# R1 audit reconciliation — Codex

Both complete actual provider responses were copied, hashed and read. Claude Opus5 High returned BLOCK for one concrete defect; Grok Heavy returned ACCEPT WITH LIMITS and did not identify it. No majority vote overrides a reproduced failure. R1 remains immutable historical evidence, not the release candidate.

## Reproduced defects and R2 changes

- Claude B1 reproduced against immutable R1: two complete captures for the same scheme/version, with retired/nonselectable versus active/selectable codeA and different enumerations, produced candidate-only. R2 checks semantic source coordinates packet-wide. All payload fields must agree for same scheme/version, profile id/version, assignment id/revision or crosswalk id/version. Equal payloads in distinct capture envelopes remain allowed. This is internal consistency, not authentication of source statements.
- Author probe found cached ancestry could accept9 profiles despite cap8. Both reviewers' general ancestry assurances missed this. R2 counts the cached ancestry length plus current traversal depth;8 passes,9 refuses.
- Claude observation11 confirmed author probe: minimum33 could pass profile-only despite instance cap32. R2 returns unsupported. Maximum above32 remains meaningful when a smaller allowed selection satisfies the minimum.
- Duplicate selected crosswalk IDs now refuse. Captured correction of a selected historical assertion and proposed assignment state now add notices without silently replacing the host's selected revision.
- Public review validates returned assessment and complete replayable packet/result against closed schemas and1MiB cap before returning. A near-capacity regression verifies explicit refusal and unchanged packet. No evidence truncation.
- Concept validity exclusions now retain outside-valid-time in detailed results. Tests cover deprecated and time-expired target, mismatched-context competitor, exact repeated capture, other source-coordinate conflicts, correction/proposal notices and empty-source packet preservation.

## Every nonblocking observation disposition

Claude1: notice adopted; historical host selection retained. Claude2: same-coordinate payload consistency adopted for all snapshot kinds. Claude3: duplicate crosswalk pin guard adopted. Claude4: no host-grant authenticity claim is added; grant provenance remains independently host-retained, now explicit in adoption/binding docs. native_records performs no authorization. Claude5: binding text now explicitly says fact identity and that separate object validation belongs to V3/host. Claude6: public result/joint-size validation adopted. Claude7: chain-requested unsupported in every mode explicitly documented. Claude8: precise concept interval outcome adopted. Claude9: proposed-state notice adopted. Claude10: irrelevant acknowledgments remain captured historical basis and may change assessment identity; documented, no implicit permission. Claude11: capacity guard adopted. Claude12: alternatives are mode/source-context dependent; absence is not proof of no correspondence, full packet retained. Claude13: failure of acceptance harness raises without a success report; only actual completed reports are evidence. No invented failure report. Claude14: review files are post-audit release bookkeeping; research artifacts explicitly linked to separate public dossier. Claude15: selected meaningful gaps added, including packet-preservation assertion.

Grok1: unused new-use flag remains private code; actual target activity checks are in migration, no new interpretation added. Grok2: coarse eligibility labels remain conservative, no candidate from unknown release or inactive target. Grok3: no applicable mapping cannot satisfy the nonempty candidate rule. Grok4-5: nonapplicable/empty-source correspondence remains in packet; test improved. Grok6: detailed interval outcomes clarified. Grok7: proposed-state notice added. Grok8-9: trusted-host and uniform deny boundaries retained. Grok10: review files generated later. Grok11: JSON is deliberately used within YAML-compatible files. Grok12: contrary to its static assurance, the independent cache counterexample was reproduced and fixed.

## Scope retained

The schema, external dependencies, source semantics and authority boundary are unchanged. No live classification, source authentication, generic reasoning engine, production authorization, grant archive, object-validator implementation, existing-Dimension migration or standards certification is introduced. The source package remains a bounded original offline review companion. Neither provider ran tests, hashed files or audited excluded upstream/generic toolchain bytes. R2 needs a separate frozen delta audit in both existing conversations after regenerated author tests and native acceptance.