Tool Use Failure in a Small Local Language Model: An Exploratory Negative Result Case Study

This artifact documents a reproducible negative result observed in a locally deployed, instruction-tuned language model with nominal web tool access. Despite explicit tool invocation and forced epistemic overrides, the tested model fails to update a simple, time-sensitive factual belief that contradicts its training priors. Under certain prompting conditions—specifically when explicit corrective assertions are introduced—the model does produce the correct answer. The study does not claim a demonstrated causal mechanism for this behavior. Instead, it records a concrete failure mode and advances several competing hypotheses, including conflict-triggered update and correction-dependent belief revision, as directions for future controlled study. The results demonstrate that tool access alone does not guarantee epistemic belief revision and that conversational alignment objectives may supersede truth commitment in some small local models. This contribution is intended as a diagnostic negative result and includes a minimal replication protocol to support independent verification.