Home › Case studies › Case 04
Case 04 · Silent updates
GPTZero, Grammarly, and Trinka silently update. Their scores are not reproducible.
Same text, different day, different score — evidence value goes to zero.
The failure · on the record
Every commercial detector using an LLM or a hosted classifier reserves the right to update silently. LLM-based detectors are non-deterministic even at temperature=0 due to floating-point ordering across GPU infrastructure (arXiv 2601.19934). GPTZero, Grammarly, and Trinka have all shipped silent engine updates in 2025–26. The score you got yesterday is not the score you can produce in court.
The numbers
| Deterministic? | Reproducible in court? | Signed audit log? | |
|---|---|---|---|
| GPTZero | No (LLM inference drift) | No | No |
| Grammarly | No (silent updates) | No | No |
| Turnitin | No | No (removed from many campuses) | No |
| Binoculars | Partial (open weights) | Only if weights pinned | No |
| Truth-in-AI | Yes (bit-exact) | Yes (5 years, 50 years) | Yes (Ed25519 per result) |
What we do differently
Every scoring event returns an Ed25519 signature covering the score, the text SHA-256, the detector version, and a timestamp. Re-run any prior signed result and get the exact same bytes back.
Public replay endpoint: paste a prior signature, get bit-for-bit the same score verified.
Proof
The replay endpoint. Replay endpoint →
Try it yourself
Score once, replay months later.
# 1. score today
S=$(curl -s -X POST https://api.trigeochiral.com/detect \
-H 'Authorization: Bearer $KEY' \
-d '{"text":"any"}' | jq -r .signature)
# 2. replay in 5 years — same bytes
curl -s https://api.trigeochiral.com/replay/$SBounty · put money where the claim is
$1,000
Prove we return a different signed score for the same text across two dates. Failed audits published.
Sources