evolve — Darwin Mode self-improvement

kimi-k3-harness ships with Darwin Mode (@metaharness/darwin, ADR-070…146): the model is frozen; the harness evolves. Each generation mutates ONE of the 7 surface files (planner, contextBuilder, reviewer, retry/tool/memory/score policy), sandboxes each child, scores it, and keeps only variants that measurably improve — building an archive of successful descendants.

Run it

npm run evolve        # real substrate: runs your test command per variant (deterministic mutator — no API key, no network)
npm run evolve:dry    # mock substrate: fast, fully offline, no test execution

Or directly:

npx metaharness-darwin evolve . --sandbox real --generations 3 --children 4

Safety (secure by default)

• Deterministic mutator is the default — no network, no API key, air-gapped.
• Every mutation passes the validateGeneratedCode gate: no new imports, network, filesystem, shell, env access, or dependencies — pure refactor/tuning only.
• Mutations run in a sandbox; only variants that pass your tests are archived.
• Nothing is promoted without measured improvement (guard against Goodharting).

See @metaharness/darwin for selection strategies (--selection, --crossover, --curriculum), statistical gates (--fdr, --bench), and the real-LLM mutator (library API).

What the benchmarks taught us (measured, full SWE-bench Lite 300)

Defaults worth carrying into how you evolve and run this harness (full evidence + CIs in @metaharness/darwin's LEARNINGS.md / bench/results/RESULTS.md):

1. Closed-loop repair is the #1 lever (~2×). Feeding test/compiler failure back and retrying took resolve-rate 7.7% → 15.3% on the same cheap model. Iterate against ground truth, don't single-shot.
2. Cheap-first + cost-aware routing. Track $/resolve, not just resolve-rate; a cheap model resolved 31× cheaper per fix than a frontier one. Reserve frontier for measured capability gaps.
3. Tier the models (Barbarian & Scholar). Cheap sweep + frontier on only the residual = 33.3% at ~6× lower cost than running frontier everywhere.
4. Put the output-format contract in a system message + example, and size prompts to the model's real context window — this alone took a weak local model from 0% to ~50% valid output.
5. Only trust batch evaluation of the final artifact — in-loop counters drift 1.5–5×.
6. The harness multiplies the model; it can't rescue one below the task's reasoning floor. Pick the smallest model above the floor, then let evolution do the rest.