[Yigtwxx]

What this changes

• Opt-in accurate NLI profile. It is the winner of the NLI bake-off: MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli, int8, run with k=2 and its own cuts (decide 0.25, medium 0.252721, high 0.999142). The cuts were calibrated on SciFact train, and rule-decided rows were left out of the tier walk. Turn it on with [models] nli = "accurate", proofpath config set models.nli accurate, /config, or proofpath check --accurate for one run.
• Consent for the 643 MB download (product rule 5). The new permission permissions.install_model (ask/allow/deny) works like the browser gate: once, always, no or never, and with no TTY ask means deny (rule 4). --accurate picks the profile but does not grant consent.
• Reported fallback (rule 6). A refused or failed download falls back to the default profile, and the run says so three ways:
  • a note with advice that fits the reason;
  • nli_requested in the report's models block;
  • an nli line in the CLI that -q does not hide.
• Tier honesty. A run whose medium cut sits on decide now says it has no real low tier. This applies to default runs too, since their medium cut is 0.008 above decide. The models block now records k.
• TUI. /config has models.nli and permissions.install_model rows, and the permission prompt also asks the model question. /allow, the buttons and quit all answer it.
• The default profile is unchanged. A test pins its verdict-cache key, which matches main byte for byte. The eval scripts are pinned to the default model.
• Docs: README "Accurate model (opt-in)" section and Measured row, spec §7.1/§12/§14, CHANGELOG, OPEN-ITEMS §20, site Measured data.

Why

The bake-off measured SciFact dev macro-F1 at 0.697 for this model, against 0.580 for the default. It is not better on news and fact-check pages (AVeriTeC readable 0.475, against 0.537 for the default model at k=2; both are below the 0.708 majority baseline), so it ships as an opt-in for scientific sources rather than as the default. The default install stays small.

Every task had an independent reviewer and a re-review. A final whole-branch review answered READY TO MERGE after one Important fix: the eval scripts had followed the user's profile. Live runs were checked twice: --accurate with the model cached (k=2, accurate cuts), and with the model missing and no TTY (no download; fallback reported three ways).

Checklist

• ruff check and ruff format --check are clean (and mypy --strict)
• pytest passes (2667 passed, 1 skipped)
• Type annotations on every new function signature
• No network access in unit tests
• If a new reported state was added, spec section 15 is updated too (no new state; a models entry and notes only)
• This change does not make the tool sound more confident than its evidence

🤖 Generated with Claude Code

https://claude.ai/code/session_01KRNnJrkUuFNdBcR5vXRgTp

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
proofpath	Ignored	Preview	Oct 1, 2026 10:32am UTC