nobrainer-autoimprove

NoBrainer Autoimprove

Improve one artifact only when a frozen evaluation proves a candidate better. This independent adaptation applies Andrej Karpathy's autoresearch measure-change-keep loop to skills, prompts, instructions and other reviewable artifacts; it is not an official Karpathy project.

Route the smallest fix

Fix an obvious typo, dead link or deterministic defect directly and cover it with the normal test. Do not start an experiment when the objective cannot be measured, target/evaluator isolation is impossible, or the change needs an unapproved production, credential, publishing, safety or irreversible action.

Treat owner correction as evidence, not permission for a global rewrite:

• task-local clarification stays in the task;
• a repeatable behavior gap becomes a regression scenario;
• a verified durable fact or preference follows LEARNING_WRITE_POLICY: AUTO_SCOPED writes one minimal sourced rule to one canonical store, ASK prepares one exact diff, and OFF keeps it task-local: do not create a durable diff in AGENTS.md, tasks/lessons.md or a wiki.

On AGENT_ERROR_CORRECTED, preserve the failing scenario and error fingerprint, fix the active result first, then test the smallest reusable prevention. One anecdote never justifies a broad user trait or unrelated write.

Frame before mutation

Freeze this contract before changing the target:

TARGET: <one artifact and allowed write scope>
OBJECTIVE: <one observable improvement>
BASELINE_VERSION: <commit/hash/copy>
DEVELOPMENT_EVAL: <representative iteration cases>
HOLDOUT_EVAL: <unseen promotion cases>
HARD_GATES: <format, safety and deterministic checks>
SCORE: <criteria, aggregation, direction and meaningful delta>
EVAL_INTEGRITY: <separate owners, access and calibration controls>
SCORE_RECEIPT: <append-only trial identity, provenance, raw evidence and decision>
NOISE_POLICY: <paired repetitions, dispersion and retry rule>
BUDGET: <finite variants, rounds, repetitions, retries, time and resources>
PLATEAU: <minimum delta and dry rounds>
ROLLBACK: <retained baseline, restore and readback>
OWNER_GATES: <action, owner and recorded evidence>