The model panel we use today

The pushback

Sonnet is here because it tends to phrase objections without becoming theatrical. It often finds the reservation that makes the memo useful.

Honest caveat

We like it as a skeptic, but the judge still has to decide whether the critique actually holds.

The organizer

GPT is the panel's sense of order. When the question is messy, it often helps put options, risks and next steps on the right shelf.

Honest caveat

The higher cost means it has to earn its place through structure, not just length.

The wider lens

Gemini is here for breadth. It is often good at widening the question: what assumptions are missing, what context have we not asked for?

Honest caveat

This is a preview model, so we watch it more closely than stable panelists.

Between disciplines

Grok has a seat as a challenger. It can be useful when a decision spans product, engineering and market at the same time, where risks often sit between disciplines.

Honest caveat

We use it for contrast, not to dominate the final recommendation.

The low-cost contrast

DeepSeek is here as a low-cost challenger and fallback. It gives the panel a different flavor without making Deep review unnecessarily expensive.

Honest caveat

V4 is tracked as a candidate; V3.2 stays until the newer line has passed our own cases.

The judge reserve

Fable is the judge's dedicated reserve: it steps in if Opus fails, and can serve as an adaptive judge when a question calls for it. It never sits on the panel itself — the reserve must not be a voice that already argued.

The Standard anchor

We use Large as the anchor in Standard. It is not there to impress with a model name; it usually gives the most coherent first assessment in the EU route.

Honest caveat

Because the alias is a latest alias, we check which concrete Mistral version it points to when the panel is updated.

The practical counter-voice

Medium is our practical counter-voice in Standard. It is often sharp enough to find objections without making a simple everyday question unnecessarily expensive.

Honest caveat

We use it as a balancing model, not as a single source of truth.

Speed and cost discipline

Small is here for speed and cost discipline. It often finds a simpler option or a cheaper objection that larger models can overwork.

Honest caveat

It is meant to challenge, not carry the whole decision by itself.

The separate judge

Magistral writes the Standard panel together. We use it as a separate judge so the same voice does not both argue and decide.