AEP Metrics (Informative)¶
These definitions are informative. They must not be converted into UHQS
module scores or letter grades. Status is valid | inconclusive | control_failed |
not_computed.
Value of Deception (VoD)¶
| Item | Rule |
|---|---|
| Inputs | Explicit utility.weights in the experiment manifest; observed defender outcomes / costs on trials |
| Units | Utility delta (study-defined) |
| Invalid uses | Substituting UHQS or delta_uhqs; omitting the utility model |
| Interpretation | Positive VoD means higher mean defender utility on the decoy arm under the declared model |
Worked synthetic example: weights {detection: 1.0, intelligence_yield: 0.5,
defender_infra_cost: -0.1} with mean utilities 1.4 (decoy) and 0.9 (reference)
⇒ VoD = 0.5.
Fingerprinting Susceptibility Vector (FSV)¶
Report per layer (network, protocol, system, state):
- Confusion counts (TP/FP/TN/FN)
- TPR, FPR, balanced accuracy
- Bootstrap intervals when sample size allows
| Item | Rule |
|---|---|
| Inputs | detector.layer, predicted_decoy, actual_is_decoy on decoy/reference trials |
| Invalid uses | Collapsing to one global FSV scalar without meeting all control requirements (AEP never emits a global scalar today) |
Dwell-Time Distortion Ratio (DTDR)¶
When any trial is right-censored, AEP uses a Kaplan–Meier median estimator and does not pretend an uncensored mean is valid.
| Item | Rule |
|---|---|
| Units | Ratio (dimensionless) |
| Invalid uses | Comparing against production assets; ignoring censoring |
| Tarpits | DTDR > 1 may be desirable even when Module E latency looks “worse” |
Worked synthetic example: decoy median 210 s, reference median 92.5 s ⇒ DTDR ≈ 2.27.
Exploit Exhaustion Rate (EER)¶
Mean unique controlled capabilities expended on the decoy per session, as a
fraction of the declared experiment budget (budget.max_unique_capabilities).
Category counts (tools, credentials, payload families, ATT&CK IDs) are preserved.
| Item | Rule |
|---|---|
| Invalid uses | Treating EER as universal attacker cost outside the declared budget |
Uncertainty¶
Bootstrap intervals use the analysis --seed for determinism. Low sample size,
high censoring, missing controls, and missing utility weights produce warnings
or inconclusive / not_computed statuses.