v5.1.0 — 7 Reference-Inspired Improvements

What's New

7 improvements from deep analysis of the 9 references in the README.

Tier 1 — High Impact

Full History Archive (Meta-Harness: 2x improvement with trace access)
New tools/archive.py saves ALL candidates (winners + losers) — diffs, proposals, scores — to evolution_archive/. Proposers can grep prior approaches for cross-iteration causal reasoning.

Pairwise Evaluation (LangSmith, Context Engineering)
read_results.py --pairwise "exp_a,exp_b" compares candidates head-to-head by per-example win count. Triggered when top 2 are within 5%. More reliable than independent scoring.

Archive Branching (DGM, AlphaEvolve)
New archive_branch lens lets proposers fork from non-winning ancestors. Prevents premature convergence by preserving diverse stepping stones.

Tier 2 — Medium Impact

Wave-Based Proposers (Self-Organizing Agents: +14%)
Proposers split into 2 waves: critical/high lenses run first, medium/open see wave 1 results before starting. Research shows +14% quality when agents observe prior outputs.

Failure-to-Regression-Gate (Traces Loop)
regression_tracker.py --auto-guard-failures adds currently-failing examples as "known hard" guards. Every fix becomes a permanent regression test.

Few-Shot Evaluator (LangSmith)
Evaluator agent loads human corrections from LangSmith as calibration examples. Scoring accuracy improves over time without code changes.

MAP-Elites Diversity Grid (AlphaEvolve)
read_results.py comparison now includes a diversity grid — best candidate per approach category. Preserves diversity even when one approach dominates in score.

Tests

12/12 passing (added test_archive).

Install

npx harness-evolver@latest

🤖 Generated with Claude Code