The Repair Loop in Operation

This practitioner field report records an expectation-driven development gate operating on one proprietary reference implementation. A trusted, fixed, versioned gold state supplies both scenario inputs and expected persisted results. The Prove lane drives a coverage-selected set through the real interface and captures current interface evidence and command shapes; the Validate lane replays the full in-scope corpus through the official command path and compares the resulting stored state with gold. Four records separate the claims. Record I establishes a clean interface-evidence baseline. Record II reports 19 first-attempt replay passes, six marker-backed capability skips, no failures or retries, and roughly three-quarters of a million cumulative row comparisons against the expected gold state. Record III reports six planted upper-layer faults exposed and repaired across five stops, including one sweep-found fault and two faults stacked in one scenario. Record IV isolates the write-path claim: seven planted create-path mapping faults passed every interface-level test, while persisted-state parity convicted all seven. The report claims no general detection rate. The faults are deliberately selected, the author reports his own instrument, and the proprietary implementation and run artifacts are not published. The result demonstrates what the comparison uniquely detected within the declared scope, not what the gate detects universally. Security, performance, concurrency, read-path rendering, external side effects, oracle errors, and behavior outside the versioned coverage model remain separate concerns.