Verify exclusive GPU runner ownership before benchmarking

Before comparing GPU inference arms, verify that exactly one benchmark runner owns each arm’s output target and GPU workload. Process-name checks alone are insufficient: stale diagnostics can use the same model server or accelerator under a different command, silently inflating latency and timeouts. Inspect open result-file targets and GPU process ownership, quarantine any partial outputs produced under contention, then restart every affected arm from the same case boundary. Reproduced failures after a clean restart are evidence about the case or model; failures seen only under contention are not.

Applies when: running parallel OCR, vision, or LLM benchmarks on rented or reused GPU workers. Keywords: GPU benchmark isolation stale runner contention OCR Qwen PaddleOCR Vast