Objective
Compare ranked repair sequences by restoration impact while preserving modeled feasibility and human authority.
Validation standard
Measure the decision, not just the outage.
Dovrane evaluates whether a repair sequence is feasible, explainable, and capable of improving restoration outcomes using only the information available at decision time.
EVENT-TIME REPLAY · BASELINE COMPARISON · OUTCOME TRACEDecision-quality framework
Evidence that can survive scrutiny.
The standard connects each recommendation to measurable outcomes, operational constraints, and the exact evidence available when the decision was made.
01Outcome impactRestoration curve, customer-minutes interrupted, and critical-load recovery.
02Modeled feasibilityTopology, crew, material, access, safety-condition, and critical-load constraints.
03Decision qualityAlternatives, uncertainty, abstention, rationale, and operator disposition.
04Operational learningFrozen-time replay links recommendations to outcomes without hindsight.
Protocol
A fair comparison starts with identical information.
Every comparator must operate on the same frozen event-time state. Hindsight is reported separately, never blended into a product result.
Primary metric
Estimated customer-minutes interrupted (CMI): sum of modeled interrupted customers across modeled time intervals.
Observed CMI
Customer-minutes derived from approved historical outage and restoration records; not interchangeable with estimated CMI.
Fairness rule
Dovrane and baseline receive identical information available at the original event time.
Frozen replay
Later evidence is excluded until its historical arrival timestamp.
Comparators
Utility-approved baseline, Dovrane engine, and analysis-only oracle with future information clearly separated.
Evaluation sets
Development, regression, and sealed evaluation sets reveal progress without contaminating final comparisons.
Reporting
Full distribution, median, spread, adverse tail, violation count, calibration, and abstention.
Modeled-feasibility invariants
Every recommendation earns its place in the ranked set.
Independent checks screen candidate sequences against the constraints that shape restoration work.
Topology and isolation state
Tracked across every evaluation set to support review while preserving utility authority.
Crew class and availability
Tracked across every evaluation set to support review while preserving utility authority.
Required material availability
Tracked across every evaluation set to support review while preserving utility authority.
Access and road status
Tracked across every evaluation set to support review while preserving utility authority.
Safety-condition prerequisites
Tracked across every evaluation set to support review while preserving utility authority.
Critical-load handling
Tracked across every evaluation set to support review while preserving utility authority.
Evaluation outcomes
From recommendation quality to operational learning.
- Sequence qualityCompare restoration impact against utility-approved baselines.
- Feasibility disciplineMeasure constraint violations, withheld candidates, and abstention.
- ExplainabilityTrace evidence, alternatives, uncertainty, and ranking rationale.
- Operational fitStudy operator acceptance, modification, deferral, and rejection.
- Resilience prioritiesExamine critical-load recovery and adverse-tail outcomes.
- Continuous learningConnect observed outcomes to future evaluation and governance.