Files

21 lines
646 B
JSON

{
"id": "reasoning.budget@1",
"description": "How much the serving model should think before it answers. State: {request, system, turns, tools, prompt_tokens, images}.",
"calibration": "cal_c6ffe2e7b0c8819a",
"questions": {
"budget": {
"type": "score",
"instructions": "How much reasoning does a correct answer to this request take?",
"criteria": [
"none: recall, formatting or a direct answer",
"short: one or two inference steps",
"medium: a chain of dependent steps",
"deep: search, proof or planning under uncertainty"
]
}
},
"thresholds": {
"budget": 0.5
}
}