How to use the Model Council — deliberation, export, records, Telegram, quota, and more.
The Model Council is a multi-model deliberation engine. Four frontier AI models — GPT-5.5 Pro, Claude Sonnet 4.6, DeepSeek R1, and Qwen 3.7 Max — debate your question across three structured rounds: Opening Statements, Cross-Examination, and Rebuttal. An orchestrator model reads all three rounds and produces a single synthesized verdict with claims, confidence scores, dissent tracking, and an auditable Evidence Basis.
Model Council is the general-purpose adversarial deliberation engine — open-ended questions, broad domain expertise, exploration and analysis. Veritas is the same engine, precision-tuned for professional expertise. It deploys domain-specific personas for legal, accounting, and consulting, with regulatory frameworks and provenance-signed artifacts designed for audit-grade professional use. Same pipeline. Same governance. Different context.
The active council seats are: GPT-5.5 Pro (OpenAI), Claude Sonnet 4.6 (Anthropic), DeepSeek R1, and Qwen 3.7 Max. The orchestrator — which reads all rounds and produces the final verdict — uses Claude Opus 4.8. Models may change as newer versions become available; the council roster is audited daily against live OpenRouter availability.
Sign in with Google or a magic link email. Type your question into the prompt input — the more specific, the better the deliberation. A question where reasonable experts might disagree produces the richest output. The Council runs for 2–5 minutes depending on model response times and how many rounds are needed.
Above the prompt input on the main page, four template chips are available: Compare, Decide, Review, Validate. Click any chip to autofill the textarea with a structured prompt. Templates are starting points — edit the prompt freely before submitting. The "Custom" chip clears to a blank input.
Yes. The prompt input accepts up to 2,000 characters. This is enough for a detailed, well-structured question — think a substantial paragraph, not a sentence fragment. If you need more space, consider breaking your query into focused sub-questions and running separate deliberations.
Yes. Upload a PDF on the deliberation page or send one directly to @NiraNexus_ModelCouncilBot on Telegram, then run /deliberate with your question. The Council cross-examines your uploaded document — models are grounded in the source text, not parametric memory. Document-augmented deliberation consumes 3 debates (reflecting the additional processing and context grounding).
Typically 2–5 minutes. A deliberation that converges in Round 1 may finish faster. A three-round deliberation with cross-examination and rebuttal takes longer. The timer updates in real-time during the debate.
After a deliberation completes, click the "Export" button in the verdict card. Choose PDF (print-optimized with NiraNexus provenance footer), Markdown (structured document), or JSON (machine-readable with _source field). You can also export from the Records page by clicking "Export" on any row.
Every deliberation produces a public URL at model-council.niranexus.com/debate/[id]. Click "Share" on the verdict card or on any row in Records to copy the public link. Anyone with the link can view the full deliberation — no authentication required. The public page mirrors the in-app viewer with the same structure: claims, confidence scores, dissent, and evidence basis.
The Evidence Basis is an auditable record of the sources each model used to support its claims. Every claim in a verdict is labeled: VERIFIED (retrieved evidence directly supports it), DISPUTED (conflicting evidence detected between models), or UNVERIFIED (insufficient or low-quality data). This is not a bibliography — it is a traceable chain of what each model found, where it came from, and how the Council resolved (or flagged) discrepancies. The Evidence Basis is what turns AI-assisted reasoning from a black box into a defensible artifact.
Confidence scores are self-reported by each model for specific claims in the verdict. Scores range from 0–100% and are categorized: High confidence (≥80%), Moderate (≥60%), Guarded (≥40%), and Low (<40%). Each model has its own color bar for visual identification — not a universal green/amber/red system. The orchestrator also reports its own confidence in the synthesized verdict. All scores are displayed as progress bars with aria-labels for accessibility. Scores below 70% trigger a low-confidence flag in the governance cockpit.
Every debate records execution metrics: models deployed, API calls executed, fallbacks triggered, whether a verdict was produced, duration, and estimated tokens consumed. These appear in a collapsible EXECUTION METRICS section below the verdict on the main page, public records, and in export files. No consumption or billing data is exposed.
Dissent is surfaced disagreement between models. When a model disagrees with a claim or verdict, that disagreement is recorded — not buried. The verdict includes a dedicated Dissent section. Claims that have dissenting models are flagged with a warning icon. Dissent is a feature, not a bug: the Council is designed to expose disagreement, not manufacture consensus.
From the main deliberation page, open your history panel to browse past debates. You can also navigate directly to /records for a paginated, searchable table of all your deliberations. Click any row to re-open that deliberation and view the full verdict.
On the Records page, click "Expunge" on any row. You will be asked to confirm — deletion is permanent and removes the deliberation and all associated rounds, responses, and verdict from Supabase.
The free tier includes a limited number of debates. Each deliberation consumes one debate regardless of how many rounds it runs. Additional debate bundles are available for purchase. When your debate capacity is exhausted, the capacity selector appears directly on the deliberation page. Debate enforcement is at the engine level: the deliberation will not start if you have no remaining debates.
When your debate capacity is exhausted, a capacity selector appears on the deliberation page. Select the bundle that matches your usage, and capacity is provisioned immediately. The council operates as self-service infrastructure — no sales pipeline, no friction.
Sign in to the Model Council, then visit the Telegram setup page at /telegram. Generate a verification code, send /link [code] to @NiraNexus_ModelCouncilBot on Telegram. Once linked, you can deliberate directly from Telegram: /deliberate [your question]. Send a PDF attachment before your /deliberate command to run a document-grounded deliberation. Other commands: /status (check remaining debates), /clear (reset pending state), /start (bot overview). The same governance and engine apply — a new entry point, not a separate component.
Yes. POST /api/deliberate accepts a JSON body with prompt, optional rounds (2 or 3), and optional model selection. Authentication via Bearer JWT from your Supabase session. Response is an SSE stream identical to the web UI. GET /api/deliberate returns interactive documentation.
Yes. Row-Level Security (RLS) on Supabase ensures you can only see your own deliberations. Public deliberation records are opt-in — you choose to share a link. No deliberation content is used to train models. Data is stored in Supabase with encryption in transit and at rest.
Each model has a fallback chain — if the primary model is unavailable, the engine tries alternatives from the same provider. If all models in a chain fail, that seat is marked as unavailable in the verdict. The deliberation continues with the remaining models. The verdict will note which models contributed.
Yes. If you ask the exact same question, the Council may return a cached verdict — faster, but the debate is still consumed per standard policy. The cache is per-user, so another user asking the same question gets their own fresh deliberation.