Four rules, and what each one is guarding against.
These are constraints, not aspirations. An instrument satisfies all four before it is used on anything that matters, and each one is here because the alternative failed in a room with a client in it.
Calculation before interpretation
Where a number can be computed, it is computed. A language model is asked for wording and for judgement about wording. It is never asked for a figure that arithmetic can settle.
Assumptions on the surface
Every forecast carries the inputs it was built from — the market, the flight, the product mix, the date the numbers were pulled — on the same screen as the result.
A person approves
Nothing reaches a client, a platform or a budget without someone signing it. Automation drafts, schedules and checks. It does not decide.
Reproducible or it did not happen
A run is saved with the inputs that produced it. Asking the same question twice returns the same answer, or the difference is explained by something that changed.
Consequence
What follows from them
Together the four rules decide most of the architecture before any particular problem is picked up. Calculation in code rather than in a prompt means the arithmetic lives in tested functions. Inputs on the surface means a result is a record, not a message. An approval gate means every automated path ends at a person. Reproducibility means every run is stored with what produced it.
A number that cannot be checked is an opinion with a decimal point.
The rules cost speed. A system that asks a model for the figure ships sooner than one that computes it and tests the computation. That trade is made deliberately, because the work these instruments feed into is work somebody has to stand behind.