Physical Trust Layer: from AI output to inspectable state
AI becomes useful only after its physical claims are grounded in equations, parameters, and constraints.
Principle
No text answer is trusted until it can be represented as a state the user can inspect.
Pipeline
User prompt -> AI Provider -> Intent classification -> Knowledge Card or SceneSpec.
Validation
Equations, units, parameters, and constraints remain visible before rendering.
Runtime
The workbench shows model state instead of hiding behind a chat response.