self-improvement-loops

Use a three-stage loop: mine failed traces into clustered weakness signatures, generate bounded proposals touching one editable surface each, then validate with a two-split acceptance gate before merging. Keep the evaluator and permissions outside anything the loop can modify.

Accept a candidate only when it does not regress on either a held-in split or a held-out split the proposer never sees, and strictly improves on at least one, under repeated evaluation. Reject trade-offs between splits even when the net score is positive.

Hide the scoring implementation from the proposer and expose only scores and traces, since visible evaluators get monkey-patched or overwritten. Enforce budgets and sandboxing in the runtime rather than in prompt text, and count detected exploits as failed candidates.

Common causes are diversity collapse from hill-climbing only the latest candidate, and degenerate improvers that silently return their input unchanged. Keep an archive of all viable candidates, discount parent selection by offspring count, and alarm on empty diffs, not just score drops.

Avoid it when evaluation is slow, ambiguous, or judge-only, when no held-out split can be protected from the proposer, or when the base model fails a capability validation run. Weaker models can degrade mean performance across iterations instead of compounding gains.

Context collapse occurs when a model monolithically rewrites its accumulated playbook, shrinking it drastically and dropping quality below the no-adaptation baseline. The fix is itemized entries with stable identifiers, incremental deltas, and deterministic non-model merge logic.