Module Counterfactuals for Attributing LLM Serving Scheduler Effects

Abstract An unchanged aggregate score can hide a scheduler component’s effects. We specify module interventions for large language model serving that pair complete runs on matched workloads and compare recommendations at identical simulated states. Beyond which component is replaced, the contract declares observations and units distinguishing opportunity, recommendation, executed action, request outcome and aggregate outcome. In controlled simulations, replacing the input-prompt component changes completion timing or deadline status for about 59 percent of requests while leaving weighted deadline attainment unchanged. Across three selected component pairs, about 99 percent of recommendation disagreements disappear before admission; additional pairs show that erasure depends on the component and surrounding scheduler. Composition interactions, workload dependence and endpoint sensitivity limit transfer of aggregate conclusions. Thus, an aggregate null does not establish inactivity, and a changed recommendation does not establish a changed action or performance benefit. On seven constructed cases, the protocol and equally instrumented ablation produce identical diagnoses. The contribution is an explicit experimental specification and scoped empirical evidence, not diagnostic superiority or a new causal estimator. Conclusions remain model-conditional; native fidelity and identical-input serving-engine confirmation remain unestablished.