For universities · publishers · government · business

Detection your committee can actually verify.

Every Proofline result is Ed25519-signed and reproducible on your own hardware — deployable on-prem or air-gapped, sized to your case count. Request a trial and run it on your own documents; we'll reach out to set it up.

Verifiable, not asserted

Signed benchmark artifacts you recompute yourself — where competitors publish a headline accuracy with no way to check it.

Your text stays yours

On-prem & air-gapped deployment. We do not sell, monetize, or train on your content.

Runs on your CPUs

No GPU, no data center, no third-party API in the detection path — which matters for procurement and security review.

Compliance in progress

We're establishing the protocols to reach formal certification; current posture and documentation on request.

Step 1 of 3 · Who you are

Which best describes your organization?

We'll reach out shortly.

Thanks — your request is in. Someone from our team will contact you to set up your trial. In the meantime you can run a document through the tester.

Use cases

Where institutions put Proofline to work

Education

Academic-integrity review that survives an appeal

A signed, reproducible result a student — or a review board — can independently verify, instead of an opaque score. Matters where detectors are documented to misflag non-native writers.

Liang et al., Patterns 2023 · DOI 10.1016/j.patter.2023.100779
Publishers & corporations

Submission & content screening at volume

Batch-check manuscripts or contributed content through your own pipeline, with an audit trail you keep — no third-party service holding your text.

On-prem / VPC deployment · content never stored or monetized
Government

Air-gapped, procurement-ready

Runs entirely on your CPUs with no external calls in the detection path — the property a security review needs before anything is approved.

CPU-only · no GPU, no data center, no external API
Any institution

Non-English coverage the leaderboards skip

Full-document scoring in any language on the verifiable pipeline, where most public benchmarks report English only.

RAID benchmark, ACL 2024 · aclanthology.org/2024.acl-long.674
What buyers askTypical AI detectorProofline
Can we verify the accuracy claim?Headline number, no methodSigned artifact you recompute
Does our text leave our network?Sent to a hosted APIOn-prem / air-gapped option
Is there generative AI in the loop?Often another modelNone in the detection path
Reproducible on our own hardware?NoYes — Ed25519 + OpenTimestamps

Comparison of properties, not a head-to-head accuracy claim. Every Proofline figure ships with its signed source — see the proof page.