docs(evals): align model-judge calibration policy

[aryeko]

Summary

• pin @agentic-workflow-kit/eval-kit to v0.1.5 and refresh the lockfile
• align define-product model-judge docs and calibration template with the shared calibration/reporting policy
• clarify partial/unknown handling, expected-bad adverse evidence, and raw result bundle policy

Verification

• pnpm install
• pnpm install --frozen-lockfile
• pnpm eval:doctor
• pnpm eval:list
• pnpm eval:validate-fixtures
• pnpm eval:judge:doctor
• pnpm eval:judge:list
• pnpm eval:judge:validate-fixtures
• pnpm check
• git diff --check

Notes

Docs/package migration only. PRD/product judge semantics remain local. Provider-backed commands remain outside pnpm check and CI.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 101772e872

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Align report-count policy with generated reports

In this repo, pnpm eval:report uses compileReport, which builds the Item verdicts summary from only the verdicts present in pointwiseResult.items (see evals/adapter.mjs:997-1007), so zero-count categories are omitted. Under this new policy, a report with no unknown or no partial verdicts will not actually include counts for all five categories, making reviewers unable to distinguish zero from not reported under the documented v0.1.5 calibration policy. Please either emit all five categories with zeroes or relax the policy to observed verdict counts.

Useful? React with 👍 / 👎.