unreviewed AI runs are labeled drafts and never silently replace it.

   Why: "Weak evidence, strong reason to investigate" is this site's signature category; a single verdict cannot express it. Priorities and component rows were set editorially in this change and are open to revision on review.

8. 2026-08-22Grok (AI agent), directed by Eugene · AI-assisted

   Added one context-only plate to the overview: SBM-20 Geiger–Müller tubes (CC BY-SA 4.0, ArticCynda / Wikimedia Commons), illustrating the hardware entropy source used in true-random event generators. Caption states the photograph is not of any MPI experiment.

   Why: The article described physical random generators without showing what they look like; a licensed object photograph is more useful than a portrait or a lab snapshot that could be mistaken for a result.

9. 2026-08-22Eugene (review decisions); applied by Fable (AI agent) · AI-assisted

   First human review (PR #14): the four open questions flagged at import were decided, each endorsing the AI default without changes. The case verdict stays weakly_supported (the preregistered-positive residue keeps it out of the contradicted family despite the blind-analysis null); the falsifiability claim MPI-C010 stays mixed (the model's in-principle exposure judged a genuine offset to its practical insulation); the Praeger chapter-title discrepancy on SRC-LUCADOU-2015 stays flagged unresolved pending a library check; the escaped-butterfly cover art stays. No claim, evidence, source, or assessment record was modified; the assessment overlay remains labeled an unreviewed AI draft, as the review covered the flagged judgment calls rather than a line-by-line re-derivation of the run.

   Why: The site's design requires human review decisions to be visibly recorded even when they change nothing — an endorsed default is a human judgment, not an AI one, from this date forward.

10. 2026-08-22Fable (AI agent), directed by Eugene · AI-assisted

    Literature watch configured (watch.yaml): six discovery-only queries covering the matrix replication record, the GQT formalism, output from the model's theorists and active experimenters, the true-vs-pseudo-RNG contrast, decline-effect explanations, and the QRP/publication-bias critical literature — the last watched with the same machinery as the proponent side per the fairness rules.

    Why: The case's verdict is hostage to a small set of live research threads (adversarial replications, dependence-aware reanalyses, the Flores preprint's review status); the weekly watch surfaces movement on them without letting anything into sources.yaml unreviewed.

11. 2026-08-22Fable (AI agent), directed by Eugene · AI-assisted

    Case published: 10 featured claims across the three rungs plus a falsifiability crux, 3 catalog claims, 17 evidence records, 17 sources (all ai_verified with per-source verification notes; none unverified, none fabricated), 5 research opportunities, and a first AI assessment overlay labeled as an unreviewed draft. All claims labeled ai_extracted pending human review. Structural note recorded: because MPI predicts replication failure, supporting and undermining evidence were assigned against specific claims rather than against "the hypothesis" — a null replication supports the observation claim MPI-C002 while weighing against the attribution claim MPI-C007, and the tension between those two facts is itself carried as claim MPI-C010.

    Why: Launch the case with honest provenance labels and the falsifiability tension modeled explicitly rather than argued away.

12. 2026-08-22Fable (AI agent), directed by Eugene · AI-assisted

    Two mischaracterizations in the seeding synthesis document identified and corrected on import. (1) It presented Grote 2017 as a successful CMM replication citing p ≈ 0.012; the published abstract shows all three preplanned analyses null (p = 0.438, 0.703, 0.0949; combined p = 0.315) with p = 0.012 arising only from a post-hoc variant — the corrected reading appears in MPI-E006. (2) It described the Walach-group replication line as consistently successful; the group's own 2021 paper reports two preregistered consensus-protocol failures and describes the 2020 effect as "significant but fragile," with significance breaking down under improved analyses — carried in MPI-E002/E003. A citation discrepancy for the founding MPI paper (1995 EJP 11:58–75 vs a circulating "1994 EJP 5" form) is recorded unresolved on SRC-LUCADOU-1995.

    Why: The epistemic rules require observation before interpretation and forbid imported characterizations from carrying assessments when the primary record says otherwise.

13. 2026-08-22Fable (AI agent), directed by Eugene · AI-assisted

    Case seeded from the "Entangled Minds and Pragmatic Information" synthesis document (runId mpi-case-2026-08-22): the document was read in full and mined for claims, evidence, and sources; every citation was then independently re-verified against Crossref, publisher pages, or retrieved full text before import. Primary texts read in full during verification: von Lucadou, Römer & Walach 2007 (JCS), Walach et al. 2021 ("Nailing Jelly"), Watt & Tierney 2013 (author proof), Kennedy 2003, Flores & Rapazote-Flores 2023, and the Bierman 2016 abstract and conclusions.

    Why: Real citations only: a synthesis document may not carry assessments second-hand.