Hosted API
Our infrastructure. Fastest to start, suitable for evaluation and non-sensitive workloads.
AI Models · Early access
Models that cite, and run where your data already lives.
General models are good at sounding right about financial text. We tune for a narrower target: extracting a claim and returning the exact passage it came from, or returning nothing at all.
The family
Three models, one behaviour.
Each is tuned to abstain rather than guess. In the domains we work in, a confident wrong answer costs more than a blank one.
Extract
Pulls structured facts out of filings, transcripts and contracts with passage-level citations attached to each field.
Compare
Diffs a document against its prior version and reports what materially changed in the language, not just the text.
Judge
Scores whether a generated claim is actually supported by its cited source. Used as an evaluation model, including on our own output.
Deployment
Three ways to run them.
Your cloud
Deployed into your own account so document content never leaves your boundary.
On premises
For firms that can't use external inference at all. Under evaluation with early-access partners.
Benchmarks not yet published.
Citation faithfulness and abstention rates are being measured against public baselines now. Numbers go on the research page when the evaluation is complete — not before.