# Evaluation contract

## Decision

- Candidate and baseline versions:
- Ship/hold/rollback decision this evaluation informs:
- Owner and deadline:

## Population and task

- Intended traffic and exclusion rules:
- Dataset source, snapshot, unit of analysis, and leakage controls:
- Environment and dependency versions:

## Metrics and graders

- Primary outcome metric and success predicate:
- Guardrail, slice, process, cost, and latency metrics:
- Grader implementations and calibration evidence:

## Experimental design

- Trial count, seeds, pairing/blocking, and clustering:
- Minimum effect of interest and power/precision target:
- Missing-data and retry policy:

## Decision rule

- Required improvements:
- Non-inferiority margins:
- Veto conditions and worst-slice limits:
- Who may waive a condition, with what evidence:

