self-improvement-loops

Use a three-stage loop: mine failed traces into clustered weakness signatures, generate bounded proposals touching one editable surface each, and validate with a two-split acceptance gate. Keep the evaluator, permissions, and budgets outside any surface 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 candidates that trade one split against the other even when the sum improves.

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 at the runtime or OS level, and count any detected exploit as a failed candidate rather than a high score.

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 rather than only score drops.

Avoid it when no fast, deterministic, automatable evaluator exists, or when the base model fails a capability validation on the task family, since recursion amplifies weak models downward. Single-loop governance, one-shot context optimization, and judge design belong to adjacent skills instead.

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