Eight percent of fields left empty beats two percent quietly wrong

Assignment is a narrower job than reading, and in the dev.to design the narrowing is the whole point. The model receives the stored text plus candidate values that a pattern pass already lifted out, and it returns a mapping rather than a value ^[20]. Errors then have to surface as a figure in the wrong slot instead of a slightly wrong figure in the right one, and only the first kind gets noticed ^[1].

Do the arithmetic on the author's own preference ^[7]. Take a thousand fields. The tolerant system leaves eighty of them empty, at roughly a minute of attention each, so about eighty minutes of clerical work ^[4][3]. The tidy system fills all thousand and is quietly wrong about twenty, and at the piece's own propagation figure of three copies apiece, that is eighty records carrying a wrong number ^[5][4]. That is the trade in plain terms: eighty minutes of scheduled work with a named owner against eighty records that need reconciling once somebody has already quoted one of them.

The edit-distance claim is checkable with a calculator. 349,000 to 340,000 is one character and nine thousand in value ^[2][1]. 132 square metres to 138 is also one character, and the author prices that gap at nine thousand euros, which implies about 1,500 euros per square metre ^[11][2]. Whether that implied rate matches your market does not matter; the ratio between how wrong the string looks and how wrong the money is stays lopsided.

The mechanism offered is digit-level. The cited research has models encoding numbers digit by digit in base 10, with errors that sit close in string edit distance and far away in value ^[9]. The author flags the gap rather than papering over it: those papers measure arithmetic and counting, not copying out of a document ^[10]. For the finding to transfer, the copy path would have to run through the same digit-level representation instead of a span lifted verbatim from the input, and nobody has shown that here. What stands without it is the weaker claim, that an error in a number need not look like an error, and the weaker claim is enough to justify the pattern pass ^[12].

The guarantee most teams already bought covers shape. OpenAI's documentation promises the response adheres to the supplied schema, with no missing required key and no invalid enum value ^[13], and the same documentation says structured outputs can still contain mistakes and that input unrelated to the schema can still produce hallucinations ^[14]. A price key will exist and will hold a number, but the schema guarantee stops there and does not check that number against the document ^[15]. A well-formed row looks audited whether or not it is.

Adoption cost lands on document variety. The candidate pass is very reliable on clean HTML and on PDFs with an embedded text layer ^[21]. On a scan the uncertainty starts at text recognition, and tables, differing decimal separators and currencies degrade the candidate list before the model is involved ^[22]. So the pattern-first layer does not fix scans; it removes one failure path and inherits another. In my context, HTML and text-layer PDFs feeding figures that get quoted downstream, that is the right trade, and the stored copy is the cheap half of it ^[17]. For photographed faxes, the honest answer is a confidence gate and a person on the digits.