Multi Model Answer Checker

Goal. Your harness runs one model, and a single model can be confidently wrong. This recipe cross-checks your host model’s answer against two or three independent models with Fusion, right inside the same Claude Code or Codex session, so a contested or incorrect claim shows up instead of hiding behind one model’s confidence. When to reach for it. A design call, a security claim, a tricky algorithm, a “which library” decision, anything where being wrong is expensive. For an ordinary question you already trust, you do not need it.

How It Works

A single model has blind spots, and it states a wrong answer with the same fluency as a right one. You cannot tell which from the answer alone. Fusion asks several models the same question at once and synthesizes their answers. The result has two parts: each model’s own take, and a synthesis combining them, which flags material disagreements it cannot resolve. Reading the participants next to each other is what tells you whether the question is settled or contested, so you can compare that against what your harness’s own model told you. Because it runs through the plugin, all of this happens inside your existing session. You do not copy the question into another tool or lose your context; you get an independent panel in the same place you are already working.

Prerequisites

Nito installed and signed in in Claude Code or Codex. See Install in Claude Code or Install in Codex.
A paid plan (Fusion is a paid feature). The single-model check below works on Free.

Step 1: Get your host model’s answer

Ask your harness as you normally would, for example:

Should this service use optimistic or pessimistic locking for its inventory counter?

Your host model answers. Keep that answer in mind for the comparison.

Step 2: Cross-check with a panel

Send the same question to two or three independent models with Fusion. Name models from different families so the panel is genuinely independent, not three variations of one lineage:

Claude Code
Codex

/nito:fusion openai/gpt-5.5 z-ai/glm-5.2 google/gemini-3.5-flash -- Should a high-contention inventory counter use optimistic or pessimistic locking? Answer in 3 sentences with the deciding tradeoff.

$nito:fusion openai/gpt-5.5 z-ai/glm-5.2 google/gemini-3.5-flash -- Should a high-contention inventory counter use optimistic or pessimistic locking? Answer in 3 sentences with the deciding tradeoff.

Step 3: Read the synthesis against your host model

Fusion returns each model’s answer plus a synthesis. Abridged and representative:

Nito · Hosted Fusion · 3 participants · judge: openai/gpt-5.5 · status: complete

1. openai/gpt-5.5        leans optimistic + retry; contention is bursty, not sustained
2. z-ai/glm-5.2          pessimistic; a hot counter serializes anyway
3. google/gemini-3.5     optimistic with a fallback to a row lock under contention

Synthesis
   Recommendation: optimistic with bounded retry, fall back to a row lock if
   retries exceed a threshold. Where they disagree: how hot the counter really
   is. If one row is updated constantly, pessimistic wins; if contention is
   bursty, optimistic wins.

Now compare against your host model:

The panel agrees with your host model: you have independent corroboration. Move on.
The panel splits: the question is genuinely contested. Read the participants to see on what, so you can decide deliberately instead of trusting one voice.
The panel disagrees with your host model: the strongest signal. Read why, and re-examine the host’s answer before you act on it.

Verify a Specific Claim

To check one concrete statement rather than re-ask the whole question, paste the claim and ask a single independent model whether it holds. This works on Free:

Claude Code
Codex

/nito:ask --model z-ai/glm-5.2 Is this correct? "A UUID v4 is guaranteed unique across systems." Answer yes or no and explain in one sentence.

$nito:ask --model z-ai/glm-5.2 Is this correct? "A UUID v4 is guaranteed unique across systems." Answer yes or no and explain in one sentence.

Automate It From the Terminal

To cross-check answers outside a chat, or in a loop over several claims, run the same check from the shell. This script takes a claim and reports a panel verdict:

#!/usr/bin/env bash
# second-opinion.sh "claim to verify"
set -euo pipefail

claim="${1:?usage: second-opinion.sh \"claim to verify\"}"

nito fusion \
  --model openai/gpt-5.5 --model z-ai/glm-5.2 \
  --web-search off \
  "Is this claim correct? \"$claim\" Each of you answer yes or no with a one-sentence reason."

./second-opinion.sh "Postgres SERIAL columns are gap-free."

Variations

Fact-check with sources. For a claim about the current state of the world, let the check use web search so it returns citations:

Claude Code
Codex

/nito:ask --web-search on What is the current stable release of PostgreSQL, with a source URL?

$nito:ask --web-search on What is the current stable release of PostgreSQL, with a source URL?

Keep the check private. If the claim involves proprietary details, verify it with private so your harness never sends it upstream.

Troubleshooting

What you see	Why	Fix
”Fusion requires a paid plan”	You are on Free	Use the single-model `ask` check in “Verify a Specific Claim”
The panel just echoes the host model	The models chosen are too similar	Pick models from different families (a frontier model plus an open-weight one)
A privacy-level error	A Confidential model was mixed with a lower level	Keep Confidential participants in a Confidential-only Fusion. See Fusion
Web search returns stale data	The search backend declined the query	Rephrase as a clearly time-sensitive question, or try again

Where to Go Next

Fusion command

Full syntax, how to pick participants, and how privacy levels mix.

Private Code Review Pipeline

Multi-model review of your changes on Private-tier models.

Choosing a Model

How to pick strong, complementary models for a panel.

Web Search

How the web-search modes behave when you fact-check.

​How It Works

​Prerequisites

​Step 1: Get your host model’s answer

​Step 2: Cross-check with a panel

​Step 3: Read the synthesis against your host model

​Verify a Specific Claim

​Automate It From the Terminal

​Variations

​Troubleshooting

​Where to Go Next