Model Assignments Locked into the Wiki
wiki/ai/models.md now documents per-step model assignments for the design-feedback-v2 pipeline, including the Gemini quota caveat and a DeepSeek migration path. The Groq decision has a permanent home.
The A/B test conclusion needed to live somewhere durable. wiki/ai/models.md now carries the full per-step model assignments for the design-feedback-v2 pipeline — what runs on Groq qwen3-32b, what still touches Gemini, and where DeepSeek slides in if quota becomes a blocker.
The section added is design-feedback-v2: fully online + parallel variants (Phase 4B enhancement). It covers three things: the confirmed Groq assignment for critique steps, a note that Gemini free-tier quota will bite you if you run parallel variants at volume, and a documented swap path to DeepSeek for those steps if that happens. Nothing in the n8n workflow changed — this is documentation catching up to decisions already running in production.
- wiki/ai/models.md — design-feedback-v2 section
What this unblocks: Any future phase that adds or swaps a model step now has a reference format to follow in the wiki. The DeepSeek migration path being written down means that swap, if needed, is a 30-minute task rather than a re-evaluation.