ADR-0010: Every reviewer model is configurable, NVIDIA's included; NVIDIA is only the default first reviewer, and no setup is blocked

Status: accepted · Date: 2026-10-19 Decided by: the owner · Written by: the advisor (claude-code, claude-opus-5-5), 2026-10-19 Extends: ADR-0008, ADR-0009 Trigger: the owner, after ADR-0009 (planning/CONVERSATION_2026-10-19.md, hop 5)

Context

ADR-0008 and ADR-0009 name NVIDIA as the default reviewer and recommend three models: A writes, NVIDIA reviews by default, B reviews special cases. They could be read as requiring NVIDIA, or as refusing a setup without an independent model family. The owner:

"And we make NVIDIA configurable too. Some customers may want to code with A and review with A, without NVIDIA, or use A and B. We don't block this. NVIDIA is just a model like the others, only the default first reviewing model."

Decision

1. Every reviewer model is configuration: the default reviewer and the special-review model alike. NVIDIA is the value proven ships with for the default reviewer, nothing more.
2. No setup is refused. A team may run:
   • A writes, A reviews, with no second model;
   • A writes, B reviews, without NVIDIA;
   • A writes, NVIDIA reviews, B reviews special cases, the recommended setup (ADR-0009).
3. Independence is recommended and reported, not enforced. Every review states which model it used and whether that model is the author's (ADR-0009, point 5), so a reader always sees which kind of independence the review had, including none beyond a fresh agent.
4. Evidence is never configurable. Every finding needs a test or a query and its rows (ADR-0002), whatever models are configured.

Consequences

• Proven fits whatever models a team already has, including a single one.
• A team that reviews with its writing model gets fresh-agent reviews, not cross-family ones, and every review says so.
• Proven's own runs at the hackathon use the default, NVIDIA, as the first reviewer.
• The open point of ADR-0009, a change written by the default reviewer's own family, becomes a configuration choice like any other: the review runs and says the reviewer's model is the author's. The owner, on the same day: "The same as class A-A: it's NVIDIA-NVIDIA. Still valuable, because the review is done by an agent with a fresh context, with no contact with the code before. A limited scenario, yet still useful."