Extraction and checks
The pipeline locates month grids and table regions, checks crop boundaries and reads each section separately. Two prompts read each crop, with a third read to resolve disagreement. YOLO detectors I trained provide a fallback for locating regions. Code checks dates, row counts and assignment of TERYT addresses to zones.
Evaluation snapshot, 18 October 2026
The development evaluation covers six scans, with 2,274 cells per run and three pipeline runs. A cell is one locality, waste type and month; every date must match the manually checked reference. These documents were also used to tune the pipeline.
Across 6,822 cell results, the pipeline accepted 5,904 with no observed errors and withheld 918, or 13.5%.
Whole-page comparison
On the same six scans, the compact whole-page model had 26.8% cell error with its best tested prompt. A larger whole-page model had 0–0.5% error across the tested prompts. The case study compares extraction methods, publication checks and coverage on this evaluated set.