Definition
The fraction of a reference (cloud/full-precision) model's prognostic capability preserved after compression and deployment to an edge runtime, combining prognostic-horizon retention and probability-distribution agreement.
The Measurement Gap
Model validation typically happens once, in the cloud, on the full-precision model, but production inference runs on quantized, pruned, or distilled versions deployed to edge gateways. F1, precision, recall, and AUC are computed against the reference model and silently inherited as if they applied to the deployed model. The failure mode is silent prognostic-horizon shrinkage: INT8 quantization may preserve 97%+ classification accuracy while reducing prognostic lead time enough to consume the organization's response buffer. Legacy operational metrics (OEE, MTBF, PM Compliance) cannot see the difference between reference and deployed model performance. A 3% accuracy degradation can translate into days of lost lead time: operational value is far more sensitive to compression artifacts than benchmark accuracy implies.
Decision Supported:
Is the edge-deployed model operationally equivalent to the validated reference model, or has compression silently degraded performance in ways that affect prognostic horizon and action timing?
Formula
Plain language: Run shadow-mode scoring with both the edge and reference models on the same input data. Compute the ratio of edge to reference prognostic horizon, then multiply by one minus the average Jensen-Shannon divergence between the two probability distributions. The result combines how much lead time the edge model preserved with how closely its outputs agree with the reference.
Notation:
EDF = (PH_edge / PH_ref) × (1 − mean JS(P_edge,i ‖ P_ref,i))
Variables:
| Variable | Definition | Unit |
|---|---|---|
| PH_edge | Average prognostic horizon from the edge-deployed model | Days |
| PH_ref | Average prognostic horizon from the full-precision reference model | Days |
| JS(P_edge,i ‖ P_ref,i) | Jensen-Shannon divergence between edge and reference probability outputs for scoring window i | Ratio (0–1) |
| mean JS | Mean JS divergence across all evaluation windows | Ratio (0–1) |
| EDF | Composite fidelity score | Ratio (0–1) |
Confidence Grading (required alongside every EDF 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 EDF 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.