Author a mode as a config object

Identity and framing. name and ticker are the only hard requirements on a create. mode_key is derived server-side from the name and made unique — you do not set it, and it is permanent once the mode is published, because it is what a deliberation call resolves against.

Whether seats are filled dynamically per request or pinned to a fixed pool. Pinning trades adaptability for reproducibility — certification cares about the latter. Every id is validated server-side; an id that does not resolve is a hard 400, never a silent fallback.

Two rules refuse a save. A new Mode must send pool_strategy; without it the answer is 400 seat_selection_required, because a silent default would pick for you. And "fixed" needs all six seats: three engine_ids and three backup_engine_ids, or 400 six_seats_required.

Round ceiling and whether seats critique each other. The ceiling is a cap, not a target — the Judge exits early when the panel has converged, which is most of the time. The four depth presets are real named configurations, not a slider:

Per-mode overrides for the three non-seat roles. Leave them unset and the mode inherits Quorum's current assignment, which moves as the field moves. Set one and it is pinned until you change it — the server validates that the engine exists, is active, and covers every input modality your own seats require, and returns 400 rather than quietly substituting a different one.

Domain grounding — injected context and jurisdiction. This is where a mode stops being a parameter tweak and becomes a real product. It is listed here because the lever set is real and reviewable, and omitting it would misrepresent the shape of the config object. It accepts no fields today.

What the Marketplace shows, and what a customer is billed to call the mode through the API. Ingredients — provider families, round band, whether seats critique each other — are derived from the config above rather than written here, so you cannot claim a panel shape the mode does not run. Your own revenue share is not settable from this object at all: the payout weight is server-forced, permanently, whatever the request body contains.

Everything above, composed. This is the exact body /api/mode-builder accepts — POST to create a draft (name and ticker required), PATCH with mode_id to update one. Omitted keys keep their defaults.

This page composes the request body from the fields above. It does not send it — copy it into your own client, or use the visual builder if you would rather save from a UI.

These two are the real API surface, and unlike authoring they take a project API key. Both reference the mode by model — the api_model_id from GET /v1/models, not the internal mode key.

/v1/tests is the metered run: you choose the buckets (light, medium, hard), up to 100 questions each and 300 in total, priced per question at a 25% discount to the normal rate. Config stays editable across test runs. Send dry_run: true to price a run without launching it. Up to 10 launches per organization per 24h.

POST /v1/tests
Authorization: Bearer <api_key>

{
  "model": "quorum-standard",
  "buckets": { "light": 10, "medium": 10, "hard": 10 },
  "dry_run": true
}

{
  "ok": true,
  "dry_run": true,
  "model": "quorum-standard",
  "total_questions": 30,
  "per_bucket": { ... },
  "total_cost_usd": ...,
  "estimated_minutes": 30
}

/v1/certify is the flat-price run: $50, a fixed 150-question set split 50/50/50 across the three buckets, judged pairwise. It writes the score the Marketplace listing carries. Up to 3 launches per organization per 24h. A certified mode whose config changes is no longer certified.

POST /v1/certify
Authorization: Bearer <api_key>

{
  "model": "quorum-standard"
}

{
  "ok": true,
  "batch_id": "…",
  "model": "quorum-standard",
  "total_questions": 150,
  "billed_usd": 50,
  "status_url": "/v1/certify?batch_id=…"
}

Real, disclosed constraint, returned in the response rather than buried: the shared execution engine processes roughly one question per minute per batch. A 150-question certification run takes about two and a half hours, not seconds. Poll status_url — GET /v1/certify?batch_id=… and GET /v1/tests?batch_id=… both return progress, and the certification verdict once status is completed.