The document prints its own checksum. So "did we read it right" is arithmetic, not an opinion.

A school timetable, literally an hours table in the law’s own vocabulary, is a table of subject by grade by weekly hours, printed inside a regulation. The regulations are what a state education ministry publishes in its gazette, and they arrive under the law’s own names: regulation, decree, circular decree, notice, amending regulation, correction notice. The numbers a researcher wants have been public for decades. They have simply been typeset rather than tabulated, and the typesetting is the work.

Reading one is not OCR. German timetables join several adjacent subjects with a brace and print one shared value for the group. The brace is a typographic glyph with no height, so when layout recovery flattens the page the shared value lands on whichever row the engine thought was nearest. Four subjects become one subject with a number and three subjects with nothing, or the value is copied onto all four. Both readings are wrong, and only one of them looks wrong.

Four more traps, one sentence each. A column that is blank on every subject row still exists as a column, because only the total row carries a value there, and dropping it slides every value one place left. Physics / biology or natural sciences is a choice between two subjects and one, not a long subject name, and only the arithmetic can tell those apart. Hours are printed as 2/3, +2/3, 1,5, (4) and 12 to 14, so the field is free text and not an integer. And one timetable in this batch is three sentences of prose telling you to replace a five with a two in one cell.

The print technology is the last of it. Twelve of the twenty delivered documents are pure page images at 300 dpi with no text layer at all, roughly 2,432 by 3,496 pixels. Typewritten curricula from 1956 sit elsewhere in the corpus. Documents printed in blackletter type, which several states’ timetable law reaches back into, were excluded from this project outright, and that exclusion is worth saying out loud: one spec does not read every printing technology these gazettes were ever set in, and this one was never asked to.

The contract is two datasets and not one. The first is the timetable as printed: subjects and groups verbatim, combined grade levels such as 5–6 and 7–9 preserved as they are shown, hours exactly as printed, ambiguity indicators included. The second is the legal text around it: titles, footnotes, contextual notes and amendment references, structured so that they can be linked back to rows. There is no automatic redistribution, no normalisation and no semantic reconciliation between the two. One explicit exception exists, at the researchers’ request: a hyphen or an empty cell is recoded to 0.

The spec is the product. Seventeen global rules at version six, co-authored across six rounds of comment and revision. In product terms that spec is a Spec, defined in the platform’s pipeline: platform configuration, not bespoke software written for one corpus. The record calls it not a field list, but a codified reading of how German timetable law is typeset, and that is the honest description of what changed hands here.

The customer’s vocabulary wins, every time. Subject names are never normalised, because the variation is itself historical data: geography alone has three spellings inside one state’s twenty documents, work and technology has six, religion and ethics has five. The group separator had to become a plus sign rather than a slash, because a subject name such as work / economics / technology already contains a slash. That looks cosmetic. It is actually about whether the boundary between group members survives the trip.

Two boundaries are design. Mapping raw subject names onto research categories is a research decision and not an extraction one, so it stays with the researchers. One regulation is never reconciled against another, by rule, which is the zero the band’s RESOLVE stage prints.

Every printed timetable regulation carries its own checksum. It prints a total column per subject and usually a total row per grade, and that turns "is this table correctly aligned" from a judgment into a sum. The check costs nothing, it runs before any structuring, and it is deterministic. Ground truth is normally the expensive part of this work. On this corpus it is printed on the page.

The checks below are not code written for one corpus. They run as agentic validation in the same pipeline the Spec is defined in, so the rule about how a page should be read and the arithmetic that holds the reading to it are a single configuration.