A restricted CNE program keeps the model from supplying the computed figure

A model fine-tuned to avoid inventing figures becomes less likely to invent one. Likelihood is the wrong currency when the output is a number someone will act on: a rare fabrication that carries the same confidence as a computed answer is difficult to distinguish without provenance.

The restricted CNE path moves the computed figure out of model recall. The model authors a program in a constrained domain-specific language, and a deterministic interpreter executes that program against the named data layer. Within that path, the figure returned by the DSL is the interpreter’s output and its provenance includes the program and inputs.

That language has no eval, no exec, no shell, no arbitrary imports, no dynamic attribute access, and no filesystem or network reach. These properties constrain what an accepted program can do; they do not make the program’s question, inputs, or interpretation correct.

Outside this restricted path, Lattice can narrate a tool result and can state a model-origin estimate. The visible trace distinguishes those estimates from tool-backed figures; it does not pretend the model can never write a number.

Where the model still earns its place

Routing a question, authoring the restricted program, and narrating a computed result in prose are all model tasks. On the restricted path, the model is not the source of the computed figure; it remains responsible for choices and narration that the interpreter does not validate.

A related evaluation result sharpened this. In a sealed assessment, a model handed a worked exemplar reproduced its structure on all 140 items — which reads as competence until the copied structure is separated from independently derived work. An evaluation that cannot distinguish those outcomes is measuring the exemplar, not the model.