Concepts/Definition

Application context: a model may infer intent, but a transfer cannot be approximately correct

A user describes a complex payment in natural language. A model is useful for extracting the goal, finding missing information, and proposing steps. Payee identity, amount precision, account state, delegated limit, and idempotent commit require repeatable decisions and reviewable evidence.

Definition

The probabilistic core handles open-ended interpretation, planning, and candidate generation. The deterministic shell handles identity, state, authority, constraints, tests, commit, and external receipts. The core proposes an action; the shell decides whether it qualifies for execution and verifies whether the external world changed.

Engineering mechanism

natural-language goal
  → model produces a structured Intent
  → schema / policy / state / authority checks
  → controlled tool call
  → receipt and after-read
  → trace and later evaluation

The shell does not require a fully hard-coded flow. A model may still choose candidate tools and steps dynamically, but only from versioned capabilities and under repeatable pre- and post-conditions.

Boundary

Low-risk content work can use a thinner shell. Tasks that affect money, data, authority, production configuration, or user rights need a thicker one. Deterministic mechanisms can also fail, so policies, schemas, and validators require their own versions, tests, and observations.