TextGrain Technical Report Explained

The central idea: dependence with an entropy budget

Sections 1 and 2 connect generated tokens to randomness derived from a secret key. Optimal transport chooses a joint distribution with prescribed marginals; KL regularization limits its departure from independence. The report relates that dependence to an average reduction in sampling entropy. This is a statistical construction, not a hidden string appended to an answer.

How block generation works

1. The model provides next-token probabilities after its sampling adjustments.
2. A secret key and preceding context divide the vocabulary into blocks.
3. A keyed cost table and entropy budget define a coupling over blocks and columns.
4. The selected column determines block sampling, followed by token sampling with the original relative probabilities inside that block.

What the detector reconstructs

Section 3.2 requires the matching tokenizer, block and column counts, context-window rule, secret key and pseudorandom construction. The detector reconstructs scores from observed tokens; it does not need the generating model\'s probabilities or entropy budget. Reading the paper, or obtaining future source code, does not supply another system\'s production key. See detector requirements.

Why the assumptions matter

The entropy bounds are averages under the stated construction, not identical guarantees for every fixed key or response. The report also distinguishes a requested budget from the achieved output of a finite numerical solver. Detection calibration assumes independence conditions and scores only the first occurrence of each distinct context. A deployed key and finite precision still require empirical calibration; a mathematical null distribution alone is not a universal real-world error rate.

A published report is not a released tool

Technical publication makes the method inspectable. Source-code availability, a runnable implementation, authorized production detection and removal verification remain separate milestones. Check open-source status and removal-method limits. This site offers character cleanup only; it does not use the report to produce watermark scores or removal certificates.

Sources and verification

Official announcement, access guide and technical report reviewed October 6, 2026. Availability below distinguishes announced rollout from verified access. This site is independent and does not offer a TextGrain detector.

• OpenAI EU text provenance announcement
• OpenAI provenance signals guide
• OpenAI content provenance API and text-access requirements
• textGrain technical report (October 5, 2026)