CAIN-42 governed self-evolution (2026-10-07)

YOUR AI AGENT CAN IMPROVE WITHOUT BEING ALLOWED TO SILENTLY EXPAND ITS AUTHORITY. LEARNING DOES NOT GRANT AUTHORITY.

What it does

An agent can propose a better skill. CAIN-42 runs the candidate in a fresh, confined process (no network), reproduces its claimed gain on known, hidden and unrelated regression tasks, puts it through the existing evolution gate and an operator approval, watches it on a canary, and then promotes it or rolls it back. Every outcome gets a governance-signed evolution receipt. Delegated authority only narrows: a sub-agent can never ask for more than its root principal delegated.

What was measured

• Task success on 24 held-out tasks: static baseline 0.3333, ungoverned self-evolving agent 0.625 with 36 security violations, CAIN-governed evolving agent 1.0 with 0.
• Of 8 proposed skills (overfit, regressing, nondeterministic, authority-seeking, data-exfiltrating, latent defect and two genuine), only genuine improvements went live; reproducibility 1.0; the canary rolled back the latent defect.
• 16 of 16 evolution receipts verify under a clean-room verifier with no CAIN imports.
• Continuous red team: 411 attacks, 411 blocked, 0 breaches. Its first run found 4 real defects in existing gates; all fixed at source.
• Action decisions against an independent policy oracle: 0 false allows, median 1.218 ms.

Tests: 127 of 127 pass. Reproduce: python3 scripts/governed_evolution/run_evidence.py --seed 5150 --red-team-seed 4242.

Files: RESULTS.json · BENCHMARK_RESULTS.json · RED_TEAM_SCORECARD.json · RECEIPT_VERIFICATION.json · junit.xml · manifest.json. Source code is not published.

Limits. Synthetic, deterministic workload: the proposals are a fixed labelled pool, not a language model, so token cost is not measured. Small n (8 proposals, 24 tasks, 40 actions). A library with tests and reproducible evidence; not yet wired into the hosted gateway. Same-project, not third-party. PRE-PRODUCTION.