A run recomputes every score from its structured inputs, refreshes derived columns, runs the self-audit and writes a new version only if content changed. fetch and read need network and the LLM; they queue proposals, never edits.
Add companies to ingestion
Upload a workbook (edited in Excel, or the first v60 seed)
Dry run lists every cell that would change. Untick it to apply: the upload is ingested, the pipeline runs, and a new version is written if anything changed. The upload wins on every differing cell (overwrite with history, not a three-way merge).