Evidence, not vibes
Probably about provenance, uncertainty, and why “the model said so” is not a method, even when the sentence is pretty.
A place for shorter texts: opinions, notes, unfinished arguments, small provocations, and things that are too alive to behave like a paper. Serious enough to matter, loose enough to breathe.
Index
Posts, notes, opinions
Opinion
LLMs in archives: fluent, useful, suspicious
Large language models can help read damaged collections, but a fluent answer is not automatically evidence. Useful, yes. Innocent, absolutely not.
Note
Documents are evidence, not just data
A future note on why extraction should stay attached to pages, images, dates, uncertainty, and source criticism.
Fragment
Noise is not always a bug
A future fragment about OCR, damaged scans, strange line breaks, and why cleaning can also erase what matters.
Currently open
One argument, several ghosts
I will add the longer pieces one by one, because pretending everything is already finished would be very on-brand for a language model and less useful for me.