feat(v5): experimental Haiku WARN cascade tier (opt-in; F1 honestly caveated)#26
Merged
Conversation
… unchanged) - hooks/no-sycophancy-warn.sh: reference cheap-judge WARN tier; opt-in, regex-negative only, cross-model Haiku, NEVER exits 2 (BLOCK floor owns blocking); bash -n ok, smoke pos/neg pass, exits 0 - evaluation/v5/SPEC.md + RESULTS.md: cascade closes the recall gap in CAPABILITY; F1=1.0 reported as circular/optimistic upper bound, not a production metric; real number needs human gold - no-sycophancy.sh BLOCK path UNCHANGED (git diff main empty) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…saturates Co-Authored-By: Claude Opus 4.7 (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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Opt-in experimental cheap-LLM-judge WARN tier above the deterministic regex floor — the v5 cascade the v3/v4 negative result + deepresearch (
.taste/research/cascade-llm-judge-tier.md) pointed to. Companion data/scorers PR: waitdeadai/agent-closeout-bench feature/v5-cascade-haiku-tier.What this adds
hooks/no-sycophancy-warn.sh— reference WARN tier. Opt-in (LDP_SYCOPHANCY_WARN_JUDGE=1); fires only on regex-negative closeouts; cross-model Haiku judge; never exits 2 (WARN-only — the deterministic regex floor owns BLOCK).bash -n✓; smoke WARNs on a positive, silent on a clean negative, exits 0.evaluation/v5/SPEC.md+RESULTS.md.no-sycophancy.shBLOCK floor unchanged (git diff mainempty).Results — read the caveat, do NOT treat F1 as a win
On the held-out (n=58) + fresh (n=35), regex∪Haiku-WARN scores F1=1.000 (regex-only 0.26/0.23), all previously-missed modes recovered, 0 control FPs. This is an optimistic/circular upper bound, reported as such: the frozen judge labels were κ-validated against the same construction gold; positives are synthetic; control set is tiny. Cross-judge check: Haiku==Sonnet (κ=1.0) on the WARN cases → cheap tier suffices, but the universal 1.0 shows the corpus saturates — too unambiguous to validate real-world precision. An unbiased number needs human gold or an ambiguous/real-trace test set.
Recommendation
Merge as an opt-in experimental tier + honest eval, not as a validated detector. The headline F1 claim is deliberately withheld pending non-synthetic validation.
🤖 Generated with Claude Code