Large language model agents are increasingly expected to handle complex, long-horizon tasks that require sustained tool use and stateful interaction. Final outcome labels in existing agent benchmarks provide only a coarse view of long-horizon agent execution. They obscure differences in the execution process among successful trajectories and collapse failed trajectories that may have achieved different amounts of task-relevant progress. We introduce the Auditable Semantic Reference Graph (ASRG), a post-hoc framework that represents task-relevant progress with task-specific Semantic Reference Units (SRUs) and matches trajectories against these references. Each SRU is a checkable task milestone with an explicit verifier. Deterministic matching quantifies process coverage and coverage-weighted first-hit timing, while an action-level diagnostic records actions unexplained by the labeling rules. We apply ASRG to 4,740 Toolathlon evaluation trajectories from 22 models across 108 tasks. ASRG reveals process-coverage differences between models with identical final-success rates and first-support timing differences between same-task trajectories with matched coverage. By relating measured progress to resource use, ASRG also supports process-efficiency analysis. ASRG complements final outcome evaluation with auditable, reference-conditioned process diagnostics.