Definition
No existing KPI measures the unintended quality drift on dimensions not included in the AI's optimization objective.
The Measurement Gap
The decision MDQDI supports is whether autonomous process control has been given a broad enough optimization objective. When an AI controls temperature and pressure to optimize tensile strength, what happens to surface finish, dimensional accuracy, or porosity rate? Those uncontrolled dimensions can drift without any alarm firing because the AI is performing correctly on its objective. The measurement gap exists because SPC systems monitor individual parameters against their own control limits, but no metric tracks whether the collective drift across uncontrolled dimensions is correlated with the AI's control actions. The invisible failure mode is optimization scope blindness. The AI optimizes what it is told to optimize. Everything else becomes a dependent variable that nobody is watching systematically. Quality on uncontrolled dimensions may degrade gradually over months, masked by the improvement on controlled dimensions that dominates the aggregate quality scorecard. The drift becomes visible only when a customer complaint or field failure traces back to a dimension nobody was monitoring relative to the AI's actions.
Formula
For each quality dimension not in the AI's optimization objective, compare the current statistical distribution to the pre-AI baseline distribution. The formula is: MDQDI = (1/D) times Sum_d of [absolute value of (mu_d_current minus mu_d_baseline) / sigma_d_baseline plus absolute value of (sigma_d_current minus sigma_d_baseline) / sigma_d_baseline].
Variables: D is the count of quality dimensions not in the AI's optimization objective. mu_d_current is the mean of quality dimension d in the current measurement period. mu_d_baseline is the mean of dimension d in the pre-AI baseline period. sigma_d_current is the standard deviation of dimension d in the current period. sigma_d_baseline is the standard deviation of dimension d in the baseline period.
Sign convention: 0 equals no drift on any uncontrolled dimension (ideal). Rising values indicate drift with no upper bound. The metric captures both mean shift and variance change.
Edge cases: if baseline data does not exist for a quality dimension, it cannot be included in MDQDI, so baselines must be established before enabling autonomous process control. Quality dimensions with inherently high variability should be weighted lower or excluded. If the AI's objective is updated to include a previously uncontrolled dimension, remove that dimension from MDQDI and re-baseline. Minimum sample: 100 measurements per dimension per period.
Proposed formula, pending validation pilots.