Definition
Degree to which plant process data transferred across cybersecurity boundaries (via approved data diodes and transfer mechanisms) accurately represents the real-time plant state, measured as the discrepancy between source instrumentation readings and the data available to AI systems after transfer.
The Measurement Gap
Nuclear plants operate under 10 CFR 73.54 cybersecurity requirements that mandate isolation between the plant's safety and security networks and any network accessible to AI systems. Plant process data reaches AI systems through approved data diodes and transfer mechanisms that introduce latency, sampling, and potential fidelity loss. No legacy metric measures whether the data arriving at the AI system accurately represents the real-time plant state. Instrument calibration programs verify that plant instruments read correctly. Data historian systems verify that data is stored accurately. Neither measures what happens to the data as it crosses the cybersecurity boundary. The invisible failure mode is that the data transfer introduces variable latency, sampling artifacts, or value corruption that causes AI models to misinterpret plant state, and no metric distinguishes between an AI model producing poor results because the model is wrong versus the model receiving degraded data.
Decision Supported
Can AI systems trust that the data they receive through the cybersecurity boundary is a faithful representation of plant conditions, or does the transfer introduce latency, sampling artifacts, or fidelity loss that could cause AI models to misinterpret plant state?
Formula
Plain language: For a set of representative plant parameters, simultaneously capture the value at the source instrument (inside the cybersecurity boundary) and the value available to the AI system (outside the cybersecurity boundary, after transfer). Compute the discrepancy. Average across parameters, weighted by the parameter's importance to AI decision-making.
Notation:
CBDFI = 1 − (Σ (w_p × |V_source_p − V_transferred_p| / |V_source_p|)) / Σ w_p
Variables:
- V_source_p = value of parameter p at the source instrument (inside the boundary), at reference time t (units vary by parameter)
- V_transferred_p = value of parameter p as received by the AI system (outside the boundary), at reference time t + latency (same units)
- w_p = weight reflecting parameter p's importance to AI decision-making (based on number of AI models consuming this parameter)
- Latency must be accounted for: compare the source value at the time it was sampled for transfer, not at the time the AI receives it
Sign convention: CBDFI = 1.0 means perfect fidelity (no discrepancy). Lower is worse. Values below 0.95 indicate significant data degradation.
Aggregation period: Continuous monitoring with weekly summary reports.
Edge Cases
- Latency is intentional and known: if the data diode introduces a fixed 30-second delay, the AI system should be aware of this and the delay itself isn't a fidelity issue. The concern is variable, unpredictable, or increasing latency, and data corruption. Measure discrepancy after aligning for the known fixed latency.
- Parameters with step changes (valve open/close): use state-change latency rather than MAPE. How long does it take for the transferred value to reflect the state change?
- Data transfer downtimes: track separately as "transfer availability". If the data diode drops out for 10 minutes, the AI operates on stale data. CBDFI measures fidelity when the transfer is working; availability measures how often it's working.
Confidence Grading (required alongside every CBDFI value):
| Grade | Basis | Interpretation |
|---|---|---|
| A: Empirical | 5+ observed events in measurement window provide direct tail observation | Measurement grounded in observed outcomes; highest confidence |
| B: Hybrid | 1-4 observed events plus Monte Carlo simulation with Bayesian updating | Blends observation and simulation; moderate confidence |
| C: Simulated | Zero observed events; derived entirely from Monte Carlo simulation | Modeled, not observed; treat as indicative, not validated |
A CBDFI of any value at Grade C is a simulation that agrees with itself. It tells you the model is internally consistent but not that it matches reality. Grade C scores should trigger proactive stress-testing against historical near-miss events and cross-asset analogues rather than acceptance as calibrated.
Proposed formula, pending validation pilots.