# Architecture Commerce-1 packages a complete native Transformers checkpoint, tokenizer, processor configuration, chat template, and decision readout. The checkpoint is derived from the pinned Qwen3.8-27B base through the governed V3 training path; it does not rely on downloading a mutable model identifier at runtime. The decision model receives structured state and typed questions. Choice maps named alternatives to a categorical distribution. Noul returns probability for a binary false/true decision. Score returns an ordered categorical distribution and its expected score. A dedicated 255-option readout supplies logits for all three primitives. Held-out, per-primitive temperatures are applied before answer projection. The release layout separates concerns: - `model/` contains the native full-checkpoint shards and model metadata. - `calibration/calibration.json` binds fitted temperatures to the artifact. - `source/autojev/` contains the exact three-file serving source closure. - `runtime.json` binds checkpoint, calibration, source, and the content-free public evaluation attestation. - `evidence/` contains content-free public evaluation and lineage attestations. Raw authorization, spend, provider-operation, and orchestration records stay in protected release controls and are not distributed with the model. - `release-manifest.json` and `SHA256SUMS` bind the complete staged package. The native backend validates those identities before importing AutoJev. It then loads only local checkpoint paths with offline mode enabled. The package is a typed decision runtime, not a claim that ordinary causal language-generation APIs reproduce its behavior.