Multi-Model Consensus
For CRITICAL decisions (medical / financial / engineering / general), query 3 best-in-domain models INDEPENDENTLY in parallel, then a synthesizer (Claude Opus) compares answers, flags agreements/disagreements, and produces one verified answer with a confidence level and dissent notes.
When to use
- High-stakes questions where a single model's error is costly.
- "Get a second and third opinion and reconcile them."
- Engineering checks, medical reasoning, financial assessments.
Quick start (load env first)
bash
bash -c '. /data/.openclaw/secrets/env.sh; python3 skills/multi-model-consensus/consensus.py ask "<question>" --domain engineering'
Command
| Command | Purpose | |---|---| |ask "<question>" [--domain D] [--models N] | Query the domain's 3 models in parallel, then synthesize. --models N limits to first N. |Domains: medical, financial, engineering, general (default general).
Domain model sets (native APIs, verified 2026-05-31)
| domain | models | |---|---| | medical | claude-opus-4-8 · gpt-5.6-sol · gemini-3.1-pro-preview | | financial | claude-opus-4-8 · gpt-5.6-sol · grok-4.5 | | engineering | claude-opus-4-8 · gpt-5.6-sol · deepseek-chat (V4) | | general | claude-opus-4-8 · gpt-5.6-sol · gemini-3.1-pro-preview |Synthesizer: claude-opus-4-8 (falls back to gpt-5.6-sol if Opus fails).
How it works
1. Each model gets the SAME question + a domain-appropriate system prompt, queried independently (no cross-talk) viaconcurrent.futures (parallel, uses the 4 cores). Wall time ≈ slowest model, not the sum.
2. The synthesizer receives all answers and emits: Agreements, Disagreements/Dissent (named per model), Final Verified Answer, and a Confidence: HIGH/MEDIUM/LOW line.
3. Plain-text only — system prompts forbid LaTeX so all output is clean (safe for Arabic-facing use).
4. Resilient: if a model fails it's marked FAILED; synthesis proceeds with the survivors (and reports the count).Output sections
- INDEPENDENT ANSWERS (per model + elapsed time)
- SYNTHESIS (agreements / dissent / final answer / confidence)
Example
bash
... consensus.py ask "Patient on warfarin needs an antibiotic for a UTI — which classes raise bleeding risk most?" --domain medical
... consensus.py ask "Required midspan Mu for a 6 m simply-supported beam, factored UDL 30 kN/m?" --domain engineering
Tested (2026-05-31)
- Engineering question (RC beam design moment): all 3 models (opus-4-8, gpt-5.6-sol, deepseek-v4) queried in parallel (~6s wall vs ~14s sequential). Synthesizer correctly found unanimous agreement (Mu = 135 kN·m, ULS flexure), reported only detail-level dissent, and rated confidence HIGH. Output fully LaTeX-free.