Definition
Statistical distance between the distribution of AI-influenced operational decisions (relative to their applicable limits) and the historical conservative decision baseline, measuring whether the plant's decision culture is systematically drifting from "well within limits" to "near the boundary."
The Measurement Gap
AI optimization in nuclear plants naturally pushes operating decisions toward technically compliant but less conservative positions. Each individual decision may be defensible, but the pattern of decisions shifts the plant's operating character from "well within limits" to "near the boundary." No legacy metric detects this drift. Capacity factor measures generation output and is blind to margin consumption. Scram rate measures unplanned shutdowns, not decision patterns. Maintenance Rule (a)(2) criteria track system performance against goals but not how close the plant operates to those goals. WANO indicators measure broad performance outcomes, not the distribution of individual decisions relative to their applicable limits. The invisible failure mode is that the AI optimization regime changes the statistical distribution of decisions (LCO durations, surveillance test margins, shutdown margins, maintenance deferral durations) from "comfortably conservative" to "technically compliant near the boundary," and no single metric captures the collective shift because each individual decision is within limits.
Decision Supported
Is the AI optimization regime changing the plant's operating character from "comfortably conservative" to "technically compliant but operating near limits," a shift that no individual decision would reveal but the pattern makes visible?
Formula
Plain language: For each category of operational decision that has a regulatory or Technical Specification limit (e.g., LCO entry duration relative to allowed completion time, surveillance test results relative to acceptance criteria, shutdown margin relative to requirements), measure where the actual decision or outcome fell on the spectrum from conservative center to the limit boundary. Compute the median position across all decisions. Compare current period to baseline.
Notation:
CDDI = (M_current − M_baseline) / (1 − M_baseline)
Where:
M_period = median of {(Actual_i − Conservative_midpoint_i) / (Limit_i − Conservative_midpoint_i)} across all sampled decisions i in the period
Variables:
- Actual_i = actual decision point or measured value for decision i (units vary by decision type)
- Conservative_midpoint_i = the midpoint between the expected nominal value and the limit for decision type i: represents the conservative operating target
- Limit_i = the Technical Specification or regulatory limit for decision type i
- M_current = median margin consumption ratio for the current measurement period (dimensionless, 0 to 1)
- M_baseline = median margin consumption ratio for the baseline period (pre-AI or first year of AI operation) (dimensionless, 0 to 1)
Sign convention: CDDI = 0 means no drift from baseline. Positive values indicate drift toward limits (bad). CDDI = 1.0 would mean the median decision has moved all the way from baseline to the limit: catastrophic.
Aggregation period: Quarterly, with trend over 8+ quarters to detect slow drift.
Edge Cases
- Different decision types have different limit structures (time-based, value-based, count-based): normalize all to the 0-to-1 margin consumption scale before computing the median
- Baseline period may have been excessively conservative (waste): CDDI measures drift from whatever the baseline was, not from an optimal position. If the baseline was wastefully conservative, some positive CDDI may be appropriate. The concern is monotonic drift, not the absolute position.
- Single decision types with very few events (e.g., reactor trips): exclude from the median calculation and track as separate high-consequence indicators
- Minimum sample: 200 decisions per quarter across all types for the median to be statistically stable
Proposed formula, pending validation pilots.