feat: MAD Confidence Scoring + HermesJudge [polished]#21
Open
AIandI0x1 wants to merge 1 commit intoNousResearch:mainfrom
Open
feat: MAD Confidence Scoring + HermesJudge [polished]#21AIandI0x1 wants to merge 1 commit intoNousResearch:mainfrom
AIandI0x1 wants to merge 1 commit intoNousResearch:mainfrom
Conversation
Statistically rigorous quality gates for skill evolution. ## New files - evolution/core/mad_scoring.py (387 lines) - MAD confidence scoring - evolution/core/hermes_judge.py (146 lines) - LLM-as-judge via hermes chat - docs/MAD_CONFIDENCE.md (70 lines) - technical documentation - tests/core/test_mad_scoring.py (179 lines) - test coverage ## Modified files - evolution/skills/evolve_skill.py - holdout MAD, proof.json, CLI flags - evolution/core/config.py - GPT-5.4 defaults, Nous API - evolution/core/fitness.py - temperature=1.0, MAD re-exports ## Key features - Confidence = |mean_delta| / MAD on holdout deltas - Labels: likely real (>=2.0x), marginal (>=1.0x), within noise (<1.0x) - HermesJudge: no API key management, uses hermes chat subprocess - proof.json: captures MIPROv2 optimized instructions + demos - --mad-trials and --judge-model CLI flags - Fix: constraint validator now checks full skill (with frontmatter) ## Proven results - Arxiv regression caught: -11.2%, confidence 3.32x (likely real) - Self-evolution: +13.4%, confidence 1.03x (marginal) - Near-optimal baseline: -2.0%, confidence 1.00x (within noise) No new pip dependencies. Requires hermes CLI for HermesJudge. See also: PR NousResearch#20 for full raw work with results and proofs.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Statistically rigorous quality gates for skill evolution.
What
When you evolve a skill, the system now tells you whether the improvement is real or noise — with proof.
Math:
confidence = |mean_delta| / MAD>= 2.0x→ "likely real" — keep>= 1.0x→ "marginal" — borderline< 1.0x→ "within noise" — discardNew files
evolution/core/mad_scoring.pyevolution/core/hermes_judge.pyhermes chatsubprocessdocs/MAD_CONFIDENCE.mdtests/core/test_mad_scoring.pyModified files
evolution/skills/evolve_skill.py--mad-trials/--judge-modelCLIevolution/core/config.pyevolution/core/fitness.pytemperature=1.0on judge LM, MAD re-exportsProven results (see PR #20 for full evidence)
Usage
python -m evolution.skills.evolve_skill --skill arxiv --mad-trials 3 --judge-model "gpt-5.4"Dependencies
hermesCLI (for HermesJudge subprocess)Companion PR
This is the clean code PR. Full raw work with results, proofs, and test scripts is in PR #20.