Beyond Prompt Tokens: Evidence-Safe Cost Measurement for Tool-Using Agents

Context-management interventions for tool-using language-model agents are often evaluated by the input tokens they remove, although an agent is billed over an entire trajectory and must still produce an acceptable result. We introduce an evidence-safe performance-evaluation protocol that joins an append-only per-call cost ledger with evidence-lifecycle records, executable oracles, and fail-closed analysis. The protocol reconstructs cost from raw provider usage and frozen price schedules rather than trusting persisted bills. We first observed prompt-side versus trajectory-level decision disagreement in 29 of 199 eligible paired blocks (14.6%; task-cluster bootstrap 95% CI: 8.5–21.0%). We then preregistered a real-provider experiment testing whether a persistent, treatment-symmetric evidence-failure opportunity increases economic false adoption: externalization lowers primary-input cost but raises total trajectory cost. The study attempted 56 paired blocks across 14 benchmark-derived tasks and two providers. Forty-three blocks had complete paired cost, while 19 of 112 branches ended in provider errors and remained missing. Contrary to the accepted hypothesis, injection did not increase false adoption. The complete-case injected-minus-control risk difference was −10.6 percentage points for DeepSeek and −36.4 percentage points for Qwen; Qwen’s worst–best missingness bound remained negative [−50.0, −7.1]. Instead, Qwen control branches already exhibited false adoption in all 11 complete pairs. Mean primary-input savings of CNY 0.001709 became a mean total-trajectory cost increase of CNY 0.005413, carried mainly by ordinary verifier repair and finalization retry. These results show that primary-input cost is not a sufficient performance metric for tool-agent optimization. Evidence safety, task completion, missingness, and complete-trajectory cost must be measured jointly. The study contributes an auditable measurement protocol, a paired decision endpoint for detecting disagreement between prompt-only and trajectory-level evaluation, and a preregistered falsification test that distinguishes injected evidence-failure opportunity from ordinary control-flow overhead. This document is a non-peer-reviewed preprint. The reported estimates are bounded by the evaluated providers, task family, missing provider responses, and the study’s six-call cap. The work does not estimate natural evidence-failure prevalence or a cross-provider pooled effect.