How CrossCheck Works

User Guide

From your question to a verified answer — in six steps

Step 1

Choose your models

Two ways to get started

BYOK (bring your own keys): Add API keys for Claude, GPT, Gemini, DeepSeek, Mistral — your keys, $0 markup. Keys are encrypted and never stored in plain text.

Managed: Skip setup — just pick which models to use. We handle the AI infrastructure. No API keys needed.

BYOK = full control, zero fee · Managed = zero setup

API Keys settings — Anthropic and OpenAI connected, Google key being added, Mistral AI available

Settings → API Keys

Step 2

Ask your question

Paste any prompt

Type or paste the same question you'd ask ChatGPT. Pick which models to verify with. Choose Quick or Standard analysis depth — longer content takes more time, but you see progress in real time.

Analysis time scales with content length and depth

Chat input with model selector and CrossCheck button

Chat → New verification

Step 3

Models analyze independently

Phase 1 — blind peer review

Each model answers your question separately. No model sees another's answer — independent analysis, like blind peer review. You see each response appear in real time with timing and cost.

3 models · independent · no influence between them

Phase 1 · Independent Analysis

Step 4

Errors surface through debate

Phase 2 — multi-agent critique

Models critique each other's answers across two rounds. Disagreements are flagged by severity — from minor nuances to critical errors like fabricated statistics or misleading benchmarks. Each model's position is shown side by side.

14 disagreements found · CRITICAL errors flagged

Phase 2 — CRITICAL disagreement with three model positions shown side by side

Phase 2 · Multi-Agent Critique

Step 5

Get your Confidence Report

Phase 3 — synthesis

All findings are synthesized into a confidence score with actionable recommendations. See what all models agree on, where they diverge, and what you should do about it. Points of consensus are highlighted in green, unresolved disputes in amber.

62% confidence · 41 claims verified · 3 models

Phase 3 · Synthesis

Step 6

See what was caught

Concrete harm prevented

The "What CrossCheck caught" summary shows exactly what would have gone wrong without cross-verification — misleading benchmarks, fabricated confidence in broken tracking, tools that increase costs for small businesses despite claiming otherwise.

Below that, the full synthesis report breaks down every verified claim, consensus point, and disputed finding in detail — your complete audit trail.

Unresolved disputes? Run Deep Verify — it adds web sources and more models to break the tie. Export the full report as PDF to share with your team or clients.

Export PDF · Deep Verify · share with clients

What CrossCheck caught — concrete harms prevented, Deep Verify CTA

Results · What CrossCheck caught

Unlike single-model AI tools, CrossCheck uses adversarial analysis — models challenge each other to surface errors you'd never catch alone.

Try it free →