[jadenfix]

Summary

• One entry point. uv sync && uv run python -m simtheory verify, or make verify, runs these checks in order:

  1. byte-compiles src/;
  2. validates every claims/*.json manifest: schema, globally unique IDs, evidence files and anchors, file.py:symbol references, nonclaims, and a cited test for every non-open claim;
  3. validates the paper's claim ledger;
  4. recomputes the real-data receipts;
  5. compares the experiment suite with receipts/experiments.txt;
  6. runs pytest and reports, per claim, whether its cited tests passed.

  --quick skips the nine slow exhaustive tests. --online re-downloads the public data. CI runs verify on Python 3.11 to 3.13 from uv.lock.

• Real data.

  • The Delft 2015 loophole-free Bell test recomputes from the released trials to S = 2.42 +- 0.20, p = 0.039, as published. The point sits right of the S = 2 chord on the canonical CHSH disk.
  • All 256 NIST beacon precommitment links verify, and the composite Bayes factor on 131072 bits never reaches ln 100.
  • Claims ST-F50, ST-F51, ST-F52, ST-T271 and ST-M83 are new. They were renumbered from ST-F32 to F34, ST-T185 and ST-M50, which the merged research PRs already use.

• Claims. Every non-open claim now cites a passing test, including a direct test for the core result ST-T01. The paper ledger (paper/identifiability-before-anthropic-bayes/claims.json) is checked for unique IDs, assumptions, nonclaims, and the test files its workflow names.

• Docs. Repairs LaTeX that escape sequences had turned into control characters (\frac became a form feed, \nu a newline) and drops em dashes.

• README. States the core result plainly. Lists results by type: 262 theorems, 83 model results, 51 finite checks and 1 open problem, across 53 manifests. Adds lanes for the merged Bayesian Boolean, latent-prior and mixture, and active-experiment work.

Verification

uv run python -m simtheory verify (full) on this branch, rebased on main after #47 to #70:

ok    53 manifests, 397 claims (51 finite_check, 83 model_result, 1 open_problem, 262 theorem)
ok    paper/identifiability-before-anthropic-bayes/claims.json: 8 paper claims
ok    Delft 2015: n=245 k=196 S=2.422+-0.204 p=0.0392 disk q=(0.856,-0.094)
ok    NIST beacon: 256 precommitment links verified, 131072 bits, max log BF 1.48 < ln 100
ok    151 lines match receipts/experiments.txt (relative tolerance 1e-9)
ok    587 passed, 0 failed, 0 skipped across 134 files
ok    404 claims have every cited test file passing
verify: all checks passed

A passing check validates code and a bounded receipt. It does not promote a finite check into a theorem, and none of this is evidence for or against simulation.

🤖 Generated with Claude Code