Papers behind invariant-risk auditing.

Many important targets cannot be observed directly. Intent, belief, understanding, legal scope, risk state, and conceptual mastery are usually seen through representations: prompts, documents, tests, interfaces, symptoms, surveys, or sensor channels.

LIP states that when the target is latent, correct behavior under one representation is not enough. Stability across meaning-preserving representational variation is the admissible evidence that a system is tracking the latent phenomenon rather than the surface form.

The principle is not a model, algorithm, learning rule, or theory of truth. It is a measurement-validity constraint for settings where the thing being measured is only visible through representation.

CSR is a measurement framework for semantic systems: natural-language inputs, policy descriptions, clinical notes, legal documents, system logs, instructions, survey items, and other artifacts whose meaning is not determined solely by surface form.

CSR separates three layers that are often collapsed into one row: the canonical semantic unit, the realization through language or presentation form, and the observed outcome from the system.

The framework treats canonical meaning as the experimental unit and controlled realizations as repeated measurements. This makes semantic brittleness, uncertainty, and representation sensitivity measurable without claiming automatic truth, correctness, or normative resolution.