I kept a local fork. At night, I would run small pipelines on tired datasets: attendance records with dropped columns, clinical logs with inconsistent timestamps, shipping manifests with encoded abbreviations that smelled of a different era. Each run produced a report that combined quantitative summaries with prose reflections: "Confidence: medium. Likely source of discrepancy: timezone offsets introduced during import. Suggested next step: consult ops notes from March 2017." The language felt human because it was — the tool encouraged humans to remain in the loop.