YAI controls model invocation boundaries. It does not treat model text as case authority or carrier execution.
- Provider runtime handles local or external model providers as bounded invocation targets.
- LAN target support is a provider/runtime boundary, not a remote authority model.
- Model carrier work binds model invocation to case, subject, policy, receipt import and observation.
- Runner/model runtime doctrine keeps model execution separate from operational authority.
Model output can become a claim, proposal, observation or result envelope. YAI must still decide whether it can affect the case, create memory, invoke a carrier or enter review.
SPINE.48 derives fact_model_behavior from explicit model records and stable
markers only. It records posture such as authority_overclaim,
unsupported_claim, review_required, refusal and filesystem operation
proposal markers. The fact is analytical: facts are not truth. model proposal
is not execution. model cannot approve. automatic proposed-op gate import is
future work. No LLM-based classification is used.
MTP/speculative decoding, native model runners and richer provider backends are future runtime work. CLORI is a possible future or parallel native neural execution component; no CLORI implementation is claimed in this repository state.
A model must not reconstruct the active case from scratch for every request. YAI plans ContextFrame base/delta state, ContextDelta invalidation, CaseModelSession binding and runner KV/cache boundaries.
CaseModelSession binds case, actor, model, runner, provider, context frame, policy scope, retrieval providers and cache policy. It is not authority.
Runner KV/cache state is an optimization scoped by model, runner, tokenizer, context hash, prefix hash and valid frame id. KV cache is not YAI memory, truth, authority or provenance.