wtf is pageledger?

PageLedger is a command line tool for tracking exactly what happened when you run OCR or document extraction on a large batch of pages. It was built for people who need to answer questions months later, such as which software engine processed a specific page, how much a run cost, which pages were too noisy to trust, and which ones still need a human to look at them. You bring your own extraction engine, Tesseract, Docling, Marker, or a cloud vision language model, and PageLedger wraps the run, enforces budgets on pages, tokens, and dollars, and writes everything it learns as plain files on disk that you can search, cite, or use to reconstruct exactly how a result was produced. There is no server and no database involved. The project began as part of a Soviet census digitization effort, where getting a model to return structured JSON was only the start of the real work of trusting that output. It is aimed at digital humanities labs, archives, historians, and anyone else who has to defend their extraction methodology later. To use it, you install it with pip and run a command pointing at a scanned PDF, using a built in adapter for plain text, born digital PDFs, or scanned PDF OCR. Each run produces a folder holding the extracted text alongside a manifest, per page evidence of where the text came from, quality warnings, OCR confidence scores, cost records, and a queue of pages flagged for review. If some pages come back weak, you can rerun just those pages through a stronger engine and compare the two runs directly. Under the hood, PageLedger classifies each page's structure, tracks quality signals such as low confidence text or likely mismatched historical spelling, grades each page from A to F with the reasoning behind the grade, and can quarantine or automatically rerun pages based on rules you set. It also checks that extracted table data lines up with the schema you expect, flagging duplicate headers or uneven rows rather than silently fixing them. The tool ships with tests and documentation aimed at both human users and AI coding agents that might operate it directly.