Audit exact failure mechanisms before retraining

Audit exact failure mechanisms before authorizing retraining; tool-family or category counts can conceal both already-supervised failures and genuinely absent negative constraints. When a failure persists despite direct examples, “missing data” is not established: first replay the frozen decision through real application boundaries to distinguish model behavior from containment or state effects. When the exact mechanism is absent, do not recycle an opened development example—freeze a new call/input-disjoint training supplement and review it independently.

Method: 1. Define each observed failure as a functional predicate, including relevant history and negative constraints. 2. Count only canonical training rows whose input and agreed target exercise that exact predicate; report thin, absent, and conservatively excluded near-matches separately. 3. If exact examples already exist, treat data absence as unsupported and test application-path containment before changing training. 4. If coverage is zero, collect a new blind training-only micro-supplement before any SFT comparison; hold recipe and evaluation fixed. 5. Keep policy/label ambiguities out of automatic training targets.

Limits: Exact coverage does not prove that supervision is sufficient, representative, or learnable. A few examples may still be too thin, and application replay cannot by itself identify the optimal model intervention.

Evidence/example (2026-09-11, bounded offline audit): three residual mechanisms had exact training coverage counts of 3, 0, and 3. One of the three-example groups contained a near-exact form of the observed failure, while the absent group encoded a negative constraint that broad category counts missed. A candidate prompt/harness change produced one gain on a small frozen slice without regressions, so the justified next action was bounded full-path replay, not immediate retraining. Development cases remained diagnostic-only; no test data, training run, or production state was touched.