Derivation boundaries: A structural source of faithfulness failure in large language models

Abstract

Faithfulness to supplied information is not uniform in large language models: some rules, thresholds, or features shape outputs correctly while others do not, with nothing in a model's own explanations distinguishing the two. An industrial classification task motivates this observation, where supplying a classifier's exact decision thresholds improved accuracy while degrading minority-class recall, raising the question of where such failures occur. Because that setting involved a closed model and proprietary data, the same question is examined in two domains: airline baggage fee and tax calculation. Across domains, models reliably handle directly available values but fail when a known fact must be transformed into a new value through a rule-conditioned lookup, points named here as derivation boundaries. Two error patterns recur there: misassignment, in which a valid value is assigned to the wrong side of a rule boundary, and fabrication, in which a value unsupported by the rules is introduced and justified as though retrieved correctly. Layer-wise analysis using the logit lens suggests a mechanistic explanation. Directly available values stabilize early, whereas boundary-dependent values commit later and less stably, with correct alternatives often remaining competitive until the final layers. Faithfulness failures concentrate at derivation boundaries rather than spreading uniformly, a pattern layer-wise analysis links to delayed internal resolution.