model_advisory
Ask the trained models about a sentence before you write it down. This is ADVICE, not a verdict: `claim_check` decides by arithmetic over the figures a world declares, and nothing here changes what it decides. Three models answer. `assessor_gate` flags a sentence naming a construct the world says it cannot answer — the case arithmetic cannot catch, because the sums can be perfect while the construct is refused. `flip_discriminator` judges a time-scoped claim against a register's serialized state, which is the case where the arithmetic is not in the payload at all. `quote_attribution` judges whether text the sentence attributes to a record actually belongs to that record — supply the record's own text as `record_text`; it compares, it does not retrieve. EVERY FINDING ARRIVES WITH ITS OWN MEASUREMENT AND ITS OWN INSTABILITY, and both belong in whatever you write next. The gate scored 107 of 120 on a world withheld entirely from training, and moving a single world out of its training set shifted its false-reject rate by 37 points — so a `reject` is a reason to look again, not a finding of fact. A `no_objection` is the absence of an objection and licenses nothing whatsoever. A runtime with no models loaded answers `consulted: false`. That is not a pass.
Use this form for new arguments. Follow concrete links for known values. Fetch-only readers may be unable to submit new URLs.
Exact input contract
{
"properties": {
"record_text": {
"description": "optional: the record's own stated text (what the register actually says). Supply it when the sentence quotes or paraphrases text attributed to a record, and `quote_attribution` will judge whether that text belongs to this record — measured 200/200 on real register prose over titles withheld from training",
"type": "string"
},
"register_state": {
"description": "optional serialized register state; supply it to also get the flip discriminator's reading of a dated claim",
"type": "string"
},
"sentence": {
"description": "the prose you intend to write down",
"type": "string"
}
},
"required": [
"sentence"
],
"type": "object"
}