Takara TLDR - Daily AI Papers
・Answer accuracy is an insufficient reliability signal for LLM data agents.
・In structured-data tasks, a benchmark-correct answer can be produced by an invalid trace.
・This paper introduces Trace Integrity, a deployment reliability criterion for evaluating whether the computation recorded behind an answer is explicit, executable, schema-valid, operator-faithful, replayable, answer-consistent, and auditable.