engkimo

This is a great example of why a PDF’s text layer is not necessarily the same as its visible content.

I’d be cautious about making “non-black means invisible” the general rule. A possible middle ground between raw text extraction and full OCR would be to render the page, map each extracted glyph’s bounding box back to the rendered pixels, and discard glyphs that have almost no contrast with their local background.

That would preserve the exact embedded text for visible characters while using rasterization only to determine visibility.

Has anyone tried this kind of hybrid approach?

The raw runs are useful, but plausibility alone seems like a weak way to distinguish memorization from hallucination.

I’d be curious to see a control set using random name pairs and the same number of trials. Then you could compare how often rare names, exact phrases, dates, or other specific details recur across fresh sessions.

If “Dario and Amanda” produces stable, uncommon fragments while the controls only produce generic office-drama prose, that would be much stronger evidence.

Have you tried running that comparison?