[ErikaOnFire]

The gate moved to HumaneBench rubric v4 in f56f03d (rubrics/rubric_v4.md, upstream 68a872e, sha256 6e81d24…, byte-identical to humanebench main). RUBRIC_DELTAS.md still said it was pinned to v3. Docs only: no change to the code, the prompts or the vendored rubric.

What changed

• New header: which rubric runs, the pinned commit and sha256, and that the gate vendors the rubric spec, not judge_prompt_v4.md.
• Each delta re-checked against judge.py, prompt.md and the v4 text:
  • Dropped because v4 now has them: abstain by default, unless, the confidence filter, evidence verification, suppressions reported as covered. They're listed under "Carried from v4, not deltas".
  • Dropped because it no longer applies: "global rule 1 is not applied". v4 narrowed the rule and prompt.md no longer exempts it.
  • Added where the gate falls short of v4:
    • no escalation for a non-floor covered item whose would_have_been is -1.0
    • no default floor when humane-policy.toml is missing
    • "single closest principle" versus global rule 10
    • unresolved questions raise the verdict to review
    • one acceptance covers every finding sharing a principle+file id
    • document-mode acceptances aren't version-pinned
    • evidence matching is looser than "verbatim"
• A short, scoped note: v3 is frozen for the published v1 benchmark, and gate verdicts can't be compared with it.
• "Not yet reconciled" now lists two open items:
  • prompt.md says any floor finding means discuss, but the code requires -1.0.
  • The prompt never maps v4's outcomes onto the gate's schema.
• README setup step 2 no longer calls the vendored rubric a hand-transcription.

🤖 Generated with Claude Code

[github-actions]

HumaneBench · advisory

🟢 Clear

This diff migrates documentation from rubric v3 to v4 and clarifies the sync-rubric workflow. It changes no code that affects users and introduces no new capability, data handling, or user-facing behavior.

---

_{🟢 clear   🟡 review   🟠 discuss   🔵 needs context   ⚪ accepted by a person. There is no red, because this check does not block anything. Scored against HumaneBench rubric v4, loaded verbatim, plus this repo's humane-policy.toml and the policy documents it names. Findings whose quoted line is not in the diff, or that the judge marked low-confidence, are dropped before posting. Deviations from v4 are in RUBRIC_DELTAS.md. Rubric 68a872e, commit ebec944.}

_{Judged 24 Sep 2026, 08:19 UTC}