stzLLMFunction

(sans commentaire dans la source : le nom se lit comme une phrase) initaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) name_aucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) set promptaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) returns numberaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) returns booleanaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) returns stringaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) returns one ofaucune extension

THE STRUCTURED RUNG. paFields is an stzOutputSchema declaration; it is compiled and judged HERE, so a defective declaration can never reach a model.QCSXTZZZIBWU

(sans commentaire dans la source : le nom se lit comme une phrase) has schemaaucune extension

A READ. Haro copies on assign, so configuring the returned schema configures a copy and nothing else -- use RefuseUnknownFields() below for the one knob that has to land on the stored one.aucune extension

Closed-world: a field the schema never declared refuses the answer instead of being reported and dropped.aucune extension

The findings the LAST validation produced, in the family's unified shape -- so a refusal can be handed to stzRuleReport.Ingest() and stand in the same CI gate as every other rule in the library.aucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) budgetaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) set retriesaucune extension

TEST / OFFLINE DOOR, named for what it is. The responder is a FAKE standing where the model stands: it receives (prompt, attempt) and returns the raw text a model would have returned. It exists so the refusal paths -- the ones that matter most and that no seeded cache can reach -- can be narrated without a GGUF. It spends budget like a real call, and Why() says "FAKE responder" on every answer it produces, so nothing it returns can be mistaken for a live one. (The name follows stzGraphRule.UseChecker(): same escape-hatch shape, same word.)QCSXTZZZIBWU

(sans commentaire dans la source : le nom se lit comme une phrase) is using responderaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) calls madeaucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) whyaucune extension

the callaussi écrite: Ofaucune extension

test/offline door: seed a known answer into the memo cache (the golden path for model-free environments; the seed is EXPLICIT)aucune extension

The prompt actually sent. For a structured function it carries the schema's own clause, so the model is ASKED for the shape rather than hoped at -- and the memo key is taken over this same text, so a seed and a call agree on what the question was.aucune extension

constrained decoding, asked and answered --------------------------- ON by default when a structure is declared and the engine enforces grammars. Turn it OFF to compare, or when a caller wants the model's unconstrained voice and the court's verdict on it.QCSXTZZZIBWU

Will the NEXT live call be grammar-constrained? Ask this rather than assuming: a nested structure, an off switch, or a build without the sampler rung all answer 0, and WhyNotConstrained() says which.aucune extension

(sans commentaire dans la source : le nom se lit comme une phrase) why not constrainedaucune extension

The grammar text itself, for a reader who wants to see what was enforced.aucune extension

Was the LAST live answer grammar-constrained? (0 for a memo hit and for a fake responder -- neither of them decoded anything.)aucune extension

The sampling a RETRY uses. Attempt 1 never sees these -- it is greedy.QCSXTZZZIBWU

(sans commentaire dans la source : le nom se lit comme une phrase) set max tokensQCSXTZZZIBWU

(sans commentaire dans la source : le nom se lit comme une phrase) retry temperatureaucune extension

golden setsQCSXTZZZIBWU

Goldens hold STRUCTURES as readily as scalars now. Two things had to change for that, and both were real defects rather than gaps: Haro's own `=` answers 0 for two identical lists, so a structured golden could never have passed; and a failing structured case that reports only "expected / got" is unreadable, so the failure now carries the FIELD that moved.aucune extension