When you eventually support natural language prompting please enable the user to instruct the model not guess or infer what the original author intended. (I see a similar comment was submitted back in Aug 2025 and I haven’t found any indication of progress in the current 10/2025 model) A simple [illegible] would be great. Even better if you would aspire to more common archival notations such as the following. Yeah, a big ask, but, c’mon, you can do it! 
Situation Transcription notation|
One completely unreadable word [illegible]
Approximately three unreadable words [illegible, ~3 words]
Unreadable portion of a sentence [illegible passage]
Probable word, but uncertain [France?]
- or even just -
Word supplied by transcriber, not visible in original [France]
To end on a more positive note: I REALLY like the ability to call up other models to do a follow-up transcription and compare the results side by side. This feature by itself will keep me coming back. Nice work!
p.s. It would be good if, when exporting to text (thank you!), you indicate that you are exporting (as I have verified} as plain-text (.txt) document encoded in UTF-8 without a BOM, using Unix-style line endings., just so I always know nothing has changed.