council-modelisted

# Council Mode ### Adversarial Parallel Comparison | Failure-Mode Surfacing | Not Consensus Council mode dispatches the **same task** to multiple model tiers and/or providers in parallel, then produces a diff report surfacing where each output disagrees. The goal is not to pick a winner — the goal is to learn where each model breaks down, so future routing avoids known failure modes. **Core principle:** > The orchestration layer's job is to surface where each model FAILS so you can route > around it — NOT to pick the "best" model. Build as adversarial comparison > infrastructure, not consensus/coordination. This is the opposite of traditional LLM council patterns (majority vote, consensus synthesis). Those patterns wash out signal. Council mode preserves disagreement because disagreement IS the signal. --- ## BRAIN CHECK — Run Before Every Council Call Brain root: resolved from `$CEREBRO_ROOT` env var or `~/.cerebro/profile.yaml`. 1. Verify advisor-mode is installed and provider-aware (`--local`, `--compare` available) 2. **Local fleet (optional):** if `fleet_primary` is set, verify Ollama is reachable at that host on port 11434. Cloud-only deployments (no local fleet) skip this check — all tiers route to cloud providers. 3. Check `{brain_root}/skills/advisor-mode/logs/daily_usage.md` if present — is remaining advisor budget enough for the council roster? 4. Load task context from the relevant department Brain if the task references one --- ## WHEN TO USE CO