Interpretation Drift

A Concrete Case

Consider a cybersecurity event. One model rates an incident high risk, another medium risk, a third low risk. Every rating is well-formed and passes the same JSON schema, yet each triggers a different, incompatible action that determines whether a breach will be contained or ignored.

Schema validation catches nothing here: all three ratings are structurally valid. The divergence is not in the format, it is in the interpretation. The same event is classified as both high and low risk by different models, and sometimes by the same model on a different run.

It Is Not Randomness

The intuitive explanation is sampling randomness. It is wrong. Large language models exhibit output variance even under deterministic decoding (τ = 0), where sampling is removed entirely. The randomness is gone, and the output still diverges.

Our research shows this variance arises not primarily from sampling or implementation artifacts, but from interpretation itself: a task that admits multiple valid readings does not force a model to pick the same one twice. Remove the randomness and the ambiguity remains.

The Formalism

Let D be an input artifact (its data together with the task instructions), and let I(D) be its interpretation space: the set of internally coherent task interpretations compatible with D. A task is determinate when |I(D)| = 1, a single admissible interpretation, so every competent system resolves D the same way.

Interpretation drift is the condition |I(D)| > 1: multiple admissible interpretations coexist, so different models, or the same model on different runs, may select different elements of I(D) and produce divergent outputs. This is divergence from multiplicity, not from model error. Behavioral prompting can bias which interpretation is selected; it does not change the cardinality of I(D).

The fix is not a larger model. It is a constraint specification S (a substrate) that restricts the space to IS(D) ⊆ I(D). Deterministic interpretation obtains exactly when |IS(D)| = 1: the space collapses to a single admissible interpretation i*, and interpretation stops being a source of variability. See Substrate Engineering for how a substrate is constructed.

Why It Becomes Risk

As AI systems move from generating content to making decisions, interpretation stops being a stylistic concern and becomes infrastructure. When the same input can lead to different actions across models, environments, and time, non-determinism becomes operational risk: systems that are individually reasonable yet collectively unreproducible, unverifiable, and impossible to govern.

A substrate collapses I(D) to a single admissible interpretation; TCP/AP then validates the resulting decision and permits it to execute, or fails explicitly. See How It Works for the mechanism, and the specification for the normative contract.