Every commercial AI detector fails Daubert. Ours doesn’t.
Federal Rule of Evidence 702 requires known error rate + reproducibility. Only signed, deterministic detectors clear that bar.
Federal Rule of Evidence 702 sets the Daubert reliability standard: testable methodology, peer-reviewed, known error rate, general acceptance in the field. Proposed FRE 707 (2026) applies these criteria specifically to machine-generated evidence. LLM detectors that update silently do not have a known error rate at any single point in time — their score is a moving target. We publish signed benchmarks that pin our error rate to a specific detector version, verifiable years later.
| Daubert factor | GPTZero | Turnitin | Binoculars | Truth-in-AI |
|---|---|---|---|---|
| Testable methodology | Partial | No | Yes | Yes (signed artifact) |
| Peer-reviewed | No | No | Yes | Yes (RAID benchmark) |
| Known error rate | No (drifts) | No (drifts) | Yes (pinned weights) | Yes (signed, per-config) |
| Reproducibility | No | No | Only if weights pinned | Yes (Ed25519 replay) |
| General acceptance | Commercial | Commercial | Academic | New — growing |
Every signed result includes an admissibility packet: the detector’s signed benchmark AUC, its version SHA, the input text’s SHA-256, an Ed25519 signature over all of the above, and a Bitcoin OpenTimestamps anchor.
Enough for a paralegal to file. Enough for a Daubert hearing.
curl -X POST https://api.trigeochiral.com/detect \
-H 'Authorization: Bearer $KEY' \
-d '{"text":"...","packet":"admissibility"}'