Didactopus/examples/arena/arena_review_queue.json

32 lines
630 B
JSON

[
{
"candidate_name": "stub-baseline",
"overall_rating": "borderline",
"overall_score": 0.733,
"needs_human_review": true,
"weak_roles": [
"mentor",
"evaluator"
]
},
{
"candidate_name": "stub-strict-grounding",
"overall_rating": "inadequate",
"overall_score": 0.547,
"needs_human_review": true,
"weak_roles": [
"mentor",
"evaluator"
]
},
{
"candidate_name": "stub-trust-preserving",
"overall_rating": "inadequate",
"overall_score": 0.547,
"needs_human_review": true,
"weak_roles": [
"mentor",
"evaluator"
]
}
]