dokimi 0.1.0

First release on PyPI. Requires Python 3.11 to 3.14.

Added

• dokimi --version; runs and scorecards record dokimi_version.
• tools preference: how much one change to a tool moves a model's choice between two equal tools; --max-cost-usd caps its spend.
• repo compare: treatment arms against a baseline arm, paired by task.
• context trace: prompt size of every model call in a Claude Code trace.
• novelty --max-usd spend cap; novelty routing --catalog tests against the real sibling skills.
• Guidance-only repo manifests: repo scan, reach and validate without tasks.
• repo run --build-timeout and repo tasks check --build-timeout (default 3 h).
• Arms artifacts store the exact text the judge read; eval cross-judge --judges re-scores on it.
• Novelty bundle records name what was measured (artifact_name, artifact_ref, measured_digest).
• Documentation site at https://intel.github.io/dokimi/.

Changed

• Default models: answers and generation openai/gpt-6.1-sol, judge anthropic/claude-sonnet-5-5, on every command. Scores differ from earlier defaults.
• --judge-provider alone picks that provider's default judge model; google, google-vertex and openrouter require --judge-model; openai-codex is no longer offered.
• Question sets are written as questions.v2 (question instead of text); v1 sets still load.
• A golden set under --questions-root is answered as-is: no question generation, no persona discovery.
• Every judge path refuses a record with no question text before the first judge call (exit 2).
• eval panel-score checks numeric claims the way the single judge does. Panel scores differ from earlier runs.
• harmful_atoms reports measured: false when the conflict check did not cover every atom pair.
• command:, Harbor, BenchFlow and explorer timeouts kill the whole process tree.
• Bundled registries and the spec schema ship inside the package.
• RAGAS is an optional extra: eval ragas and answers generate --ragas need pip install 'dokimi[ragas]', and exit 1 before any model call without it.
• beautifulsoup4 and anthropic are no longer direct dependencies.

Removed

• Python 3.10.
• The top-level cli module is no longer installed; use dokimi, python -m dokimi or dokimi.cli.

Fixed

• eval panel-score judged every answer with an empty question and no context. Panel scores differ from earlier runs.
• Judge and panel replies with non-finite or out-of-range scores are rejected and not cached.
• The trust gate's retrieval hit rate read a field nothing wrote, so it reported 100% for runs with no documents.
• tasks run could score a failed cell as passed (relative --output-dir, or a stale run dir that could not be cleared).
• Question id 0 was read as missing.
• The agentic arm ran a tool on {} when its arguments did not parse; failed calls are now recorded.
• Novelty no longer counts truncated or failed replies, non-atom rows or non-boolean verdicts as results.