Visual summary checks fail for LLM-edited formulas at scale; need diffs and dependency checks

saimatrixxx · reddit · 2026-08-31

The author argues that checking if a "Summary tab looks right" is insufficient when an LLM edits 5,000 formulas. Potential risks include broken references, shifted ranges, hardcoded values, and logic mismatches.

The author suggests implementing before/after formula diffs, dependency checks, and automated reconciliation tests before sign-off, asking what constitutes a reasonable acceptance test at this scale.

Original post →

More from coding & agent

coding & agent channel →