Methodology
A model is only useful if the people it has to convince can take it apart. This is how ours are built, what we check before handing one over, and — just as importantly — what we do not yet claim.
Calculations are deterministic. Every figure is produced by a formula evaluator that accepts arithmetic and a fixed whitelist of maths functions, applied to named inputs. The same inputs always produce the same outputs. A language model never computes one of your numbers — it can help draft an explanation of a result, but it cannot create the result.
Free calculator results are recomputed on our servers. The figures shown in the browser are recalculated independently before we send you a breakdown, so what you receive does not depend on anything editable in the page.
Sensitivity is measured, not asserted. We perturb every input across its plausible range — several hundred samples — and rank inputs by how strongly each moves the headline result. That ranking is what tells you which assumption to argue about first.
You can check all of this. The public sample decision room runs the same engines, shows its assumption ledger, and lets you move the deciding assumption until the recommendation flips.
Before a model is delivered, an analyst who did not build it reviews it against this list. It is published so you can hold us to it — and so you can run the same checks against a model built by anyone else.
A methodology page that overstates the method is worse than none. These are on the roadmap and are not shipped:
An NDA is signed before any work begins. Your data is not used to train models and is not shared across engagements. Controls and the current subprocessor list are on the security page. You own every deliverable outright.
Five questions and you will see the scope, price band and timeline — before giving us your details. Decision Model Sprint from $8K–$20K.