Create 10-20 synthetic .xlsx fixtures that cover messy spreadsheet cases:
merged headers, side-by-side tables, notes above tables, sparse sheets,
date/year headers, totals, formulas, and multi-table layouts.
For each fixture, add expected outputs for:
- structural anchors
- compressed prompts
- detected ranges
- coordinate remapping
Why this is highest leverage: it turns vague "paper-inspired" claims into measurable behavior and protects future changes from regressions.
Add a lightweight formula graph that captures:
- formula cell
- formula string
- cached value
- referenced ranges
- cross-sheet references
- formula errors
- repeated-formula summaries
Why this matters: formulas are one of the biggest remaining spreadsheet-specific reasoning gaps, and this work is relatively isolated compared with full benchmark reproduction.
Make evaluation records explicitly label results as:
syntheticreconstructedpaper-original
Fail or warn if a run attempts to claim paper-comparable results without:
- dataset metadata
- split metadata
- model/backend metadata
- prompt serializer
- coordinate mode
- baseline status
- metric definition
Why this is leveraged: it prevents unsupported research claims and makes the repo methodologically defensible without requiring the original paper datasets.