# Consensus Groups (@leumas/consensus)

Panels of AI members that deliberate and synthesize ONE answer. Generalizes the legacy ollama-consensus pipeline over every provider, and ports the MIT council-of-high-intelligence deliberation...


Panels of AI members that deliberate and synthesize ONE answer. Generalizes the legacy
`ollama-consensus` pipeline over every provider, and ports the MIT `council-of-high-intelligence`
deliberation protocol as code.

## Model

A group (`/db/consensus_groups`, zero-backend CRUD) = `members[]` + `chairman` + `mode` + `settings`.
- Member = a `chatbots` row (`kind:'chatbot'`, e.g. the 18 seeded `council-*` personas) or inline
  `{ provider, model, system }`. Optional per member: `domainSeat` (×1.5 vote weight),
  `research: { enabled, loops, searchApi }` (runs @leumas/research before answering).
- Chairman = the synthesizer; **never a panelist** (enforced).
- Modes: `panel` (parallel + judge; fast) · `council` (restate gate → blind Round 1 → ANONYMIZED
  Round 2 cross-exam → Round 3 `STANCE: x | CONFIDENCE: high|med|low | DEALBREAKER: yes|no` →
  confidence-weighted vote: consensus iff ≥⅔ of total BASE weight, else the honest split →
  Chairman verdict leading with unknowns) · `quick` (2 rounds + vote) · `duo` (dialectic).

## How agents call it

| Surface | How |
|---|---|
| Chatbot / Leviathan | attach the `run_consensus` functioncall (executor `action:run-consensus`) |
| Automation rule | action `run-consensus` `{ groupId, prompt, mode? }` |
| MCP client | `functioncalls` server → `run_consensus` tool |
| HTTP | `POST /api/consensus/run[/stream]` with `{ groupId }` OR an inline `{ group }` (ad-hoc panels — AiLab does this) |
| Studio | Intelligence → Consensus Groups / Group Chat; Builder → AI Lab (inline local panels) |
| Code | `runConsensus({ group, prompt, mode, onEvent, signal, providerRegistry, getBot, research, connector })` |

## SSE vocabulary

`start · phase(restate|round1|round2|round3|vote|synthesis|panel) · member_start · member_done ·
research (nested per-member) · stance · vote · synth_start · consensus · done · error`.
Runs persist to `/db/consensus_runs`, checkpointed per phase (dropped streams lose nothing).

## Gotchas

- Panel synthesis imports `buildConsensusPrompt` + `CONSENSUS_SYSTEM` from `@leumas/a-ollama` — never copy them.
- All-Ollama panels default to concurrency 2 (one GPU serializes); Studio prewarms via `/api/ollama/prewarm`.
- Stance labels are cleaned of markdown/quote decoration before tallying (models emit `** monorepo`).
- Providers without a step `adapter` (cursor-agent, codex, slm) answer one-shot — fine for members.
- Council cost ≈ members × 4 calls + chairman; Group Chat shows the projected call count pre-run.

## Gating

`requireAuth` + `requireMembership('intelligence')` (admins bypass) + PassNode guard
`feature:consensus.run` (author a `passnode_rules` doc to meter). Smoke: `pnpm smoke:consensus`.


---
Source: shared/services/knowledge/build-knowledge/consensus-groups.md
Canonical: https://docs.leumas.tech/p/how-to/consensus-groups