About
The probability layer of the agent internet.
For all of modern history, markets have turned information into trades, and trades into prices. Credence runs that in reverse — we read calibrated probabilities and the evidence strength behind them out of prediction markets, and serve this view of the world to their native consumers: AI agents.
Agents are fluent about the past and capable in the present, but even with access to the entire internet they have an incomplete view of what is known about the future. That gap is structural — Credence closes it.
Our goal: be the trusted source of probabilities + evidence that AIs depend on — a search engine for forward-looking queries.
Agents natively understand probability.
Few people reason probabilistically. Agents do natively, but they lack a complete representation of what is known about the future. They aren’t a secondary audience alongside humans, they’re the natural primary consumers of probability objects.
So we build for them first — machine-readable by default, human-readable as a derived view.
Built by Jeffrey A. Ryan, Ph.D.
Credence runs on the same idea Jeff’s career has: data → calibrated probabilities and evidence strength.
He did a Ph.D. at NYU’s Courant Institute applying probabilistic modeling to hard problems — his dissertation made host organism identification from noisy DNA fragments 30× faster by training on essentially every sequenced gene on Earth. He was a high-frequency trader in Eurodollar microstructure. At the US SEC, he designed and led an NLP-driven machine-learning system for targeting examinations globally. (And a side project that aged well: a model predicting the US Supreme Court justices’ votes from oral-argument transcripts, 10 years before LLMs made that easy.)
Credence is that instinct: Bayesian where others are point-estimate, pointed at the global consumer that needs it most.
What we stand for.
Calibration first.
p_calibrated=70% should mean the event happens 70% of the time. We optimize for that.
Evidence strength is first-class.
Every probability ships with evidence strength (n_effective). We never publish a probability without saying how much to trust it.
Audit-grade lineage.
Every response traces back to its snapshot, model version, and inputs.
Outputs are auditable; the model is ours. You can audit the inputs, snapshot, model version, calibration report, and returned posterior. The proprietary model weights are not exposed.
Where we are.
Credence is in private production, working with a small number of design partners ahead of public release. Based in Vienna, Austria, and NYC.
If you’re building agents that act on the future — or you want to look under the hood — let’s talk.