Corrective learning evals

Pass criteria

• At least one recurring exception class produced a recorded change to an upstream object
• The change is a state transition on a policy, category, workflow, or authority object, not a case note
• The trace can be reconstructed from records the operator retains

Fail indicators

• Every instance was closed as resolved and no upstream object changed
• The pattern is visible only in anecdotes; no aggregation exists
• Handling changed the individual outcome and left the generating rule intact

Pass criteria

• The operator can compute an absorption share from its own records
• Corrective effort is attributed to the object that produced it
• Exception classes with high absorption share route to review

Fail indicators

• No way to tell which corrective effort fed any upstream object
• Corrective labor is invisible in the operator's own records
• High-absorption classes keep recurring without review

Pass criteria

• A named trigger connects challenge volume or exception classes to a policy review
• The trigger has fired within the review period and produced a recorded decision
• A reasoned refusal is a valid outcome; silence is not

Fail indicators

• Volume is reported to governance but no rule routes it to review
• The trigger exists on paper and has never fired
• Reviews happen on a calendar unrelated to what the challenges said

Pass criteria

• A named instance exists and is verifiable in records
• The change is to a rule, authority, or burden allocation, not only to a model or dashboard
• The instance is recent enough that learning is a live capacity, not a founding story

Fail indicators

• Learning claims rest on model metrics: retraining, accuracy, optimization
• The named instance predates the current delegation
• No instance can be named at all

Pass criteria

• A rescue register exists and covers the review period
• Each entry names the near-miss condition, the person who caught it, the labor spent, and whether the source condition was fixed
• Rescue counts and incident counts are reviewed as one series, never reported alone
• A condition that produced more than one rescue appears as a standing item with an owner who can change it
• A rescue record cannot be closed without either a fix or a recorded decision not to fix

Fail indicators

• Only incidents are counted, and a low incident rate is presented as a safety result
• Rescues are recorded only when they failed, so a successful catch leaves no trace
• Repeat rescues against one unfixed condition are cited as evidence of reliable staff
• Falling incidents with rising rescues is read as improvement rather than as unchanged risk carried by people
• No labor figure accompanies a rescue, so the cost of catching lands on the person who caught it

Pass criteria

• A workaround inventory exists with frequency per class
• Each recurring class has an investigation outcome: rebutted with evidence, or fixed upstream
• Workaround data reaches the people who own the object the workaround compensates for

Fail indicators

• Workarounds are known informally and inventoried nowhere
• Adaptation is cited as evidence the system scales
• The same workaround has recurred across review periods without investigation

Pass criteria

• Action capacity and corrective capacity are stated on separate axes
• The gap is reported with an owner, not only surfaced after incidents
• Every scope expansion rechecks corrective capacity before it proceeds

Fail indicators

• Only action metrics are tracked; correction staffing is folklore
• The appeals function has been flat while decision volume compounded
• Expansion decisions cite no corrective-capacity figure