Model cost provenance

Contextual rules can cost money. MCMR keeps their usage per model turn and per verdict instead of averaging it across a run.

Each contextual result can record the backend, model, reasoning effort, input tokens, cached input tokens, and output tokens. Passing and failing verdicts both keep usage because both required model work.

A batched response may answer many criteria. Its turn is counted once, then linked to the answers it produced. Cached input remains separate from fresh input so prompt reuse stays visible.

Rollups

Rule jobs keep totals per codebase, usage for the latest run, and usage across every codebase that runs the rule. A deterministic rule publishes no cost keys. Missing cost data therefore means the verdict was computed locally, not that a paid call cost zero.

Measure a candidate model before changing the project default.

mcmr model-sweep . --backend openrouter --model deepseek/deepseek-v4-flash-0731
mcmr contextual-experiment labels.json

model-sweep runs the live contextual catalog without editing project configuration. contextual-experiment compares backends against a reviewed label corpus.

See Rules and lanes for the backend contract.