What should survive when AI compresses a document? My attempt at a reader with reversible levels of detail

I work on human–AI alignment, and I keep coming back to a reading problem: it is easy to accept a compact explanation before checking what was left out. Unceta and colleagues describe a related concern as "epistemic debt": a gap between the output we use and the understanding we can defend. Their Nature Machine Intelligence comment is here: https://doi.org/10.1038/s42256-026-01294-w I've been exploring an interface response in TextFold, an open-source reader I build. The older idea behind it is Ted Nelson's stretchtext from 1967, where the reader could choose how much of a passage to expose. The useful design distinction, for me, is between selecting source wording and generating a shorter explanation. TextFold keeps those separate: Full contains the original paragraph. Brief selects one to three of its sentences verbatim. Key, Takeaway and Topic are AI-written layers with progressively less detail. You can reopen an individual paragraph while the rest stays folded. The levels are prepared first, so changing depth does not trigger another model call. There is a trade-off even in the verbatim layer: keeping the wording doesn't guarantee that the selected sentences preserve every caveat. An AI-written layer can introduce further errors. Keeping the source accessible makes checking possible; it doesn't prove that readers will check it. I haven't established that this improves comprehension or reduces cognitive load. The video shows the interaction and its Stretchtext background, then examples of web pages, PDF and Word documents. I'm interested in whether the interface helps readers notice when they need the fuller argument, and how to measure that against a normal document plus a summary. Source and setup (Apache 2.0): https://github.com/chenxiachan/textfold It can use a local Ollama model or a hosted provider; with a hosted provider, the text goes to that provider.