@leumas/consensus

Consensus Groups: panels of AI members that deliberate and synthesize one answer.

• A group (/db/consensus_groups, zero-backend CRUD) = members + a chairman + a mode + settings.

A member is a chatbots row (kind: 'chatbot' — the 18 seeded council personas work out of the box) or an inline { provider, model, system }. Any member can be research-augmented (research: { enabled, loops, searchApi }) — it runs @leumas/research before answering.

• Modes
• panel — parallel answers → judge synthesis (the legacy ollama-consensus pipeline,

generalized over every provider; synthesis imports buildConsensusPrompt from @leumas/a-ollama).

• council — the full council-of-high-intelligence protocol (MIT): restate gate → blind Round 1 →

anonymized Round 2 cross-examination (anti-conformity directive) → Round 3 STANCE: lines → confidence-weighted vote (⅔ of total base weight; domain seat ×1.5; honest split when unmet) → a separate Chairman model synthesizes the verdict, leading with unknowns.

• quick — 2 rounds (rapid analysis → anonymized stances), same vote + chairman.
• duo — 2-member dialectic (no anonymization), chairman adjudicates.
• Surfaces — createConsensusRouter at /api/consensus (POST /run + POST /run/stream SSE),

the run-consensus action (rules/functioncalls/MCP), and Studio's Intelligence domain.

• Gating — logged-in + intelligence membership (admins bypass) + PassNode feature:consensus.run.

Runs persist to /db/consensus_runs (checkpointed per phase). Smoke: pnpm smoke:consensus.