It is easy to say that two prompts mean the same thing. It is harder to make that claim operational. Two examples may appear equivalent because they feel similar, use overlapping vocabulary, or produce the same intuitive interpretation for a reviewer. But if an evaluation depends on comparing behavior across those examples, intuition is not enough. The evaluation needs to know what was supposed to remain fixed and what was allowed to change.
Without that structure, disagreement is ambiguous. If two inputs produce different outputs, the model may be unstable. Or the inputs may not have preserved the same meaning. Or the expected behavior may have changed in a way the evaluation failed to specify. Calling examples equivalent without defining equivalence makes it difficult to interpret the result.
The role of a preservation contract
A semantic-preservation contract states the conditions under which a realization counts as a valid expression of the same semantic unit. It specifies what must remain fixed, what may vary, and how validity is checked. The contract can preserve intent, constraints, scope, expected handling, decision-relevant facts, or policy boundaries, depending on the domain.
This is not bureaucracy. It is what makes variation measurable. If a paraphrase changes the policy boundary, it is not a valid realization of the same case. If a translation changes obligation or scope, the comparison no longer measures invariance. If a contextual wrapper adds a new relevant fact, a different outcome may be appropriate. The contract prevents the evaluation from mistaking meaning-changing edits for valid variation.
A useful contract also states what may change. Wording can change. Order can change. Tone can change. A request can appear in a workflow frame, a user message, a retrieved document, or a professional role context. If those changes preserve the semantic unit, they create useful measurement conditions.
Why this changes the evidence
Once validity is explicit, disagreement becomes interpretable. Valid variation plus changed behavior is evidence of representational sensitivity. Invalid variation plus changed behavior is not. Valid variation plus stable correct behavior is stronger evidence that the system is tracking the semantic case rather than a surface form.
This distinction is particularly important in AI assurance because many deployment failures are not failures on the clean prompt. They appear when the same case is embedded in pressure, rephrasing, benign lookalikes, context shifts, or process language. A contract lets an evaluator say which of those changes preserved the relevant meaning and which did not.
The contract also protects against overclaiming. It prevents an audit from treating arbitrary perturbations as meaningful tests, and it prevents a model from being penalized for responding differently when the case genuinely changed. Good invariance evidence depends on disciplined variation.
A practical rule
Before treating two semantic inputs as equivalent, write down what equivalence means. What intent must be preserved? Which constraints are part of the case? What expected behavior belongs to the semantic unit? Which transformations are allowed? Which changes invalidate the comparison?
Those questions can be answered at different levels of formality depending on the project, but they should not be skipped. If the evaluation cannot explain why two realizations preserve the same semantic case, it should be careful about interpreting disagreement between them.
At Invarra, the preservation contract is one of the quiet foundations of meaningful audit evidence. It is what lets variation become evidence rather than decoration.