Introduction
Interpretation Drift
Large language models produce different decisions for identical inputs, even under deterministic decoding (temperature 0). The cause is not sampling but interpretation.
When an input D admits more than one valid reading, |I(D)| > 1, models select different elements of the interpretation space and diverge. Every output can be well-formed and schema-valid, yet resolve to incompatible actions. Full argument and formalism →
# one incident, three models model_A → { "risk": "high" } model_B → { "risk": "medium" } model_C → { "risk": "low" } # all valid. incompatible actions.