Live eval harness: run every skill's refusal evals against a model (split rubric, CI-gated)

オープン
#61 コメント 7 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
20/100
issue の種類
機能追加
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
github-actions, python

調査の方向性

Start with scripts/test-ci-gates.sh, docs/roadmap.md, and the existing .claude/skills/*/evals/evals.json files to understand the current validation and eval formats. Then map out the schema changes, live eval runner, scheduled and dispatch workflow, and offline PR gate described in the issue. Done means all evals are covered, hard and soft axes use their specified checks, spend and judge separation are enforced, and a known-bad fixture fails the gate.

索引モデルが issue の本文から書いたものです。

説明

enhancement

The suite's load-bearing safety property — per-skill refusal evals — is only shape-validated today (scripts/test-ci-gates.sh), never executed against a model. Turns docs/roadmap.md Loop 2 (deterministic gates for boundary/non-fabrication + a decorrelated LLM-judge for soft axes) into built infrastructure.

Where: new scripts/run-evals.* consuming every .claude/skills/*/evals/evals.json; a scheduled/dispatch workflow; a schema extension (axis tag hard|soft + a machine-checkable predicate per hard expectation) across all evals + the shape validator.
Done when: harness runs each eval prompt against a configured subject model; deterministic predicate on hard axes, LLM-judge on soft only; aborts when judge==subject model id (decorrelation by identifier, no vendor name); a per-run spend cap is enforced before each call; writes a dated results artifact (per-eval pass/fail, model ids, spend); model choice in env/config only — no vendor/model name in any skill's prose.
Verify: live run on schedule/dispatch against main (where the secret lives), never on PR; a separate deterministic offline PR gate asserts the results artifact exists, is not stale, and covers every eval id — failing closed (not skip-when-key-absent). A golden known-bad fixture must be rejected by its predicate, proving the gate can fail.
Why: today a skill could silently start fabricating a TAM, CWE, or compliance verdict and every gate stays green. Single biggest hole in the safety story. Grounded by two independent review agents this session.

主要言語
Python
スター
2
フォーク
1
平均マージ
1時間 29分
マージ済み PR(30日)
551

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

protocol/perun のほかの issue

protocol/perun の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。