The one distinction that organises everything
Two places to put the model. Not points on a spectrum — different systems with different failure modes.
Lesson 01 of 06 · 10 min
There are exactly two places to put the model. Compile time: it reads intent and emits an artifact in a language with defined semantics, a human reviews it, and a deterministic engine executes it forever. Run time: it reads intent and input together and emits the decision.
These are not points on a spectrum of how much AI you are using. They are different systems with different failure modes, different audit stories and different cost curves.
intent (NL) ──▶ [ model ] ──▶ candidate artifact (DSL / policy / table)
│
[ verification suite ]
[ human review gate ]
│
▼
signed, versioned artifact ──▶ [ engine ] ──▶ decision
│
▼
decision record
(input, artifact version, trace)The model appears once per rule, not once per request.
Which difference between the two matters most in a finance-ops context?
The rest of this lesson unlocks when you commit above.