Evidence from 1 October 2026
Two correctness controls and a Qwen runtime comparison show the scope and evidence behind each result. These are public workloads, not customer checks.
The DSpark and FA4 correctness controls each ran three baseline arms and two fixed-source arms on one physical NVIDIA H100 NVL, with repeated A/A outcomes and eight scoped QA checks. The source change, protocol and exclusions are recorded in each report.
Speed is not calibrated. These controls do not establish model answer correctness, end-to-end inference, full-engine behavior, production suitability or release safety. No customer workload was tested.
A minimal Qwen FP8 test found the strict structural tag present on vLLM 0.27.1 and absent on 0.28.0. Eight GPU requests completed with schema-valid arguments; the report separates that result from the missing enforcement tag.
One H100 PCIe reproduced extra boundary ties: 4160 kept tokens instead of 1913. Adjacent kernel commits, an isolated fix and batch controls narrow the finding.
The source-derived CUDA fixture reproduced the expected stale-block assertion in all three baseline arms. Both fixed-source arms passed the allocation-bound check. No model weights or full DSpark engine were run.
The known defective and fixed source states produced their expected repeatable numerical signatures in three baseline and two fixed-source arms. This is a control-classification result, not a claim about latency or model output.
All 32 measured requests completed with the required 32-token output length in each of three vLLM 0.29.0 runs and two vLLM 0.30.0 runs on one H100 NVL. This checks runtime completion and output length; answer correctness and speed were not evaluated.
Report structure
One status for each listed check, with explicit untested areas and a separate infrastructure-error status.
Exact versions and environment, the same workload for both versions, A/A calibration, repeated observations and a trustworthy oracle where available.
Timestamped results, logs, manifest ID, machine-readable JSON and the attributable compute cost.