Once the output became the object of inspection, failure had to be named before it could be repaired.

The failure probe organised output defects into structural classes rather than treating all incorrect outputs as one category.

The resulting taxonomy distinguished complete loss, partial damage and identity-level substitution patterns, with null controls retained to test false activation.

This changed the repair problem from 'rewrite the translation' to 'identify the failure class and determine whether a bounded local intervention is admissible'.

Module 4.0 tested class-aware span-local reconstruction. The mechanism first detected whether a protected anchor was structurally damaged, then attempted only the repair permitted for that failure class.

Across the fixed probe sets, detection correctness was 1.000 with zero false positives. Repair success differed sharply by class: the easiest loss class was recoverable, partial-damage cases were only partly recoverable, and the dominant identity-substitution class resisted neural repair.

The dominant failure mode was not merely missing surface form. A fluent or plausible alternative could occupy the correct position while remaining the wrong canonical identity.

Form recovery does not imply identity recovery.

This is the point where unrestricted neural repair becomes conceptually insufficient. A model can propose a candidate string, but it cannot establish that the candidate is the normatively admissible form for the document family.

The bounded repair stage therefore routes cases rather than silently rewriting everything.

The output-failure work exposed a new question: if repair cannot define canonical identity, where should canonical authority live?

The answer was external deterministic control.

From Repair to External Authority