A test plan may adopt a sequential or adaptive design such as group-sequential or SPRT to cut sample sizes while holding pre-specified type I and II error rates; performance claims are framed as non-inferiority or equivalence against the indicator threshold and judged by one-sided confidence intervals or Bayesian credible intervals with pre-declared priors; demographic reporting is stratified, hierarchical models may borrow strength but stratum-specific lower bounds are still reported against local minima; the plan registers in advance its endpoints and thresholds, its margins, its stopping rules, its analysis method, its priors, alpha or coverage, stratification and missing-data handling; risk-based sampling sets per-stratum floors and adds samples to higher-risk strata; and where external evidence or operational monitoring reduces new sampling the plan shows where the data came from, how its quality was controlled and how it was statistically combined. Default floors for reduced designs are Basic 60 total and 10 per key stratum, Effective 150 and 25, Highly Effective 400 and 50 across at least three skin tones and two genders, Strict 1,000 and 100 across at least three skin tones, two genders and two device types, with key strata reflecting declared contexts and known risk factors. Acceptance rules follow the examples given: accept a level when the one-sided lower bound meets the threshold, stop early only on a pre-set margin, flag any stratum below its local minimum and require mitigation or scope limitation, and allow reduced confirmatory samples at recertification where monitoring shows sustained performance under control limits. Vendor-supplied datasets are locked and split into design and confirmatory sets, peeking outside declared interim looks is prohibited, and datasets, seeds, code versions and environments are recorded for reproducibility.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.