Blind panel / verified synthesis

Five models. One verdict you can trust.

A blind panel of models + a judge for decisions where being wrong is expensive. fusion verifies before it synthesizes and counts agreement by model family, not by head.

Inspect the method Install fusion

hover a panelist / trace its path

And still wrong.

Models share training priors. Three matching answers can be one correlated error repeated three times. Naive averaging launders that error into false confidence.

  1. 01

    Blind, non-voting judge

    The judge writes its own answer before reading the panel. It sets a quality floor without anchoring and never counts itself as another vote.

  2. 02

    Verify before synthesis

    Load-bearing checkable claims enter a ledger and get checked with real tools, even when every panelist agrees.

  3. 03

    Consensus by family

    Two wrappers around the same model family remain one opinion. Agreement is reported across families, never as a flattering raw headcount.

  4. 04

    Calibrated confidence

    Each answer names what would change its mind. The judge weighs verified arguments and live disagreements, not volume.

  5. 05

    Neutralized voice. Honest outcome.

    A normalizer strips stylistic fingerprints before judging. When the families genuinely split, the verdict is “inconclusive,” not a manufactured winner.

From question to inspectable verdict.

The mechanism is sequential by design. The panel stays blind, the judge commits first, and checkable facts cross a verification boundary before synthesis.

01

Classify

Skip trivial facts. Flag shared freshness blind spots.

02

Prompt blind

Send one identical prompt to every panelist in parallel.

03

Judge blind

Commit the judge’s independent answer before exposure.

04

Normalize

Anonymize voices, build the claim ledger, verify facts.

05

Synthesize

Count families, floor-check, then cross-family red-team.

06

Verdict

Return confidence, evidence, disagreements, and risks.

Your models. One method.

fusion is a skill plus a zero-dependency setup script. It detects the model CLIs already on your PATH and writes a local config.

Claude · GPT / Codex · Gemini · Grok
MiniMax · OpenCode · Ollama · any CLI

fusion / setup
# Detect installed model CLIs
$ node setup/detect.mjs --write
wrote fusion.config.jsonc

# Ask a decision-grade question
› /fusion <your hard question>
Bring your own model access. Tokens stay in your environment and are never hard-coded into the config.

Honest limits.

It is a decision aid, not an oracle. fusion raises the floor and surfaces the real tradeoff. It cannot guarantee the right answer.

Diversity and cost are real. You need at least three model families, and a panel means N model calls. Use it when the decision is worth more than one answer.

Trust the mechanism, not the chorus.

Read the method. Inspect the prompts. Bring the model CLIs you already use.

Open fusion on GitHub