AI Governance & Assurance · live demo
The spreadsheet retires only when the replacement proves it, cell by cell.
Critical close calculations live in unmanaged Excel with no version control and one departed owner who understood the formulas. Retirement projects stall because nobody can prove the replacement produces the same numbers. This demo draws the trust boundary the way I believe every EUC remediation should: a risk register scored in code from parsed workbook structure with fixed, visible weights, a deterministic differ that recomputes every cell of the governed micro-app against the spreadsheet's stored values, and a cutover gate that is a boolean over max absolute difference at a tolerance shown on screen. The model only quotes cells and drafts the memo, and a named finance owner dispositions every non-zero difference before the lever unlocks. Public dataset: Enron spreadsheets corpus (CC BY 4.0) + one synthetic close workbook · runs in your browser
The EUC risk register · scored in code
8 real Enron workbooks + this close workbookfixed weights · hardcodes per formula (cap 0.75) 20 · external links 15 · hidden sheets+cols 15 · volatile 10 · nesting depth 15 · formula mass 10 · key person 15 · no model touches the score
The retirement differ · workbook vs governed micro-app
Human inspection view what a reviewer sees
| GST F5 workings · Q2 FY26 | C · base | D · tax / result |
|---|
Deterministic differ recomputed vs stored
| cell | close-pack output | spreadsheet | micro-app | diff | status |
|---|
The retirement memo · drafted from register and diff results only
Real rows: Enron Spreadsheets and Emails corpus (Hermans & Murphy-Hill, CC BY 4.0), fetched via the SheetJS raw-file mirror, parsed statically · the close workbook is seeded synthetic and marked as such · the "model" here is a deterministic stub, so every run reproduces exactly · amounts in integer cents