Failures belong in the scientific record

A research system that keeps only its best run destroys information. Without the failures, a reader cannot tell a new attempt from a repetition, and the record quietly implies that the successful path was obvious.

HowHow retains failures, rejected candidates, unresolved contradictions and inconclusive outcomes. Each failure record preserves the symptom, attempted action, diagnosis, evidence, resource cost, lesson, avoid-scope and whether a later result superseded it.

Retries are automatic only for idempotent, retryable failures. Unknown side effects, ambiguous dispatch, security failures and budget exhaustion block instead.

Inconclusive is a valid research outcome rather than a softer word for success. That distinction is load-bearing: without it, the pressure to round up is applied at exactly the moment the record matters most.