Quality standards for AI evals and software tests
Three proposed standards. Two say what a suite must contain; one says how much human review is enough to trust it. CARBON productizes them: scan a folder, see what is covered, what is missing and what to add.
AI Eval Coverage Confidence
CARBON Eval² · the mix of eval questions: supported, false-premise, boundary, insufficient info, grounding, FP/FN challenges.
Software Test Coverage Confidence
CARBON Software Eval · the mix of test cases: positive, negative, boundary, errors, state, security, accessibility, performance.
Review Confidence
CARBON Human Review · sampling, intervals and evidence sufficiency for reviewing questions, tests and their results.
These are proposals for industry discussion, not certifications. Percent targets are configurable starting points to calibrate per application. Machine-readable profiles: coverage-targets.json.