Put a certified risk bound on any AI claim verifier.
eleata verify is a calibration + conformal-abstention layer over any claim verifier. It detects over-assertion (when a verifier says Supported/Refuted on irrelevant evidence) and emits a signed certificate with a conservative statistical bound — the guarantee the dashboards don't give you.
Open-source CLI · free tier · no credit card.
A score is not a guarantee.
Every hallucination-detector returns a number between 0 and 1. None of them tell you the bound that holds. A verifier blind to relevance over-asserts on distracting evidence — and you find out in production.
- ▸ Verifiers that ignore relevance over-assert 16–82% on irrelevant evidence.
- ▸ Competitors (Cleanlab, Vectara HHEM, Patronus, Galileo) sell scores and dashboards.
- ▸ Nobody else ships a signed, recomputable risk bound.
$ eleata certify --verifier my_nli --data labelled.jsonl fitting calibrator … ECE 0.018 clopper-pearson exact bound … ✓ certificate issued ε ≤ 0.0064 (acceptance, 95%, n=1200) bound is conservative by construction signed · recomputable · input-hash bound
How it works
Calibrate
Feed your verifier's raw scores + a labelled set. We fit a low-ECE calibrator and a Clopper-Pearson exact bound — conservative by construction (it never understates the risk).
Certify
Run over your independent units. eleata emits a signed certificate binding the bound ε to the exact inputs (hash) — recomputable by anyone, not a number you have to trust.
Guard
At inference time, the Guard composes the verifier and forces abstain when the certified bound can't be met — routing the hard cases to a human instead of asserting.
The certificate is the product.
A verifiable, recomputable certificate that a hollow product could not emit. The signal sells without a salesperson — and every certificate carries the bound, not a vibe.
Pricing
Start with the open-source CLI. Pay when you need the hosted API or a signed report.
- $299 one-time
- Signed PDF + bound + input hashes
- For a regulated review
Questions
Is this another hallucination score?
No. A score ranks; a certified bound guarantees. We report ε [lo, hi] with a confidence level, conservative by construction — not a single number you have to believe.
What's the math?
An instance of Learn-Then-Test (Angelopoulos et al.): a fixed grid + Bonferroni + exact Clopper-Pearson. The bound never understates the true risk.
Does it work on my verifier?
Yes — it wraps any claim verifier (NLI, LLM-judge, entity scanner). You bring the verifier; we calibrate and certify it.
Founding access?
A founder price for the hosted API while we onboard the first teams. The CLI stays free and open-source.