Hundreds of millions of domain experts encode real institutional knowledge as spreadsheets: underwriting models, pricing engines, capacity plans, actuarial tables. Those programs run only inside Excel's interpreter, unversioned, untested, single-user, and invisible to the whole agent ecosystem.
Organon changes the runtime. It compiles a workbook's calculation logic into a deterministic, typed engine and exposes it as an MCP server, with automatic cell-level parity verification against the source workbook.
The governing principle is a hard line I do not cross: AI decides, deterministic systems execute. The model does semantic understanding and UI composition. It never computes a number.
Every compiled model ships with a signed parity certificate that states exactly how many cells were verified against Excel's own cached values, at what tolerance, and which cells were excluded and why. You can trust it because it shows its work.
Then an app engine builds an interactive application around the compiled artifact. The spreadsheet stops being a fragile single-user file and becomes a real, versioned, testable, multi-user program that agents can actually use.