Brief — Panel — Depth — Staff — Mandate — Review —
What is this mode for? Defaults below mirror Quorum's own Balanced mode. Change what you need in Panel and Deliberation, then come back here for the name.
Mandate
Continue → Panel
Panel — who deliberates Curate a specific engine pool, or let Quorum's live scoring pick the best available engines within your filters each run.
Pool strategy
Dynamic
Quorum's live ranking picks the best available engines each run, within your filters.
Fixed
You curate the exact engine pool. This mode always runs on what you pick here.
Seats
Pick who sits in each of the three seats. All three from one maker is refused — three engines from one lab agree with each other by construction.
Minimum access tier
Which tier can use this mode. If you pick a fixed pool with a frontier-only engine, this is raised automatically to match.
Free
Plus
Pro
Teams
← Back Continue → Deliberation
Deliberation — how they argue Depth is the biggest cost dial. Critique policy controls whether seats challenge each other before the answer is finalized.
Depth
Round budget and per-seat token ceiling.
Resolves to 2 rounds · 1.0× tokens
Critique policy
The real maestro.policy field qroute.js reads -- three real values, not a spectrum.
Never
Seats draft, synthesis runs directly on the drafts. No critique pass at all.
Gated
Critique only fires when seats actually conflict. The real equivalent of "escalate on disagreement."
Always
Every seat critiques every other seat's draft before synthesis. Highest quality ceiling, highest cost.
Panel routing
After the seats answer, one judge scores how far apart they are. This decides what happens next — writes follow_up_min_divergence and polish_min_divergence.
Not set
Keeps today's behaviour. The mode is not routed by the judge at all — nothing is assumed on your behalf.
Ships fast
Only argues when the seats really disagree. Most questions send the strongest answer straight out. Cheapest and quickest.
Balanced
Another round on real disagreement; combines the answers when they hold different useful pieces.
Argues readily
Sends the seats back over small differences. Most expensive, and the right choice when being wrong is costly.
Seats stating contradictory facts always go another round, whichever you pick.
← Back Continue → Staff
Staff — internal roles Pick which real engine plays each internal role for this mode. Leave on Default to use Quorum's shared staff. Candidates are filtered to engines compatible with what your panel already handles.
Synthesizerwrites the final answer
Pre-scorerclassifies difficulty
← Back Continue → Mandate
Mandate — domain grounding This is where a mode stops being a parameter tweak and becomes a real product. Not built yet -- shown so the lever set can be reviewed before that work starts.
Injected context Attach documents, a style guide, or a rubric the panel can reference.
📎 Requires document ingestion -- not built.
Jurisdiction Grounding sources and injected regulation for a specific region.
⚠️ Requires a real regional grounding source -- not built.
← Back Continue → Review
Review before publishing Nothing here is a performance claim. Those come from a certification run — a held-out-question test against your published listing, launched from the Certified stage in the rail once this mode has one.
Disclosure
Status: Draft, unpublished -- usable by you personally once saved. Not visible or usable by anyone else until published and approved.
Testing
Testing lives in its own tab above, so a run and the diagram it exercises stay on one surface. Open the Test tab
← Back
Save the mode to unlock the test rig.
Runs your saved mode through the real Optimization Lab rig -- real deliberation, real baseline comparison, real judging, charged to your own account.
Small
Medium
Large (Certified)
Certification runs against your published Marketplace listing -- publish this mode first, then run certification.
Customize settings
Questions per bucket
Judge repeats
Judging always runs pairwise (each judge votes Quorum vs. each baseline independently) -- the more reliable protocol, and the only one this rig runs going forward.
Run test
Cancel
Share this canvas
Not published yet — this canvas has no public link. Publish the mode to mint one, and the QR below the diagram becomes real.
Copy link
↓ Save as PDF
Certification runs a held-out question set against this mode and writes the badge a stranger sees. It needs a listing first: publish the mode, then launch the run from the Certified stage in the rail, which says what the run is before you press it, shows the score while it runs, and shows what it wrote when it finished.