Proofs, Not Promises: Governed Candidate Improvement for Agent Control Runtimes

Agents can appear to improve after failure while the path of improvement disappears. A prompt repair, memory edit, configuration change, or runtime update may make the next answer look safer without preserving what failed, what change was proposed, who challenged it, where it was applied, or what became worse. This paper studies that missing object: the governed candidate lifecycle. Civitas 6.7B treats attempted adaptation as an auditable artifact rather than as an unrecorded correction to an answer. Repeated governed-agent failures become bounded candidate artifacts; candidates are bound to Cassius challenge evidence, policy and trust-region gate evidence, and sandbox-only application state; and archived candidates are evaluated on sealed Australian finance held-out and stress tasks before a result claim is made. The evidence is a locked Australian finance paper bundle from a local Ollama/Gemma lane, not a smoke or stub run. It contains 250 curated AU task rows: 100 training failure rows, 100 held-out rows, and 50 stress rows. The bundle reports zero production mutations, zero unauthorized promotions, eight sandbox-only promotions, eight Cassius-required and Cassius-passed governed approvals, zero Cassius-unavailable claim-supporting approvals, and receipt completeness of 1.000000. Its clearest behavioral result is a held-out unsupported-claim rate reduction from 0.110000 to 0.057500. Held-out policy-violation, contradiction-to-policy, and unsafe-action rates also move downward in narrow or directional ways. The same archive preserves unfavorable outcomes: correct-controlaction rate falls, held-out regression rate rises, governed stress regression exceeds comparator stress rates, eight-of-eight approval leaves gate selectivity unresolved, and replay reproducibility is not available. The manuscript therefore makes a systems claim, not a leaderboard claim. Civitas 6.7B does not prove general safe self-improvement. It demonstrates that attempted agent adaptation can be made inspectable as governed candidate improvement: auditable, Cassiuschallenged, sandbox-contained, and empirically falsifiable under a bounded AU evidence contract. The result is proof of a chain of custody for improvement attempts, not a promise of broad behavioral superiority, production promotion, live production self-modification, API portability, EU results, statistical significance, replay reproducibility, or proven gate selectivity.