Stress-test one AI detector with formatting changes

This optional browser demo compares one open model's score for your text as written with scores for formatting variants of the same words. A difference can show formatting sensitivity; no score or comparison establishes authorship, accuracy, or misconduct.

Why the earlier demonstration was withdrawn

An earlier Draft Record experiment described five passages as historical controls without traceable titles or source URLs, and its preprocessing changed paragraph structure before the baseline score. Those data could not support the published false-positive claim, so the claim and table were removed rather than relabelled as evidence.

What this bounded stress test does

The 120-word minimum is Draft Record's interface guardrail for this comparison, not a vendor accuracy threshold. The model is the public onnx-community/tmr-ai-text-detector-ONNX model. Its model card names English-language and domain limits; a local run does not validate those limits or make the model suitable for a decision about a person.

Run the formatting comparison

Hugging Face's pinned model file listing shows the q8 ONNX weight file as 126 MB. The first run downloads that quantised model from Hugging Face. The model runs in this page; text entered in the box is not included in the request for the model and is not sent to Draft Record analytics.

0 words

This is a model stress test, not an authorship test or accuracy benchmark.

What this cannot tell you

  • Revision-save identifiers do not reconstruct a complete or ordered writing history.
  • The markers do not identify an author or establish when text was written.
  • A document with few or no markers is not evidence of wrongdoing; software and conversion paths differ.
  • This is not an AI-detection result and cannot determine whether AI assistance was used.