Inference Sessions

Runs go to a TEE-attested model on 0G with fallbacks disabled, and every run is content-addressed: a model hash over the model reference and an artifact hash over model, prompt and output. Those digests are what a later dispute is argued from — the output itself is never authority.