summary.json: state transitions, correctness contracts and five MI300X benchmark rows.gates.json: focused and full CPU/HIP/sanitizer evidence.environment.txt: build/device/measurement contract.hip-benchmark/*.json: untouched unspecified-build diagnostic records.hip-release/*.json: untouched Release records with 2 warmups and 10 repetitions.
The benchmark is deliberately a negative performance result. It proves the scheduler semantics and shows why the divergent-row serial model oracle must be replaced before claiming serving speedup.