Short pieces about what actually happens when a model is asked to reproduce a document. Each one is built from this site’s own job records, with the query or code path named so a reader can check it.
Published 2026-10-28 · 6 min read
During testing, a model asked to transcribe a table turned 循环贷款 into 循环贷款等 — one character more, invisible when skimming, and wrong. This is why we check every cell against the source instead of trusting the output.
Published 2026-10-28 · 7 min read
A real conversion log from this site: read, convert, render back, fail validation, optimise, validate again. Why a conversion that fails costs nothing, and why the successful ones are slower than the failed ones.
Published 2026-10-28 · 8 min read
We filtered a corpus of freelance job listings down to 495 paid, productizable PDF and table jobs, then measured what they actually request. Accuracy and "flag what you cannot read" beat every feature count.