{
  "slug": "evaluation-protocol",
  "title": "AI evaluation protocol",
  "kind": "Synthetic evaluation-method sample",
  "purpose": "Defines a versioned dataset, acceptance threshold and failure handling before a model is released.",
  "content": {
    "version": "1.0",
    "dataset": "Fictional CivicAnswer corpus and 50-question benchmark",
    "decision": "Can the system select an approved source or refuse an unsupported question?",
    "measures": [
      "Expected source/refusal agreement",
      "Source-text fidelity",
      "Unsupported-question refusal"
    ],
    "acceptance": "Expected source/refusal agreement is assessed against the published cases. Review every failure; do not infer general model quality.",
    "humanReview": [
      "Confirm the expected cases",
      "Inspect false refusals and unsafe answers",
      "Approve deployment boundary"
    ],
    "limits": "A deterministic extractive demonstration is not model inference, a client result or evidence of general AI performance."
  }
}