Architecture

Attributes → brief validation → cache check → template select → WRITE (model A) → QC CHECK (model B) → capped retry → publish. The brief, cache, template and the QC floor (exact banned-word match, required-facts presence, claim-vs-attributes, language, format) are all deterministic and run at $0 in both cloud and OSS modes — they are authoritative for every hard rule.

The metered stages are the writer and the checker. Cloud uses claude-haiku-4-5 for both roles; OSS uses Qwen3-8B (writer) + Qwen3Guard-Gen-4B (checker) recorded on local M4 hardware ($0 replay). A fail-closed budget cap degrades cloud to $0 OSS. The deterministic QC remains authoritative; the cloud checker adds richer semantic claim-flagging on top.