Let multiple model perspectives debate a decision, then verify the verdict.
@alembic/council implements a board of members — each with a role, weight, trust, and adapter — that deliberates over a question. It scores decisions on five axes (feasibility, revenue, customer experience, time-to-market, risk) and returns a verdict: GO, PIVOT, or NO_GO.
The verifier then acts as a maker-checker: it proves atomic claims against deterministic evidence, not model prose. If the panel rejects a claim, the decision is blocked.
Think of it like… a startup investment committee: each partner brings a lens, the average score drives the decision, and due diligence verifies the claims.
A board is defined in YAML: members, workflow phases, and contrarian ordering. runDebate runs phases serially and members within a phase in parallel. The Karpathy council variant adds blind peer review and chairman synthesis. verifyPanel runs coherence, faithfulness, and domain lenses with a hard-veto rule: one rejection rejects the whole panel.
# run a council gate before a mission
alembic run mission.json --council --offlinerunCouncilGate in @alembic/mission builds a default YAML board and calls runDebate. The CLI --council flag injects this gate before the swarm runs. A NO_GO verdict fails the run closed and writes a failed index.
Write a one-paragraph goal and ask the council offline: alembic forge "<goal>" --online (or offline) and inspect the SCOPE.md decision reasoning.