1 of 6
One page, twelve months, hundreds of numbers
This is the Wiśniew waste schedule, a scanned printout. A resident needs one thing from it: which bags go out on which days.

2 of 6
Where it can't see each day, the model fills in a pattern
In Wiśniew, bio waste goes out once a month. For August the document has one date, the 24th. The whole-page prompt returned a weekly rhythm.
Wiśniew, sector I, Borki-Kosiorki, bio waste, August 2026.
3 of 6
So the pipeline cuts the page into months
The model marks the month frames. Code moves each border into the white space between months and rejects a crop if the last day it can see is wrong, so no week falls off the edge.
0 of 12 crops
Each green frame is a separate request with 31 days in it instead of about 370.
4 of 6
Two reads per month, a third if they disagree
Two prompts read every crop, with the legend attached and a small copy of the page that shows where the month sits. Matching reads go through. If they differ, a third read decides by majority.

daysdays✓2 of 3–Above: January cut out of the Czerna (Nowogrodziec) schedule. For mixed waste the pipeline returned the 9th and the 22nd, as in the document.
5 of 6
Code checks what needs no interpretation
Sundays and holidays have red numbers in the calendar. A model hunting for the red of mixed waste across the whole page also finds them: the prompt gave mixed waste on 1 January and 6 January, New Year's Day and Epiphany.
The pipeline's month prompt says the colour of a digit is not a mark, and code flags every date on a Sunday or a holiday. Such a date only passes if a separate check of the crop confirms it. Code also checks gaps in the rhythm, row counts, and that every address in the national register lands in exactly one zone.
6 of 6
Checks, coverage and results
Development evaluation, 18 October 2026: six scans used for tuning, 2,274 cells per run, three runs. Every date in a cell must match the reference.
The larger whole-page model had 0–0.5% cell error across the tested prompts on the same scans. The full case study compares the methods and publication checks.