Per-record sensitivity propagation. Compounds across multi-client engagement work.
bench · sensitivity →The model layer
your stack is missing.
You've automated the rest — AP, AR, the close — with Claude Code, MCP, and your ERP's API. The financial model is the last thing still living in a spreadsheet. Flatland is the typed, recomputable, audit-trailed layer that fixes that: assumptions with names and types, a real dependency graph, institutional-grade exports your board, investors, diligence partner, or auditor expect.
Largest compiled DAG in the benchmark suite. Whatever your client throws at you fits.
bench · scale-warm →Same IR always produces same outputs. Audit-defensible against any skeptical board.
determinism →Inputs · typed assumptions
Outputs · compiled in topological order
Assertions · guardrails
The automation layer is done.
The model layer is the last frontier.
Intake
Your compile runs on our servers and is saved so you can share it. It also feeds the live heartbeat: counts, timing, and a coarse region — never the model itself.
Bill the same.
Deliver more.


