Private-evidence publication audit — 2026-07-22

• Issue: #175
• Public repository base: 85b0e40053c56f004fda960a00a00b72ee4fd1f8
• Companion repository inspected: fakoli/anvil-serving-notes at 7b46ceb6ae62252f8f808f6c065706a24e7970bb, after fetching origin/main
• Evidence type: repository inventory, sanitized publication, and offline deterministic rerun
• Outcome: public grounding restored; one unavailable raw-log set recorded as an explicit gap

What was published

Three load-bearing narratives were present in the companion repository and are now public:

• Anvil integration audit
• planning-capability evaluation
• harness intent-routing research

They also remain recoverable without private access. The June 28 narratives were introduced at 21f9a81f9be98dab3be15b07395ab34749d852b6, and the harness research at b0a68c64482774a719da76a62a745e095effda1e. The strongest single public snapshot is 31d95adaf68157b81318325356516cef9569b10f: immediately before relocation, its complete bundle tree 6e26ebc80b853e151075e8807c0bba084480f823 exactly matches the fetched notes mirror. The notes revision is an inventory input, not the public chain's sole provenance.

The planning evaluation's complete bounded bundle is public under eval-data/2026-06-28-planning-capability/: 21 canonical source files totaling 159,272 canonical Git bytes before three enumerated publication-safety edit categories. It includes both rendered prompts, all six model outputs, generation metrics, deterministic structural scores, four blind-judge records, the anonymization map, aggregates, and the historical harness. No file exceeds 1 MiB and the bundle is below ADR-0027's 5 MiB limit.

The pre-publication bundle matched no user-home paths, emails, credentials, authorization values, session identifiers, or secret-like values. The integration narrative's one machine-local checkout path was replaced by its public repository and pinned commit. Source SHA-256 digests and every publication edit are recorded alongside the bundle.

What ran

From the source snapshot and again from the public copy, python grade_struct.py followed by python aggregate.py reproduced the tracked structural scoreboard and judge aggregates. The rerun left the generated JSON/CSV unchanged. This proves the offline calculation path over the published inputs.

The 2026-06-28 local/frontier generations and blind judge calls were not rerun. Their models and topology are historical; the publication does not relabel them as a current model qualification.

Mirror claims checked

PR #174 removed these three public files while saying their home was the private notes repository:

• docs/PLAN-advise-and-defer.md
• docs/REVIEW-2026-07-02-architecture-and-models.md
• docs/OPENCLAW-LIVE-VALIDATION.md

None is present under its exact path or as a byte-for-byte mirror in any fetched companion-repository ref. Related descendants do exist: the advise-and-defer PRD, the v0.4-to-v0.7.2 reflection, and docs/findings/2026-06-30-openclaw-live-validation.md, which records results from the removed validation runbook. The exact removed files remain recoverable from the public product repository at cf09b06bd63d6d4013a67b59d581a8689a804cf9. They are planning/review records, not the load-bearing evidence published above. Current behavior is covered by public reference docs, ADRs, and findings, so no public claim depends on pretending those files were mirrored.

Unavailable evidence

ADR-0008's speculative-decoding raw benchmark logs were never committed to the product repository. A search of the fetched companion history for its exact throughput, cache-hit, model-class, and flag signatures found no matching record. The evidence-gap finding makes that absence public and limits how the historical ADR may be used.

Current-doc impact

The quality-gated router reference, Primary Node runbook, OpenClaw plugin docs, relevant ADRs, and changelog now link public evidence or explicitly classify private material as non-load-bearing. Private design history remains supplementary under ADR-0027.