Determine the mechanism underlying composite-directive cancellation

Determine whether the lower target-selection rate produced by the composite of a criterion and record identifier on certain Claude models is caused by a trained disposition, a preference for reasons over pointers, the composite directive's added length, or the inherited store's presentation of the verification-priority field.

Background

The paper reports that, on Opus 5, Fable 5, and Fable 5.1, combining a criterion that identifies a record with the record identifier itself can reduce target selection relative to the criterion alone. Follow-up studies test exact suffixes, labels, ratification language, explicit target fields, and larger verification budgets, but these interventions are bundled and therefore do not isolate a causal mechanism.

The authors explicitly state that the experimental instrument cannot distinguish among several competing explanations: a model's trained behavioral disposition, a preference for natural-language reasons over opaque pointers, the additional length introduced by the composite directive, and the way the inherited store presents the directive field. Consequently, the mechanism remains unresolved rather than being established by the reported descriptive contrasts.

References

We do not know whether a trained disposition, a preference for reasons over pointers, the composite's added length or the store's presentation of the field is responsible; the instrument cannot separate these.

Plan Pointers and Record-Directive Form in Budgeted Verification of Inherited Agent Memory  (2609.03450 - Nakayashiki, 3 Sep 2026) in Discussion, paragraph “The three Claude endpoints, decomposed”