AlarmShift: Identifying the Value of Early Intervention in Adaptive Inference

Adaptive-inference gains are routinely attributed to early warning even when the trigger, available interventions, checkpoints, and compute all change together. AlarmShift resolves this ambiguity with a counterfactual protocol that compares alarm families under an identical intervention stack, then delays each selected alarm on its original instance while preserving alarm identity, dispatch opportunities, and realized compute. TriSignal, a four-parameter hysteresis trigger over entropy drift, attention relocation, and retrieval agreement, supplies a low-capacity instrument for this intervention. Across GSM-Hard, SWE-bench Verified, RULER-32K, AIME-24/25, and LiveCodeBench-v4, TriSignal improves Llama-3.1-70B-Instruct success by $+3.97/+4.24/+2.82/+6.86/+5.90$ percentage points over an entropy-matched policy; every paired bootstrap interval is positive, and monitoring consumes $2.4$ of runtime. The result reproduces on Qwen-2.5-72B-Instruct and remains positive against a frozen verifier change-point null on both backbones. On both backbones, same-instance replay removes roughly two-fifths of the margin within ten tokens and brings the thirty-token response to statistical overlap with the matched null, while random-instance timing controls remain flat. These experiments establish intervention timing as an independently measurable source of adaptive-inference performance and make AlarmShift a reusable test for separating warning quality from controller capability.