Recall verdicts
SPM separates what it found from whether the evidence is sufficient to answer.
Only evidence_decision.status is authoritative for that decision.
{
"evidence_decision": {
"schema": "evidence-decision-v1",
"status": "supported",
"sufficiency_judged": false,
"sufficiency_source": "gate_only",
"support_failure_kinds": [],
"why_no_answer": null
}
}
| Field | Meaning |
|---|---|
status |
supported, insufficient, contradicted, or unknown |
sufficiency_judged |
true only when a deep selector judged the evidence sufficient |
sufficiency_source |
deep_selector or deterministic gate_only |
support_failure_kinds |
Why evidence was refused, such as cjk_floor, identifier, entity_binding, or capture_origin_unverified |
why_no_answer |
Machine-readable {code, message, action_hint} for non-supported outcomes |
Three rules are stable:
- Only
supportedmeans the evidence answers the question. - An empty result never reports a passed or sufficient gate.
gate_onlyis honest deterministic acceptance, not a claim that a model independently judged sufficiency.
Proxy captures whose speaker cannot be attributed may remain visible as
evidence, but they cannot stand alone as a supported answer. Explicitly save
important facts with MCP remember when provenance matters.
See MCP, Memory, and the Provider Proxy API for integration details.