Vela AI
The lab

Research first, product second.

37 researchers, 24 published papers, and one bet: that auditable reasoning wins in the industries that cannot afford a confident mistake.

Fig. 01 - How traces work

A decision you can hold to the light.

Every Vela answer goes through four steps. The output you read is the last of them.

01
Ground

The model is given your policies, contracts, and historical decisions - nothing else. Sources are chunked, indexed, and attributable.

02
Reason

Vela produces a step-by-step chain, each step constrained to cite the exact line or record it used. No leap may go uncited.

03
Verify

A separate verifier re-runs the chain against the sources and scores it. Unverifiable chains are flagged, not shipped.

04
Export

The full trace exports as an immutable audit document - timestamped, hashed, and attachable to the case record.

Fig. 02 - Selected publications

We publish what we claim.

Traces, Not Vibes: Verifiable Reasoning for Regulated Decision-MakingICLR 2026
Citation-Grounded Claims: Making Language Models Point at Their SourcesACL 2025
Human Oversight Without The BottleneckNeurIPS 2025
Reasoning Under Audit: A Framework for Reproducible Model DecisionsarXiv 2025
Fig. 03 - By the numbers

Quiet, measurable progress.

24

papers published

6

industry pilots

98.7%

trace completeness

37

PhD researchers

Partner with the lab

Run a pilot on one of your hardest decisions.

Pilot slots include access to the research team and your data never leaves your boundary.