adjudication-sheets

# Human Adjudication Sheets from LLM-Pipeline Data ## Problem Adjudication sheets built from pipeline intermediates tend to carry truncated "excerpt" columns (previews made for machine diffing, not human judging). A human asked to rule "does this text show X?" on a fragment produces invalid rulings: the decisive sentence may sit past the cutoff. Separately, multi-line text in CSV cells renders as broken rows/empty cells in Excel, so the labeler reports "there is no data here" even when the column is populated. ## Context / Trigger Conditions - A labeler says the sheet has "no information" — check for embedded newlines in CSV fields first. - Excerpt lengths cluster at an exact value (all 1,204 / 1,500 / 800 chars) or end with "[…]" — that is a hard cap, not natural length. - The ruling question is "does the record/text show X?" — the ruling basis must be byte-identical to what the classifier/raters saw, or the ruling grades a different object. ## Solution 1. **Trace the true rating basis from the scoring code, not from intermediates.** Find the line where text is passed to the model (e.g. `prompt = ... + text`) and reproduce that construction exactly. Verify: stored excerpt == full_text[:cap] (0 mismatches), or cache-key match against the LLM request cache. 2. **Deliver .xlsx, not .csv**, for any sheet with multi-line text: wrap_text + frozen header + generous width on evidence columns; data-validation dropdown on the `ruling` column with the exact allowed codes (e.g. E_pos...