{
  "audit_title": "Where the Checking Stops",
  "audit_version": "2.0",
  "as_of": "2026-08-31",
  "auditor": "Brennan McCloud",
  "source_repository": "openai/ten-proofs",
  "source_commit": "94bc0feb6a9ff12c7d31d6de640a725c9d43d2b6",
  "public_commit_date": "2026-08-02",
  "announced_publication_date": "2026-08-01",
  "lean_toolchain_pinned": "leanprover/lean4:v4.32.0",
  "result_advances": 10,
  "formal_result_bundles": 12,
  "comparator_configurations": 12,
  "theorem_targets": 38,
  "solution_sorry_count": 0,
  "permitted_axioms": [
    "propext",
    "Quot.sound",
    "Classical.choice"
  ],
  "nanoda_requested_in_configurations": 12,
  "definition_hole_bundles": [
    {
      "bundle": "H_GapCVP",
      "theorem_targets": 4,
      "definition_holes": 4,
      "definitions": [
        "GapCVP.Comparator.gapCVP400Promise",
        "GapCVP.Comparator.binaryNearestCodewordPromise",
        "GapCVP.Comparator.binarySyndromeDecodingPromise",
        "GapCVP.Comparator.finitePGapCVPPromise"
      ],
      "implication": "Comparator validates their names, types, safety, axioms, and kernel acceptance; its documentation requires an additional verifier for definition-hole solutions."
    }
  ],
  "findings": [
    {
      "id": "A",
      "title": "A different validation boundary",
      "summary": "The GapCVP bundle includes four definition holes in addition to theorem holes, so its assurance path is structurally different from the other bundles."
    },
    {
      "id": "B",
      "title": "Verification freshness is part of the evidence",
      "summary": "The public repository pins Lean 4.32.0; a Comparator-relevant kernel issue was fixed in 4.32.2 before the public repository commit. All configurations request nanoda, but the repository does not pin the external checker version or publish a signed recheck attestation."
    },
    {
      "id": "C",
      "title": "Artifact volume is not assurance burden",
      "summary": "The prior 99.8% versus 0.2% comparison was a byte ratio, not a measure of confidence, effort, or consequence."
    },
    {
      "id": "D",
      "title": "Semantic faithfulness is instrumentable",
      "summary": "Recent work shows that typechecking and provability are insufficient, while independent probes can detect many forms of statement drift. Final domain acceptance remains an accountable human or organizational decision."
    }
  ],
  "limitations": [
    "This is an assurance and release-artifact audit, not peer review of the underlying mathematics.",
    "This audit does not claim that any of the twelve formal result bundles is false or exploits a soundness bug.",
    "The audit does not independently rebuild every formalization from source.",
    "The Faithfulness Gap results referenced are from a 2026 preprint."
  ]
}