← Về thư mục
📄 / / root / ceo / academic-research-skills / .github / pull_request_template.md

Summary

Eval impact

The eval harness (.github/workflows/eval-harness.yml) runs automatically on PRs that touch scoring / generation logic or the gold sets (see the Delta 3 path filter in that workflow). Most PRs do not affect eval metrics — leave this section as "No eval impact." if that applies.

If your change alters ranking / scoring / generation behavior and moves a gold-set metric:

  1. Declare each affected metric, one per line, in the exact form:

Affected metric: <task>.<class>.<metric>

e.g. Affected metric: citation_extraction.aggregate.accuracy (use class aggregate for the headline metric; otherwise the per-class name).

  1. If a metric regresses (polarity-corrected signed_lift < -0.05, or any zero-baseline metric changes), the gate blocks unless you add BOTH:

  2. the acknowledgement token (on its own line):

    • [eval-regression-acknowledged] — for the CI deterministic gate, and/or
    • [ranking-regression-acknowledged] — for scripts/check_ranking_lift.py
  3. a link to an OPEN follow-up GitHub issue, e.g. https://github.com/Imbad0202/academic-research-skills/issues/NNN

No eval impact.

Platform port?

If this PR adds a new <platform>/ directory, read the Platform ports policy and open a design issue first. Otherwise, leave this section as "Not a platform port."

Not a platform port.

Checklist