Execution-Path Qualification and Realized Costs in Speculative Decoding on Consumer Systems

Abstract

Numerical execution paths can change greedy output; a mismatch alone does not predict the remaining trajectory. On the tested B570 stack, we select a draft-free target-only micro-batch-one predictor using exposed discovery data, then freeze complete generated-ID predictions before reserved live speculation. It matches 72/72 capped sequences on 24 repeat-stable prompts, including all 12 runs of four reference-divergent prompts; the same-set reference matches 60/72. The predictor is sufficient for this path, but it neither restores qualification nor identifies a unique numerical cause. Separate B570 same-state forks test intervention effects; RTX proposal/replay programs test reproduction while sharing numerical routines. We assess cost separately. Newly direct-ID-qualified RTX serialization is slower within its instrumented post-prefill boundary ($0.646\times$ target/serialized latency). At fixed measured progress and target work, 358H remains below parity even with draft and residual costs removed. Qualified M4 Guard beats target-only but loses to fixed speculation. These results give testable diagnoses and engineering decisions about lossless compatibility and complete cost on the measured paths; they establish no portable decoding or control gain.