- RAID · ACL 2024
- Ed25519 · FIPS 186‑5
- Bitcoin anchored
We built an AI-text detector without the AI inside.
That isn't a marketing line, it's a structural one. Almost every detector in use asks a language model how likely a language model was to have produced the text. That is circular, and the circularity is why they fail.
You cannot establish ground truth about machine-generated text using a machine that generates text.
So we took the generative model out of the measurement path entirely. Proofline measures structure — arithmetic computed directly from the characters. No inference, no learning model for meaning, no generative technology. Same text, same numbers, any machine, forever.
Measured on 2,000 human and 2,000 machine-written English student essays, held out from the reference set — Quick Check mode, English:
- ROC-AUC 0.9872 — see the signed DMITVA runs →
- 95.5% detected at a 5% false-positive rate (95% CI 94.6-96.4)
- 78.8% detected at a 1% false-positive rate
Both modes run at the same 5% false-positive setting — about 95 in 100 real documents cleared, only 5 wrongly flagged.
Accuracy isn’t about being right — it’s about not being wrong. That is why those false-positive rates matter to us more than any headline: a wrong flag isn’t a lower score, it’s a person.
So the number we hold ourselves to is the false-flag rate. Absolute Verification runs the whole document through Document Review and its qualifying section through Quick Check (English); when both agree, fewer than 2 in 100 real documents are flagged by mistake — across 1,858 documents.
We also measure one of the hardest, least-served cases: AI-generated code — something teams increasingly need to know. On the RAID code benchmark (Document Review mode) Proofline holds ROC-AUC ≈0.88, signed and recomputable like everything else. See the signed code runs →
What we need is the world's assessment and choice of adoption.