The card claimed identifiers are validated by checksum, as 'a correctness guarantee general LLMs lack'. No repository in the OpenNER family ships validation code, and a ModernBERT encoder cannot run a check digit, so the claim was false for every model in the family.
The held-out F1 is now explained rather than headlined: train and test come from one generator, and two models from the same pipeline (filingtag 0.6444, ibandetect 0.9488) show it can report below 1.0, so a 1.0 means the generated task was trivially separable.
Also removes a pointer to an OpenNER benchmark that is not published, and adds a 'Trained on' line naming the actual domain.
