Provider-neutral evaluation hub (judge::evaluate, judge::cancel, and the internal judge::models::list) that forwards typed Noul, Choice and Score questions to the selected judge-<provider> worker; the default provider is a Console setting seeded from JUDGE_PROVIDER (typesafe). Also resolves worker-defined @<name>(id=…) chat mentions for agents (judge::mentions::resolve and a harness pre-generate hook).
TypeSafe evaluation (judge-typesafe::evaluate), model listing (judge-typesafe::models::list) and caller-scoped cancellation (judge-typesafe::cancel), with typed usage, provider errors, per-call retries and worker-owned credentials.
laya provider for the judge hub (judge-laya::evaluate, judge-laya::models::list, judge-laya::cancel) — the ModernBERT typed-decision checkpoints (English, multilingual, typed-decisions) downloaded from the Hugging Face Hub and run in-process with language routing; no API key or external service.
decider provider for the judge hub (judge-decider::evaluate, judge-decider::models::list, judge-decider::cancel) — decider-4b v2 (a Qwen3.5-4B-Base fine-tune for typed decisions, GGUF) downloaded from the Hugging Face Hub and run in-process through llama.cpp, answering from the next-token logits of the option labels at its fitted temperature; picks Metal (macOS), Vulkan (Linux, when a driver is present) or the CPU automatically; no API key or external service.
clef provider for the judge hub (judge-clef::evaluate, judge-clef::models::list, judge-clef::cancel) — Cloudflare's Clef-Flash decision model (a Qwen3.5-9B backbone and its joint schema head in one GGUF) downloaded from the Hugging Face Hub and run in-process through llama.cpp's clef architecture, deciding every question of an evaluation together in one forward pass; picks Metal (macOS), Vulkan (Linux x86_64, when a driver is present) or the CPU automatically; no API key or external service.
SemIf provider for the judge hub (judge-semif::evaluate, judge-semif::models::list, judge-semif::cancel) — a frozen open LLM (Qwen3.5-4B GGUF) downloaded from the Hugging Face Hub and run in-process through llama.cpp, answering from the next-token logits of the option letters; picks Metal (macOS), Vulkan (Linux, when a driver is present) or the CPU automatically; no API key or external service.
