BenchmarkProof ledgerAudit and governanceEcosystemKrenn·IQAtlas
Review layers

The review consists of three re-runnable layers.

Two of the three layers are run by Krenn·IQ; only the §6 cross-model adversarial review is external. The layers: a constitution that fixes what may be called proven; a cross-model adversarial review, in which a model from another company challenges the public claims with executed commands; and reproducible conformance, an invariant suite whose CI result carries signed provenance. The Gate App’s decision record lives under Decision log.

Layer 1 · Governance

A written constitution fixes the vocabulary.

GOVERNANCE §2 fixes the states a claim may hold (proved, derived, empirical, conjecture…). §6 fixes the constructor ≠ verifier rule: the builder of a claim cannot declare it proven. On this site: three labels. CLOSED: a locked test that cannot regress. BACKED: regenerates from a command and its hash matches. SCOPED: a declared estimate with its corpus and the artifact that would upgrade it. The proof ledger (28 entries) and the claim registry regenerate from code; the release blocks if a public number has no label.

Charter: GOVERNANCE.md, binding in direction, not yet legally executed · Responsible disclosure: SECURITY.md.

Layer 2 · §6 cross-model adversarial review

A model from another company attacked 7 of 10 public claims with executed commands and refuted 3.

The §6 package holds 10 public claims, each with its evidence and a refutation procedure. A model of another family, xAI grok-4.7, directs; a script runs its commands under a read-only allow-list, no network, no writes, and returns the raw output unedited. Eight runs on 2026-10-13, about 4.5 million tokens; a ninth aborted when the credit ran out.

  • 7 of 10 claims tested; 3 refuted, each corrected and frozen as a test or sealed file the same day.
  • C6-02 (scope false-accept): refuted three times. Two regex fixes failed; the structural closure holds: a bounded ACCEPT requires the scope declared in a field.
  • C6-04 (unsealed figure): 52/81 was an unsealed recodification, re-sealed as 53/82 and 51/63 with hash.
  • C6-05 (unversioned grading rule): now a script that reproduces the exact counts.
  • Pending: C6-08, 09 and 10 (the 2026-10-13 results) were never attacked; the reviewer credit ran out. One run, invalid by a sandbox fault, is recorded as such.

Transcripts and runner: research/section6_package/ · refutations frozen as tests: benchmarks/test_section6_refutations.py. Nothing is marked proved; a claim not refuted by one reviewer family stays empirical.

Layer 3 · Reproducible conformance

Ten invariants, one reproducible result hash, signed provenance on neutral infrastructure.

The conformance suite runs ten invariants: fail-closed, robustness, no model in the verdict, research-layer decoupling, quantum commitment, pharma admissibility, proof ledger, HOLD resolution, audit hash, surface isolation. It produces a reproducible result hash. In CI, on GitHub infrastructure, the result is signed keyless (Sigstore, Rekor) and published as an artifact. That signature is signed provenance: it shows which CI run produced the result. A signature over a CI result is not a scientific validation.

# run it yourself
python3 benchmarks/conformance.py

# verify the CI-signed result (artifact: capas-conformance-attestation)
cosign verify-blob conformance_result.json \
  --certificate conformance_result.pem --signature conformance_result.sig \
  --certificate-identity-regexp 'github.com/fomv9354lve/capas-inteligentes' \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

Workflow: .github/workflows/conformance.yml · OpenSSF Best Practices badge: passing, self-certified with static and dynamic analysis · CodeQL and Scorecard weekly.

Limits

Limits of the review layers.

  • No SOC 2 / ISO 27001 certification; roadmap item.
  • The blind cross-domain adjudication study (Gap-1) has not run; its registration and gating pass.
  • The governance charter is not legally executed; the trustee is an open item.
  • No independent external audit has been done.
  • The cross-model adversarial review uses one model family (xAI) so far; the second waits on credit.
  • The blind numbers on real claims are the sister-corpus figures: 53/82 and 51/63 on X1–3, and 17/21 and 16/17 out of sample on X4+X5, with intervals.