HarnessBank: gene-bank search plus gated screening for trustworthy harness self-evolution

Luo, Xue, Wang, Hu, and Deng treat harness self-evolution as a trust problem, not just a search-budget problem. Greedy “keep what looked better once” loops collapse into prompt-safe edits, retain inert patches, and credit noise. HarnessBank pairs a frozen task agent with a separate evolver and keeps a Harness Gene Bank indexed by semantic coordinates — where an edit acts (prompt / knowledge / runtime / config) and why (targeted failure pathology) — so complementary mechanisms stay available for reinvention and recombination instead of being discarded by a single lineage.

Offspring do not get a free full-set eval. Gated Harness Screening runs a cheap subset through four sequential gates: validity (protocol-complete ledger, not sandbox-crash as “fail”), activation (the patch must actually fire — beacon required), paired significance (task-level Δ clears z ≥ 1.96), and gain over parent. Only survivors pay for full training-set evaluation and compete for cell admission. On seven domains with frozen Qwen3.6-27B, train-selected harnesses lift sealed test Pass@1 by +5.1% to +15.4%; recurring wins are runtime recovery for thinking-runaway empty turns stacked with verify-finalize self-checks. Under the same protocol and paired-2σ ruler, HarnessBank credits four of five sealed comparisons versus one for DGM and none for prompt-only GEPA. Cross-model runs show the transferable asset is the diagnosis–search–verify process: a patch matched to one model’s pathology is near-zero or harmful on another.

Why it matters for code review

Continuous factories that evolve harnesses overnight need merge criteria stricter than “mean went up.” Review harness PRs the way HarnessBank screens candidates: require an activation proof, a paired significance receipt, and a semantic reason the edit belongs in the bank — not a one-shot spike on a tiny subset. Steady harnessed throughput compounds when FOMO for more evolution rounds is spent on gated, recombinable mechanisms rather than bursty prompt thrash that never fires in production.