Files

31 lines
1.2 KiB
JSON

{
"id": "router.model@1",
"description": "Which model tier should serve a request: the smallest that answers it correctly. State: {request, system, turns, tools, prompt_tokens, images}.",
"calibration": "cal_c6ffe2e7b0c8819a",
"questions": {
"tier": {
"type": "choice",
"instructions": "What is the smallest model tier that would answer this request correctly?",
"criteria": {
"small": "a small fast model: lookup, formatting, extraction, short factual answers",
"medium": "a mid-size model: ordinary code edits, summaries, one or two reasoning steps",
"large": "a large model: multi-step reasoning, non-trivial code, long context",
"frontier": "the strongest model: hard proofs, research-grade problems, long plans under ambiguity"
}
},
"long_context": {
"type": "noul",
"instructions": "Does answering require reading a long document, many files or a long history?"
},
"private_model_required": {
"type": "noul",
"instructions": "Must the request stay on a private, self-hosted model because of the data it involves?"
}
},
"thresholds": {
"tier": 0.5,
"long_context": 0.5,
"private_model_required": 0.3
}
}