pilot.jsonl: one fresh Qwen/microLLM and PyTorch pair at context 512.profile/: candidate Qwen process-wide Kernel/HIP API aggregates.comparison.json: Experiment 051 baseline, candidate and discard calculation.profile-summary.json: exact aggregate and split-Kernel counts.
The candidate source was removed after the first process and profile both failed their gates. The T=256 CPU/HIP Q/K/V correctness test passed before removal.