Identity
Record provider, exact model version, alias, date, region, tools, and price.
Evidence first.
Preference second.
A good-looking artifact cannot outrank a broken one. A high score cannot escape its task, persona, conditions, or date.
Synthetic evaluation using demonstration outputs. Not real customer demand, adoption, or revenue.
Locked task
Freeze brief, conditions, success criteria, retries, and publication rules before execution.
Repeated runs
Execute comparable runs and preserve outputs, errors, latency, cost, and environment.
Objective gates
Correctness, schema validity, coverage, accessibility, and citation validity can disqualify outputs.
Persona judgment
Persistent disclosed personas evaluate fit without pretending to be customers, voters, or demand.
Confidence
Coverage, run variance, evaluator agreement, and freshness determine the confidence label.
What this does not prove
Synthetic evaluations do not establish real customer demand, market adoption, revenue, retention, or a statistically representative human preference. They are repeatable decision support attached to inspectable evidence.