Methodology
Inference energy analytics
Neutral, request-level energy attribution foundations for internal analytics and future reporting mappings.
Canonical unit: Joules
Event unit
J
Each inference event is stored in Joules before any derived KPI.
Derived views
J/token
Dashboards derive J/token and Wh/1k tokens from neutral event data.
Quality overlay
Confidence-aware
Method tier, uncertainty, and exclusions stay attached to every event.
What changed
Request-level attribution
The agent samples power during live inference windows and emits one auditable event per request.
Neutral metrics
The platform stores Joules, tokens, measurement method, confidence, and exclusions without embedding compliance claims.
Future compatibility
Additional reporting mappings remain possible later by transforming retained evidence instead of rewriting the raw ledger.
The previous public benchmark index is retired. Operator analytics now use auditable request-level Joule events with quality and confidence metadata.