Independent evidence audit

Scope and independence

The publication review on 18 September 2026 checked the saved evidence without sending new API requests. Separate implementations reconstructed literal graph truth and statistical estimates without importing historical generators or analyzers. The checks were performed by coding agents under the same project direction. They are independent implementations, not external peer review, independent data collection, or human annotation.

All review choices were made after the historical outcomes were available. They are sensitivity analyses and implementation checks. They do not replace the historical primary analyses or create new confirmatory evidence.

Literal evidence checks

All expected-answer checks pass. These checks are computations on literal inputs, not additional model observations. Every attempted request matches its frozen plan, all attempt identifiers are unique, and recorded chronological dispatch matches the planned order. The single failed rewiring response remains failed.

The audit also verifies inverse naming substitution, exact-copy identity, graph and question preservation across locality and header views, four changed parent references per rewiring intervention, preserved root-basin sizes, and identical edited facts in stale and refreshed layouts. The five worked-example requests reproduce the reported exact answers and saved responses.

Independent statistical reconstruction

All principal point estimates match the historical analyses. A separate 50,000-resample forest bootstrap gives the following post hoc intervals. The article and paper retain their original 10,000-resample intervals.

The grouped-root naming effect is sparse: 10 positive, 2 negative, and 84 zero forest effects. Omitting one forest can reduce its mean to 2.76 points. It is weaker evidence than the large shuffled-root effect. The ancestor locality result also retains the historical candidate-depth shortcut.

Inference and schedule sensitivities

Historical sign-flip tests assume symmetry or exchangeability of forest contrasts under the null. The naming statistic need not be symmetric even when both representations have identical serving distributions. Exact sign enumeration removes simulation error but cannot establish that scientific null assumption. Its grouped-root raw value is 0.005859, compared with the historical Monte Carlo value of approximately 0.004500.

A separate conditional permutation redistributes the ten selected labels within each world and layout among the five two-copy name conditions. It supports the naming contrasts under the stronger assumption that individual serving labels are exchangeable across those calls. With 50,000 permutations, three endpoints have no exceedances and grouped root has six; plus-one values are 1/50,001 and 7/50,001. These are post hoc sensitivities, not tests of cache independence or service stationarity.

Resampling the 24 balanced naming schedule quartets gives a shuffled-root interval of 50.20–59.57 points. Cross-name disagreement remains much larger than repeat disagreement across the observed time-gap bins. Resampling 24 chronological locality blocks gives a root interaction interval of 34.38–47.40 points. These checks make a simple schedule-only explanation less plausible, while leaving hidden service dependencies unobserved.

For the failed rewiring request, every possible selected label gives an all-16-forest grouping effect between 71.09 and 71.48 points. This is a deterministic missing-label bound conditional on the observed records, not a confidence interval.

Recorded local freezes

The naming and rewiring protocol hashes match their pre-run manifests. The locality/header manifests contain their prospective comparisons and inference rules but do not hash-bind the explanatory Markdown protocol. Its current version includes a disclosed post-freeze shortcut paragraph. Local timestamps and hashes are provenance records, not external preregistration or independent certification of timing.

Reproduce the review

Extract the evidence archive and install its pinned NumPy requirement. From the extracted directory:

python3 review/design/audit_literal.py \
  --evidence-root . --output audit-output/literal-audit.json
python3 review/statistics/independent_audit.py \
  --evidence-root . --output audit-output/independent-results.json
python3 review/literature/check_estimator_identity.py
python3 review/supporting/check_supporting.py \
  --evidence-root . --output audit-output/supporting-results.json

The package includes the audit source, original audit results, historical protocols and manifests, and the supporting raw records. Numerical output should agree apart from source paths and exported-file hashes, which differ by the disclosed transformation. The checks require no API key. The MMD identity check exhaustively verifies all 256 four-label outcomes and 16 rational distribution pairs; this verifies algebra, not serving assumptions.

Remaining limits

No comparison model, new domain, independent deployment corpus, or architectural observation was added by this review. The completed evidence supports a controlled service case study. Stronger claims about general representation practices require the prepared anchor-fixed renaming study, additional layout controls, comparative models, and independently authored application data.