KEEP THE COLUMN HEADING WITH ITS CELL

This targeted follow-up investigates Q3 only. The four tuned answers all
cited OUT-2 but omitted the cabinet and drawer. We kept those runs' three
original page IDs and their order for each language, the same model and
session settings, and the same task instructions and four original questions.

The PDFs were built from a known table grid. table-layout.json records the
printed labels and rows of all eight original PDFs. The included script first
verifies that each page contains all of its headings and cells in the
extracted text, then replaces each table with one line per data row. Every
cell is paired with its own printed column header. All other page text remains
verbatim. This fixture-level exercise uses a known grid; it does not establish
that automatic extraction could recover every table in real business PDFs.

Original OUT-2 cells (one item per extracted line):
Geschier
Plaz nom Gebrauch
Kalibréierléier
Schaf B2, Tirang 4
Reserveléier
Schaf B2, Tirang 1

The two rows sent to the model:
Geschier: Kalibréierléier | Plaz nom Gebrauch: Schaf B2, Tirang 4
Geschier: Reserveléier | Plaz nom Gebrauch: Schaf B2, Tirang 1

No translation, added fact, example answer, score key or revised question
was inserted. We applied the serialization to every table on the three
supplied pages in each call. Indexing and retrieval were not rerun.
All four model answers now supplied B2 and drawer 4 and cited OUT-2.
See answers.csv and reviewed-answers.csv for the actual wording. This
four-question diagnostic was chosen in response to an observed failure.
We did not test the serialization on all questions or on an unseen corpus.

From the kit directory:
python 05-table-diagnostic/label-table-cells.py
python 03-recorded-run/answer-multilingual-rag-lab.py --kit . --rankings 04-dataset-tuning/rankings.csv --configuration current-and-headings --case Q3 --page-text 05-table-diagnostic/page-text-labelled.json --out new-table-answers.csv
