Experiment 238 compares serial and early trainer startup. Both arms already use the same accepted 360M kernel, batch-invariant native serving and mixed-decode repair. The variable is startup ordering.

Ten complete processes

There are five control processes and five early-start processes, interleaved in ABBA blocks and a final pair. Each performs two full-parameter alphabet-sort updates and saves a final checkpoint. The paired rows label runs, not execution order; the ledger preserves the actual interleaving. Median reduction is 11.745%. Every early process is below every control in this sample.

Match the work before trusting the clock

All ten runs finalize exactly 166 rollouts, 83 tasks and 18,138 tokens, with matching loss, gradient norm and mismatch KL at both trainer updates. Common prompt/seed requests return matching token IDs and counts. The benchmark’s deterministic per-prompt seed hook is private research machinery; normal CLI sampling remains stochastic. Timed shutdown tails still differ. Actual generated work spans 172–175 successful requests and 5,047–5,158 generated tokens. The gate permits only a bounded difference and retains that work in elapsed time. Some stored logprobs differ on discarded zero-signal groups, so this is not universal serving-logprob determinism.

Why time fell

The early-start path overlaps CPU-side trainer startup with vLLM startup, then waits before initializing trainer CUDA. One-second log intervals from inference launch to initial policy broadcast have medians around 27 seconds for serial and 21 seconds for early startup. Checkpoint I/O still costs 5–8 seconds. Timing includes startup, rollouts, updates, refreshes, final checkpoint I/O and shutdown. Post-run checkpoint hashing and cleanup are excluded. All guards passed with zero job swap; sampled peak GPU memory was 7,228 MiB.

The limit of the claim

This is a startup gain in an opt-in two-update recipe. It does not isolate a new kernel gain, compare against every unchanged default, or establish long-run learning throughput. In Experiment 239, ten-update control and early runs took 158.865 and 160.390 seconds, but generated-token work differed by 8.755% and trainer trajectories diverged. That pair establishes neither a speedup nor a regression. Batch waiting dominated its recorded trainer time; the next measurement needs comparable work and a cause for the divergence. Experiment 238 and its controls Continue with the learning experiments.