Real 2x H200 SXM expert parallelism over DeepSeek's own DeepEP.
Correctness gate passed: perfect top-1 and mutual top-5 agreement vs. a
single-GPU reference. The measured answer disproved this project's own
hypothesis -- Phase 1's naive-vs-persistent kernel crossover doesn't
reproduce under DeepEP's real per-expert token counts (naive wins at
every tested count). Cost: $13.03 of a $25 cap.

Full account: docs/findings/phase-3/2026-09-16-phase-3-multi-gpu-ep-run.md