konsiliumlisted

# /konsilium — the Council Question/decision: **$ARGUMENTS** The council gives **5 independent perspectives + mutual critique + synthesis** (karpathy/llm-council mechanism: independent opinions → anonymous peer review → chairman synthesizes). It does NOT compute data — it pressure-tests decisions and conclusions. So it's useful where **the cost of error is high and the answer does NOT follow unambiguously from the analysis**. > **Hard rule.** The council decides WHAT to test and whether to ACT; your harness (tests / backtest / live data) computes the RESULT. Pairing: council raises a hypothesis / surfaces a risk → you run an honest test → council helps decide whether to commit. Neither works alone. --- ## PHASE 0 — relevance gate (mandatory, before convening) Before spending the council — check it's actually its job. **✅ Convene:** - Go/no-go before real consequences (ship this config to production, raise the size, change the parameter). - Whether to trust an analysis / critique of methodology (sample, look-ahead, regime-dependence, i.i.d. assumptions). - Choosing between competing approaches with a real tradeoff. - Pressure-testing a conclusion you/the user are attached to. - Designing a risk control (circuit-breaker, exposure limit, safety switch). **❌ Don't convene — do it directly:** - Real log/data analysis — the council has no data access, compute it yourself. - A factual lookup (how an API works, a code detail) — that's lookup, not debate. - A mechanical code e