fabius-conciliumlisted

fabius's cross-model deliberation layer — convene a COUNCIL of heterogeneous models on ONE question, then aggregate their answers into one evidence-aware synthesis: each model answers independently (first opinions) → each ranks ALL anonymized answers blind, its own included, while the backend removes its self-score (anonymized peer-review) → a chairman synthesizes the field. This is ensemble epistemics — it uses model diversity to expose possible single-model error and bias; whether it improves an answer must be measured for the task. Use when the user says "council" / "llm-council" / "ask several models" / "panel of models", or when a high-stakes question has already survived N samples of the single strongest model and they failed the SAME way — correlated error, not mere disagreement, is what earns at most 3N+1 completion calls including retries (M10). Distinct from fabius-cohors (which splits the WORK across task-specialist agents); concilium aggregates one ANSWER across whole models.
shear559/fabius · ★ 2 · AI & Automation · score 64

Install: claude install-skill shear559/fabius

# Fabius Concilium — convene the council, synthesize one answer *Concilium* — the summoned assembly, the council called to deliberate before the decision is taken. fabius's whole identity is *one set of rules above every model*; concilium seats several models, lets them answer and judge each other blind, and has a chairman fuse the field. Its aim is to expose disagreement and reduce idiosyncratic error; it does not guarantee superiority over the best single seat, so measure that on the task. Patterned on karpathy's *llm-council*. ## The lean gate first — does this need a council at all? (`fabius-parcus`) A council is the **heaviest** thing in this whole system on cost and latency. With N configured seats, M surviving first opinions, and R malformed-ballot retries, it makes exactly **N + M + R + 1** completion calls: N first-opinion attempts, M initial reviews, R retries, and one chairman. Because `0 ≤ M ≤ N` and `0 ≤ R ≤ M`, reserve at most **3N + 1** calls; the clean all-live/no-retry path is `2N + 1`. So the question comes before the council does: - **Convene when** the answer is high-stakes and a wrong one is expensive (a design/architecture call, a medical/legal/financial *analysis*, a contested factual claim, a research synthesis) **and** rung one below has already failed the *same* way, **or** the user explicitly asks f