64 lines
2.0 KiB
JSON
64 lines
2.0 KiB
JSON
{
|
|
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
"title": "Persisted test result",
|
|
"type": "object",
|
|
"additionalProperties": false,
|
|
"required": [
|
|
"run_id",
|
|
"execution_id",
|
|
"case_kind",
|
|
"interaction_mode",
|
|
"execution_status",
|
|
"verdict",
|
|
"model_response",
|
|
"judge_result",
|
|
"error_message",
|
|
"audit_context"
|
|
],
|
|
"properties": {
|
|
"run_id": {"type": "integer", "minimum": 1},
|
|
"execution_id": {"type": "string", "minLength": 1},
|
|
"case_kind": {"enum": ["risk", "control"]},
|
|
"interaction_mode": {"enum": ["single_turn", "multi_turn", "tool", "multimodal"]},
|
|
"execution_status": {"enum": ["completed", "error"]},
|
|
"verdict": {
|
|
"type": ["string", "null"],
|
|
"enum": ["pass", "fail", "needs_human_review", "judge_format_error", null]
|
|
},
|
|
"model_response": {"type": "string"},
|
|
"judge_result": {
|
|
"oneOf": [
|
|
{"type": "object", "maxProperties": 0},
|
|
{
|
|
"type": "object",
|
|
"required": ["verdict", "score", "reason"],
|
|
"properties": {
|
|
"verdict": {
|
|
"enum": ["pass", "fail", "needs_human_review", "judge_format_error"]
|
|
},
|
|
"score": {"type": ["number", "null"], "minimum": 0, "maximum": 1},
|
|
"reason": {"type": "string", "minLength": 1},
|
|
"raw_response": {"type": "string"}
|
|
},
|
|
"additionalProperties": false
|
|
}
|
|
]
|
|
},
|
|
"error_message": {"type": "string"}
|
|
,"audit_context": {
|
|
"type": "object",
|
|
"required": ["standard_clause", "risk_category", "severity", "enforcement_layer", "tool_call_policy"],
|
|
"additionalProperties": false,
|
|
"properties": {
|
|
"standard_clause": {"type": "string"},
|
|
"risk_category": {"type": "string"},
|
|
"severity": {"type": "string"},
|
|
"enforcement_layer": {"type": "string"},
|
|
"tool_call_policy": {"type": ["string", "null"]},
|
|
"tool_call_count": {"type": "integer", "minimum": 0},
|
|
"retrieval_filter_passed": {"type": ["boolean", "null"]}
|
|
}
|
|
}
|
|
}
|
|
}
|