The Evidence Behind an AI Review Verdict

The AI review verdict is determined through a deterministic process, without the involvement of a model. The verdict relies on evidence collected during the checks, and it follows fixed rules. If there is drift or failing evidence, the verdict fails. If there is no verified evidence, the verdict is inconclusive because an empty result is not a pass.

If some of the required checks could be verified, the verdict is partial. Only when the evidence is fully verified does the verdict pass. The model does not participate in this decision, ensuring stability and explainability in the result. Every evidence item carries a reproducible command and the raw process output, allowing for auditing and verification of the verdict.

The four evidence classes are Static Typecheck, Dynamic Test, Contract Schema, and AI hypothesis. The verdict is computed based on the class of each finding, the command that produced the result, and the raw process output. A reproducible command and raw output are crucial for trust, as they enable the comparison of results to ensure consistency.

The verdict reflects the actual evidence present, such as missing or partial checks, and does not round up to a pass when mixed results are present. The AI model can provide hypotheses, but it cannot make a check pass on its own. Only fully verified evidence can contribute to a pass. The system keeps model-only notes as hypotheses and treats empty results as inconclusive, preventing the system from interpreting silence as success.