tokonomix-consensus

Tokonomix consensus skill

When a wrong answer is expensive — legal/compliance interpretation, a security-relevant code change, a fact that's about to ship in a customer-facing document — don't trust a single LLM. Ask two or three frontier models in parallel and have an independent judge synthesize the result — by default the judge is selected from the same multi-vendor pool (Anthropic / OpenAI / Google / open-weight), is a different model than the proposers it scores, and is never asked to score its own answer, so the synthesis step doesn't inherit one vendor's blind spot. You can pin the judge or run a multi-judge panel (prefer a cross-family one).

Its real edge is cross-vendor diversity: Claude, GPT and Gemini answer the same prompt side by side, then a judge synthesizes them. Your host agent's own review tooling or sub-agents can only give you one vendor's perspective — a cross-vendor check is the one thing they cannot reproduce. Reach for it when the cost of a wrong answer exceeds the few cents the check costs.

What consensus does — and does not — give you

Consensus reduces single-model error and surfaces disagreement a single model would hide. It does not guarantee correctness. Frontier models share heavily overlapping training data, so on a shared blind spot — post-cutoff facts, niche domains — they can agree and be uniformly wrong, and the judge will hand you a confident wrong answer with a clean "blind spots" section that found nothing. Treat agreement as a strong signal, not proof; for out-of-distribution or post-cutoff facts, still verify against a primary source.

Read it as a recall amplifier that feeds your judgment — not a truth oracle or a blind merge gate. Its strength is recall: with several different vendors, you only need one to catch the timing side-channel / the missed edge case, and it gets surfaced. The cost is precision: more models means more flags, some of which are nitpick, so a human adjudicates. That trade is worth it for rare, high-asymmetric-cost decisions (auth, migrations, GDPR, money) where one real catch pays for a lot of false alarms — and it is not worth it for routine work, where it adds cost and noise. Agreement raises confidence, not correctness.