Definition
The proportion of an asset's credible high-consequence failure modes that are represented in the AI model's training data at a level sufficient for detection, graded by evidence strength per mode.
The Measurement Gap
There is no metric that tells you which high-consequence failure modes the model has enough training signal to detect, which means detection confidence is highest exactly where failure history is richest and consequence is often lowest.
Formula
Plain language: Start from the asset's engineered failure mode list (FMEA, HAZOP, RBI). For each high-consequence mode, grade the model's evidence: how many labeled examples of that mode (or validated physics-based synthetic equivalents) exist in training. A mode is covered if evidence meets a defined floor. RECS is covered high-consequence modes over total high-consequence modes, with partial credit for synthetic-only coverage.
Notation:
RECS = (M_covered + 0.5 × M_synthetic_only) / M_high_consequence × 100
Variables:
- M_covered = high-consequence failure modes with ≥ N_floor labeled real-world training examples (count, DERIVED; proposed N_floor = 10)
- M_synthetic_only = modes covered only by validated physics-based synthetic data (count, DERIVED)
- M_high_consequence = failure modes rated high-consequence in the asset's FMEA/HAZOP/RBI (count, MANUAL from reliability study)
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
For the failure modes that would actually hurt us (the ones in the HAZOP and the RBI study) has the model ever seen enough examples to detect them, or is it confidently monitoring for the failures that don't matter?