I wanted a loop where each change had a baseline, a result and a reason to keep or reject it. The ledger grew to 240 numbered experiments. It includes failed candidates, interrupted measurements and corrections, not just accepted wins. The original tiny-prl hillclimb figure, preserved without redesign. The figure is the project’s existing export. Its panels have different units and workloads. They are not one cumulative speedup product. Open the image to inspect individual annotations; the chapter below explains how to read them.

First make the complete loop fit

The initial work bounded context, sequences and process concurrency, and made trainer and inference coexist. Checkpoint saving exposed host-memory peaks that a trainer-only benchmark missed. Smaller staging units and per-tensor writes addressed those peaks. CUDA IPC then proved that a local full-parameter policy could refresh without writing a full checkpoint payload to disk. Correct transfer did not by itself prove faster whole-RL execution.

Then make a supported decoder faster

The kernel work moved through specialized decode, attention splitting, reduction-order investigations and lossless packed weights. Numerical admission was separate from timing admission. Several candidates matched greedy tokens but failed hidden/KV equality, or passed small fixtures and failed wider contexts. Some smaller probes did not transfer to the composed kernel. Packed activations, metadata changes, PTX staging and shared-memory reductions all needed their own controls. Experiment 240 is a useful ending point: a resource count improved, while timing remained neutral, so the change was rejected.

Finally ask what the user waits for

IcePop KV-cache work and score-centering integration also received runtime checks. The compact integration report records successful updates, sampler-support evidence and trace joins, while explicitly declining a meaningful centering, reward or speed claim from those runs. An algorithm name or an upstream idea was a direction to test, not evidence to adopt it. Earlier measurements remain historical records. The graph’s accepted points should move only when a new run passes its corresponding gate. In particular, a decoder result cannot move the main wall-time panel merely because it looks promising. Full ledger · IcePop/score-centering record Continue with the wall-time experiment.