# Verification record

Publication handoff prepared on 17 September 2026.

## Completed

- Reconstructed all model-question scores directly from raw grades and gold. Maximum difference from the saved original scores: 2.842170943040401e−14.
- Verified the archived system prompt and all 100 exported user-prompt hashes.
- Checked served model/provider fields for successful candidate calls against saved run headers.
- Preserved and enumerated the 20 malformed coverage-ID records and one missing full-pass grade.
- Verified 1,700 unique model-question outcomes, 1,693 successful generations, 3,385 completed grades, 751 source rows and 29 publishers.
- Recomputed both original 5,000-resample analyses in a disposable copy of the packaged snapshot. The resulting full-field data and production comparison JSON matched the archived outputs exactly.
- Recomputed publication tables in a separate output directory and compared the resulting JSON objects for exact equality.
- Checked score aggregation, failure zeros, component decomposition, model settings, claim evidence paths and Markdown links.
- Checked ZIP CRC integrity, safe relative paths, content manifests and archive hashes. Environment files are excluded; the packaging scan found no OpenRouter key literal or private-key block in included text files.
- Workspace `pnpm typecheck` passed. Application and website code were not changed.

## Limits of verification

The checks establish internal reproducibility and artifact integrity. They do not authenticate remote model weights independently, certify factual correctness, establish source redistribution rights or substitute for the pending native-Nepali/human gold review.

The ten figure assets are declarative Vega-Lite specifications, not pre-rendered figures. Their embedded numerical data and references were checked; the website session should render and visually verify them in its chart stack. The complete copied HTML report was browser-tested in the preceding analysis work; the new editorial draft and new length/rank/kappa analyses are additional publication assets, not new sections silently added to that earlier report.