Deep Audit — three models, cross-verified
LiveRun a contract past Claude, GPT and Gemini in parallel, then reconcile the three readings into one report.
- ✓Three model vendors reading the same contract independently
- ✓A reconciliation pass that merges them into one report
- ✓Per-upload choice, so the cost is spent where it matters
- ✓Roughly three times the token cost of a standard analysis
What it is.
Switch a single upload into deep mode and the analysis does not go to one model. Claude, OpenAI and Google each read the contract in parallel, and a reconciler pass merges the three readings into one report — so a finding that only one model saw is visible as exactly that, rather than presented with the same confidence as one all three agreed on.
It costs roughly three times the tokens of a standard analysis and takes longer, which is why it is a per-upload choice rather than the default. Use it on the contracts where being wrong is expensive.
Three steps.
End to end.
A per-upload setting, not an account-wide one. Standard analysis stays the default.
Claude, OpenAI and Google each analyse the contract independently, section by section.
A merging pass produces a single cross-verified report rather than three you have to compare yourself.
What you get.
- ✓Three model vendors reading the same contract independently
- ✓A reconciliation pass that merges them into one report
- ✓Per-upload choice, so the cost is spent where it matters
- ✓Roughly three times the token cost of a standard analysis
- ✓All three vendors — Anthropic, OpenAI, Google — disclosed in the privacy policy's sub-processor list, and all three hard-disabled in air-gapped deployments
- ✓On the standard path the Claude model is picked per request: Sonnet 5 by default, Opus 5 for long-input review, and the top tier only for deep research or a retry after a low-confidence answer
Quick answers.
It is more cross-checked, which is not the same claim. Three independent readings surface disagreements a single pass would have hidden; we are not going to put an accuracy percentage on it, because we have not measured one.
Yes. The Anthropic, OpenAI and Google calls on this path each write a cost row with the model that answered, so a deep run shows as roughly three times a standard one in the usage view. What is still outside that log is embeddings and meeting audio — see Cost transparency.
More in Review & Risk Analysis.
One card per clause type with the quote behind each finding, a playbook for the negotiation, and a renewal pack built from the record.
Three-tier extraction cascade so PDF, DOCX, and scanned bilingual contracts all work — one accept list for every upload door, and a report page that says which step the analysis is on.
Red, amber, or green for every clause, with an explanation and confidence score.
One number, 0-100, summarising your organisation's legal position across four dimensions.