A Repair Cannot Borrow Its Test

Current state

A failed review can tell a factory what kind of obligation it missed. It cannot become permission to copy the withheld case into the next candidate or its regression test.

That distinction matters because a repair can look precise while quietly learning the answer to the test it is supposed to face fresh. The next review would then say less about the candidate than it appears to say.

Changes

One completed engineering cycle made the repair path keep the failed-review provenance while limiting the repair input to an already visible related case. The resulting source revision preserves the earlier baseline cases and adds a new regression that states the failed classification boundary without reconstructing the withheld request.

The repair remains a new development candidate. Its earlier failed review stays failed, and the new candidate needs a separately authorized review with a new withheld partition before it can claim behavioral improvement.

Failures observed

The cycle exposed a tempting shortcut: a repair path could receive a redacted review summary yet lose the fact that the failure came from a bounded review. That makes it easier to generate a broad correction with no clear connection to the failed obligation, or to rebuild a regression from information that was meant to stay withheld.

Assumptions removed

We cannot assume that a repair remains honest merely because it names a failed rule. We also cannot assume that a structurally improved candidate has repaired the behavior that a fresh review would test.

Useful repair evidence needs both links: back to the exact failed obligation, and forward to a new review that does not reuse the withheld input.

Evidence

The completed cycle bound a repair proposal to the failed review, preserved the baseline cases unchanged, derived one added regression from a previously visible related case, and carried the approved correction into an immutable development revision. It intentionally stopped before a new live review, qualification, promotion, publication, or deployment.

Limitations

This is one deterministic engineering observation. It does not show that the new revision will pass a fresh review, that public regressions cover every failure mode, that repair guidance generalizes across agent shapes or targets, or that a bounded repair improves real work.

Next hypothesis

Run a separately authorized fresh review for repairs that are bound to their failed obligation but derived only from visible regression material. Retain every attempt, preserve the withheld partition, and compare the new result with the prior verdict without allowing either one to overwrite the other.