Definition
Independence of the QA batch release decision from the AI's displayed prediction, computed as one minus the correlation between AI prediction and release decision on borderline batches. A low score means the reviewer is ratifying the dashboard rather than reviewing the record.
The Measurement Gap
Before a QA reviewer opens a batch record, the predictive dashboard shows "96.4% probability of conformance." She reviews the record, confirms specifications are met, and releases the batch. The right decision. But the question nobody asks: would she have released it if the dashboard showed 72%? Would she have spent longer reviewing? Would she have caught a marginally trending quality attribute?
21 CFR 211.22 requires independent QA review. If the reviewer's judgment is systematically influenced by the AI prediction seen before review, the review is procedurally independent but cognitively compromised. Legacy metrics cannot detect this: batch release rate tells you how many batches were released, not whether the decision was independent.
Decision enabled:
When BRIS drops below 0.3, the organization should implement a "predict-then-reveal" workflow where QA documents their initial assessment before seeing the AI prediction. When combined with low QPCI (miscalibrated model), low BRIS represents the highest-risk scenario in the predictive manufacturing cluster.
Formula
BRIS = 1 − r(AI_prediction, QA_decision | borderline_batches)
Where r is the point-biserial correlation coefficient between the AI's pass/fail prediction and the QA release/reject decision, calculated only for borderline batches (quality attribute within 10% of specification limit or AI confidence between 70–90%).
Worked Example:
Over two quarters, a solid dosage facility produces 78 borderline batches.
AI predicted pass on 52: QA released 51 (98.1% agreement).
AI predicted fail/flagged concern on 26: QA held 24 (92.3% agreement).
Overall agreement rate = (51 + 24) / 78 = 96.2%.
Point-biserial correlation = 0.89.
BRIS = 1 − 0.89 = 0.11
A BRIS of 0.11 is well below the 0.3 investigation threshold. Out of 78 borderline cases where independent judgment should be most visible, QA diverged from the AI exactly 3 times. This does not necessarily mean QA is wrong. If QPCI confirms the model is well-calibrated, high correlation may be appropriate. But if QPCI shows miscalibration, BRIS of 0.11 means QA is anchoring to bad information.
Sign Convention:
Higher is better. 1.0 means complete independence. Below 0.3 warrants investigation.
Proposed formula, pending validation pilots.