Cassandra finds what is broken, writes the correction, and proves every one by recalculating the file. Nothing reaches you that it has not first proven to itself.
A real run: 8 findings, 5 proven, 114 seconds, nobody involved after the file landed.
They allocate budgets, price deals, and set national policy. Almost none are under test, and a wrong formula looks exactly like a right one.
That last gap is the whole problem. The errors are everywhere, they are consequential, and the people who wrote them are certain they are not there. Reinhart and Rogoff shaped global austerity policy off a sum that omitted five rows. JPMorgan's London Whale loss involved a copy and paste error inside a $6.5B hole.
Every other tool emits a warning and asks a human to judge it. Cassandra writes the correction, applies it to a copy, recalculates the whole workbook, and refuses to show you anything that did not survive.
Parse into a formula dependency graph. No model calls: a parser does this better.
Five deterministic detectors, plus an agent that reads labels against formulas.
Real defect, or a legitimate modelling choice? Series seeds get dismissed.
A fix that fails recalculation goes back with the reason. Three tries, then a human.
Moved exactly as predicted, nothing else touched. Only then does it reach you.
A workbook landing in a Cloud Storage bucket starts the audit. The human bottleneck is the root cause, so it is removed.
A hardcoded rate equal to the assumption it replaced is right today. Cassandra perturbs the driver and shows the two workbooks diverge.
Ordered by blast radius through the dependency graph, weighted toward the figures people quote in meetings.
Every verified correction applied, anything uncertain left alone, the original never touched. A repair, not a report.
Drop an .xlsx, paste a Google Sheets link, or let a bucket event start it. Sheets needs no OAuth.
None can read a file, change a workbook, or reach the network. Injected text can at worst cause a wrong judgement, which recalculation rejects.
Detection, the dependency graph, clustering, and every verification are plain Python. The model is asked only what a parser genuinely cannot answer, and the Verifier that decides whether it was right asks the model nothing at all.
Applies the patch to a copy and recalculates every cell. Did the target move exactly as predicted? Did anything move outside its blast radius?
Shown with its diff, and written into the corrected workbook you download.
Returned to the Patcher with the reason it failed. After three attempts, quarantined for a human.
Every answer is checked by recalculation.
Spreadsheet error detection is a mature field. The contribution here is the composition: autonomous, continuous, self verifying, and regression aware.
Drop a workbook in. Two minutes later you get the corrected file and the proof.
Open Cassandra