# Customer run transparency

MCP completed execution results now include a versioned `summary`: every attempted provider, model, effort, tier, outcome, validation status, duration and available provider-reported token counters. Escalation counts and per-provider totals include failed attempts. Each aggregate reports whether usage coverage is complete, partial or unavailable; missing values are never converted to zero. Provider counters remain separate because cache semantics differ.

Actual billed cost and subscription quota remain unknown. Completed MCP runs now request a bounded server-side API-equivalent estimate and attach it as `summary.cost.serverEstimate` when sufficient usage and configured prices are available. Missing or incompatible evidence produces an explicit unavailable or partial result. Named-baseline savings selection is not yet exposed in the UI; this is not a finished savings dashboard. A server-side estimator must use versioned prices, explicit usage semantics, context/cache pricing conditions and a named baseline. Counterfactual costs must be labelled estimates, account for every retry, and must not imply equivalent quality or measured subscription savings. A measured benchmark against a baseline requires comparable tasks and validation outcomes.

Verification after the computer restart: four focused summary tests passed, covering partial failed-attempt usage, cache counter separation, invalid/overflow counters and no-attempt results. Portable relocated MCP verification and deterministic stress generator tests also passed. No model inference was invoked by those checks.