I'm doing it in stages. I'm not using AI to fill document gaps per se. I want to recover the original, not let it hallucinate content.
AI made one pass to convert images to text. Now I'm making another pass to try to get the AI to fix up alignment issues. Future passes will fix spelling and English language issues, any known semantic problems. Final pass will convert to markdown from a recovered "LPT capture" since that's a tractable format for both human programmers and AI. The overall sub-project: I'm using AI to do OCR that preserves ASCII art and text formatter layout for this document. Ultimately this document becomes part of a RAG framework for model T development. Tesseract or something like it was probably the source of the original OCR that's embedded in the PDF. It's something, but having messed with trying to correct assembly listings and ASCII art using it, it's not great. Maybe that was a skill issue on my part, but the AI is doing a very good job in comparison to what I was getting. As you say I suspect the original OCR problems are from skew, noise, the dot matrix font, etc. As to crowd-sourcing fixes: reminds me, long ago I launched a re-pub of Thinking Forth to bring it back into publication. That job would have been so much easier today. Anyway it's coming along and I think it's good enough that we won't need to crowd source a re-pub. Part of this project is just to give me something to build skill with AI tools. All of this is new to me. Building a really good corpus of documentation including Model T BASIC programming best practices, advanced techniques is where the community might want to start thinking about what's important. AI or not, having a corpus of information for ramping up Model T programmers is valuable no matter what. BTW, I don't doubt I am upsetting some people with this AI stuff. No one has to use it that doesn't want to and I am aware of all the feelings and experience some myself. We all approach the intersection of this hobby and modern technology a little differently and I respect other positions. I talk about it here just to let people know what I'm working on, but everyone is welcome to do things their way and no one has to buy into it that doesn't want to. FWIW I am keeping AI at arm's length. I will only use AI that I can control. I don't intend to set it loose to pollute the world. -- John. On Tue, Oct 20, 2026 at 1:24 PM B9 <[email protected]> wrote: > On October 20, 2026 11:05:33 AM PDT, "John R. Hogerhuis" <[email protected]> > wrote: > >On Tue, Oct 20, 2026 at 10:11 AM > > >So barring some new discovery we will have to infer what is missing. > > > >I have a custom perl runner and agent pushing my prompt and the pages > >through gemma4-31b. It's taking a while but the results are pretty good > and > >at least now it's doing it without my intervention. With OpenClaw it kept > >stopping every 3-5 pages. > > > You are (a) just trying to OCR the document? > > Or are you (b) using AI to fill the gaps of those cut off pages you > mentioned? > > If the former, (a), using tesseract may be easier since this document is > typed, not handwritten. Ocrmypdf (which uses tesseract) gave me a > reasonable text approximation. What errors were you seeing? If you need to > tweak the OCR, straightening the page helps enormously. Or, you can add a > custom dictionary. And if there is a persistent misconstrual, training > tesseract on a specific typeface isn't as hard as it used to be. > > If the latter, (b), maybe we should go full retro and collaborate on this > list to reconstruct the missing parts. > > --b9 > >
