Model Configuration
Run the shared model roster and the role matrix — adding models, controlling vision, assigning every seat, and rolling out changes safely.
Settings → Administration → Models controls the AI setup for the entire workspace. It has two halves that work together: the roster decides which models exist here at all; the role matrix decides what each one does by default.
The model roster
The roster is the allowed pool. Nothing outside it can be selected by any user, pinned in any override, or assigned to any seat.
For each model, admins can:
- Add it from the gateway catalog. Use the Dr Moot Model Catalog to compare every available model's capabilities, limits, and release date before adding it.
- Enable / disable it. Disabling removes it from every picker immediately — and any user override pinned to it silently reverts that seat to the system default.
- Mark vision support. This flag decides whether image attachments can reach a seat filled by this model. Get it wrong in either direction and users either lose image support they should have, or send images into a model that can't see them.
- Reorder the roster — this is the order users see in pickers.
- Remove it entirely.
The role matrix
The matrix assigns a default model to every seat in every mode:
| Mode | Seats |
|---|---|
| Relay | Model A · Model B · Synthesizer |
| Debate | Model A · Model B · Synthesizer |
| Moot | Generator · Sceptic · Specialist · Chair |
Plus two supporting roles that run behind the scenes:
- Auto-mode router — the model that reads each Auto question and decides Relay vs Moot, rounds, and whether the question is high-stakes (which triggers verification). A small, fast model is the right choice here — it runs on every Auto question.
- Moot consensus checker — decides after each Moot round whether the panel is settled or needs another round. Same guidance: fast and cheap.
Fallback chains are shown read-only: if a seat's model fails mid-run, these define what steps in.
Assignment guidance
- Generator / Model A — your strongest reasoning model; it does the constructive work.
- Sceptic / Model B — ideally a different provider than the Generator. Models from one family share blind spots; cross-provider pairings argue harder.
- Specialist — a model with strong factual grounding.
- Chair / Synthesizer — a model you trust to weigh arguments rather than add new ones.
What your choices change for users
- Which models appear in every picker (composer overrides and settings overrides alike).
- Which model fills each seat for everyone who left "Use system default" — that's most users, most seats.
- Whether image attachments work for a given mode (via the vision flags on the assigned models).
- The cost and quality profile of Auto routing and Moot consensus, via the supporting-role assignments.
Rolling out a change safely
- Add the model to the roster (disabled or enabled, but unassigned).
- Set its vision flag correctly.
- Pin it yourself first — use a personal override in your own settings and run real questions through it.
- When satisfied, assign it in the role matrix.
- Test each affected mode in the Lab after saving — one Relay, one Debate, one Moot run.
The one destructive mistake
Saving a roster with no enabled models disables AI for the whole workspace. Dr Moot warns before applying that change — treat the warning as a stop sign, not a formality.