Struct QwenExpert 

pub struct QwenExpert { /* private fields */ }

A safety expert logit source for contrastive steering (ADR-013): a second Qwen engine — in production base + a safety LoRA; here any same-tokenizer Qwen GGUF — loaded through the ADR-006 provenance gate and primed with the turn’s prompt so its logits align with the base engine’s. The session feeds it the committed tokens via ExpertLogits::logits.

The weights are loaded once and kept resident like the base model (ADR-018 expert persistence): the provider holds it across turns and calls reprime per turn (reset_cache + prefill, no disk reload). State lives behind a Mutex (not RefCell/Cell) so the resident expert is Send + Sync and can sit in the Send + Sync provider.

Steering is bounded to the early-token window (ADR-013), so the expert runs only for the first steer_window tokens. When the base engine rolls back (committed output shrinks), the expert re-primes to the prompt and re-feeds the retained prefix, so its contrastive context stays aligned with the base rather than serving logits from the abandoned branch. Pointing this at the chat model itself yields ~zero contrast (a no-op); a safety-tuned Qwen GGUF gives real steering.

Implementations

impl QwenExpert

pub fn from_path_primed( path: impl AsRef<Path>, eos: Token, prompt: &[Token], permit: LoadPermit, ) -> Result<Self>

Load the expert GGUF, gate it (ADR-006 — permit is required, not optional), and prime it with prompt so its KV state matches the base engine’s post-prefill state.

pub fn reprime(&self, prompt: &[Token]) -> Result<()>

Re-prime the resident expert to a new turn’s prompt without reloading the GGUF (ADR-018 expert persistence): discard the prior turn’s KV and prefill the new prompt on the same loaded weights. The expensive part — reading + parsing the GGUF — happens once in from_path_primed; this only re-runs the (cheap, bounded) prompt prefill.

Trait Implementations

impl ExpertLogits for QwenExpert

fn logits(&self, committed: &[Token]) -> Vec<i32>

Expert next-token milli-logits given the committed (generated) context.