Construct AI triage input and evidence gates

Repair ai-triage-rules.json using hitPolicy unique or first and 1..24 distinct {id,when,then} rows. Inputs: data public,internal,restricted; purpose triage,model-training; provider approved,unreviewed; grounding corroborated,unsupported,injected. when nonempty category arrays combine by AND; omitted fields are wildcards. then must contain input (proceed,minimize,withhold), handling (purpose-review,provider-review,private-route,quarantine-output,analyst-review,verify-evidence), autonomy (boolean true or false). Apply this local policy in precedence order: model-training always yields withhold/purpose-review. For triage, unreviewed provider always yields withhold/provider-review. Triage at an approved provider with restricted data always yields withhold/private-route. For approved triage with public/internal data, injected grounding always yields withhold/quarantine-output. Remaining public data may proceed; remaining internal data must minimize. For both, corroborated grounding requires analyst-review; unsupported grounding requires verify-evidence. Earlier permission gates determine the handling reason even when grounding is injected; an input permission failure cannot be bypassed by a grounded-looking result. Autonomy is false for EVERY context. No row authorizes model-driven containment, incident closure or evidence deletion, including corroborated public input. A model-training opt-out or provider label cannot replace permitted purpose/classification. Evaluate all 36 contexts. Unique requires exactly one matching row. First permits deliberate overlap when earlier permission gates precede later evidence fallbacks. Every context needs exact outputs; every row must contribute. Use only supplied fields, typed values and meaningful reachable priority. Disjoint gates or ordered broad fallbacks are valid. Run tests, inspect actual input/handling/authority mismatches, save and repair without Reset.