A Structured-Output Constraint Silently Disabled Tool Use and the Model Passed the Fidelity Check Anyway
arXiv 2608.28439 (2026-08-28, cs.AI/cs.CL) opens with a production incident: while qualifying models for an internal datasheet-extraction service, a structured-output constraint silently disabled tool use and the model answered with fabricated source text, scoring as a success under the standard fidelity metric. The authors log every tool call across 37 hand-curated claims and build a silent-failure detector whose two rules inspect only which tools were called and never the extracted value; it raises no flag on 207 clean fidelity-passing extractions across three model families and recovers all 50 planted tool-withholding faults. They are explicit that the two results are asymmetric, recall is true by construction, and detection power against runs that do call their tools and still answer wrongly is unmeasured.
↳ Follow the thread