Problem
The prompt notes pre-1996 OCR artifacts (l↔1, O↔0, dropped commas, shifted columns) but provides no concrete grep strategy for handling them. A model reading a warning cannot act on it — it needs tool call patterns.
Suggested Fix
Replace the warning with actionable patterns in the TABLE READING section:
# For pre-1996 files, use alternates to handle OCR l/1 and O/0 confusion:
grep -iE "pub1ic|publ.c|pubiic" /app/corpus/treasury_bulletin_199[0-5]_*.txt
grep -iE "[0O]utlays|0utlays" FILE
# For shifted columns: extract a wider sed window (+/- 5 lines) to catch misaligned data
sed -n -5,5p FILE
Also add: "For pre-1996 bulletins, always verify extracted numbers by checking the row sum against a known total in the same table."
Impact
Low-medium — pre-1996 documents are a known accuracy gap; concrete patterns would help the model recover some of the always-fail questions in that date range.
Problem
The prompt notes pre-1996 OCR artifacts (
l↔1,O↔0, dropped commas, shifted columns) but provides no concrete grep strategy for handling them. A model reading a warning cannot act on it — it needs tool call patterns.Suggested Fix
Replace the warning with actionable patterns in the TABLE READING section:
Also add: "For pre-1996 bulletins, always verify extracted numbers by checking the row sum against a known total in the same table."
Impact
Low-medium — pre-1996 documents are a known accuracy gap; concrete patterns would help the model recover some of the always-fail questions in that date range.