cratefield-adapter-classifier-llm

Classifier port over the harness's own TextModel port for the Cratefield harness (issue #456). The third of the three Classifier adapters and the only one with no vendor of its own: it reaches whichever model the venture's TextModel tiers are wired to — a local model included — asking for JSON-schema structured output. When neither a purpose-trained classifier nor a vendor binding is wired, this is the honest fallback.

Usage

use cratefield_adapter_classifier_llm::ClassifierLlm;
use std::sync::Arc;

// `text_model` is the `Arc<dyn TextModel>` the venture already wired —
// the same router its drafting modules use. No new account, no new key.
let runtime = runtime.classifier(Arc::new(ClassifierLlm::new(text_model)));

// A model whose context window cannot carry the default 96_000-char
// state budget says so here, the only place that knows it:
let adapter = ClassifierLlm::new(text_model)
    .tier(cratefield_core::ModelTier::Fast)
    .max_state_chars(24_000);

Calibration is the language model's

This adapter reports Calibration::LanguageModel, and that is not a technicality. A probability elicited from a general language model is not the same number as one from a purpose-trained classifier. A 0.8 this adapter returns is the model's own rating of its certainty — sharper when the model is confident, floppier at the decision boundary, and drifting with every model version the venture re-wires the tier to. A threshold tuned against a Calibration::Classifier adapter is simply wrong against this one. A module that thresholds on Answer::confidence documents the calibration it tuned against, and re-tunes when the adapter swaps.

Behaviour

• One TextModel::complete call per ask, carrying all questions — asking as a set is the port's whole point. The prompt requests structured output through Prompt::json_schema, with a schema built per call that pins, for every question id, the chosen label and a probability for every label the question actually offers.
• Structured output is a request: a provider that cannot honour the schema answers plain text (Completion::json is None). The adapter then parses Completion::text as JSON, tolerating the markdown code fence real models wrap it in. Neither yields a JSON object with an answers mapping and the call fails Transport — with a message that does not quote the completion back.
• The model's answer is validated, never trusted: a missing or invented question id, a chosen label the question never offered, a probability under a label the question never offered, or a distribution that is not finite and normalisable is Rejected. A distribution that sums to something other than 1 is normalised; NaN and infinities never pass.
• state longer than max_state_chars is truncated with the port's deterministic, char-boundary-safe cut, and a tracing::warn! fires. The default is the port's conservative ceiling; the real ceiling is the underlying model's context window, which this port cannot see — set .max_state_chars to match whatever the tiers are wired to.
• Errors map one-to-one from TextModelError: NotConfigured (the tier is unwired — how "no classifier vendor wired" surfaces honestly), Rejected, Transient { retry_after } with the back-off preserved, and Transport. Provider text only ever rides inside variants whose Display scrubs it.

MIT. Built in the open for Cratefield, a Factory Zero venture.