Repository navigation
feat(gates): ground judge verdicts in the diff (arXiv:2606.28050) + 2026-06-29 feed triage - #413
Merged
Conversation
…eed triage ArXiv 2026-06-29 triage surfaced "Can LLMs Judge Better Than They Generate?" (2606.28050): judges attend to context 3-5x less than generators and barely read the candidate, and eval-tuning induces over-acceptance. Audited our judge prompts: the witness APPROVAL path required no grounding (skim-approve hole) and the critic rationale didn't force engaging the diff. Conservative, prompt-only grounding hardening on both: read the whole diff before deciding; approved=true asserts you checked the real changed lines; the witness summary must state what the diff changes; the critic rationale must cite specific changed lines. No schema change. Regression-checked via critic-calibrate — hardened 93% / 0 false-approve / 2 false-reject vs baseline 89% / 0 / 3 (false-approve held at 0; within-noise, so: no regression, not a measured win). Full suite 3078 passed. Also logs the curated 2026-06-29 feed triage to mind/research (the two relevant papers; the GitHub-issues source reframed as foreign-PR fuel, gated on allowlisting). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
elementalcollision
added a commit
that referenced
this pull request
Jun 29, 2026
…eign issue (#414) Chronicle entry for the day: (1) arXiv:2606.28050 audited our own judge stack and exposed the witness approval-path skim-and-approve hole → diff-grounding hardening, regression-checked (#413); (2) first externally-filed foreign issue resolved as Chimera — safetensorstogguf#6 (llama.cpp Model→ModelBase), diagnosed + fixed as a draft PR with the can't-run-it boundary named. Reach plus honesty. Docs-only under mind/**. Co-authored-by: Chimera-Agent <agent@chimera.local> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
From the 2026-06-29 arXiv triage: "Can LLMs Judge Better Than They Generate?" (2606.28050) finds judges attend to context 3–5× less than generators and barely read the candidate, with eval-tuning inducing over-acceptance — a direct hit on our judge-heavy gate stack.
Audit → fix. The witness approval path required no grounding (skim-approve hole); the critic rationale didn't force engaging the diff. Conservative prompt-only hardening on both: read the whole diff first;
approved=trueasserts the real changed lines were checked; the witnesssummarymust state what the diff changes (grounds approvals); the criticrationalemust cite specific changed lines. No schema change.Regression-checked via
critic-calibrate: hardened 93% / 0 false-approve / 2 false-reject vs baseline 89% / 0 / 3 — false-approve held at 0, false-rejects didn't rise. Honest: single-shot n=28, so it's no regression, not a measured win (the benchmark doesn't exercise the skim-approve failure the change targets).Also logs the curated 2026-06-29 feed triage to
mind/research/— the two relevant papers (this one + "Govern the Repository, Not the Agent") and the GitHub-issues source reframed as foreign-PR fuel (gated on allowlisting + the B.4 envelope).Full suite 3078 passed, ruff clean.
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com
🤖 Generated with Claude Code