04 / MACHINES
The machines.
The environment matters. Hardware, software, versions, and changes belong beside every result.
The first test bench
WMSI’s primary research machine is an NVIDIA DGX Spark with 128 GB of unified memory. It is the standalone WMSI workstation for local model inference, automation, and longer-running experiments.
Why the baseline matters
Before changing a runtime or testing a model, we preserve the system state: operating system, driver, CUDA, firmware, containers, storage, and measurement tools. Each later run should identify the exact environment that produced it.
Verified apparatus checks · September 26, 2026
- GPU computation: 65,536 of 65,536 integer additions correct in the separate CUDA check.
- First local model: Qwen3-4B-Instruct-2507 Q8_0, running in a pinned llama.cpp container, used the GPU for all 37 model layers.
- Bounded response check: 9 of 9 exact answers across three fixed prompts repeated three times, with no truncations.
- Evidence integrity: all 38 files passed the saved checksum manifest check for the successful inference record.
Observation: These figures summarize the recorded apparatus tests. The test server shut down after the run; that check did not create a persistent inference service.
Nine small checks do not establish broad reasoning ability, sustained throughput, or readiness for all 90 studies. Larger reference workloads and the general experiment platform remain ahead.
Source records
Derived from the September 26 CUDA computation result, first-model inference result, and Master Build Plan v1.2. The sealed baseline retains its August 30 identifier; capture and sealing occurred September 6. Raw machine records remain in the local research archive.