Mesh proof harness.
Fire N real inference requests and measure honest routing rates — local/mesh vs. cloud fallback, verified against persisted logs.
Test Configuration
1–100 per wave.
mesh_ok allows cloud fallback.
Fixed for on-device WebGPU.
concurrent = parallel users
parallel requests per batch
repeat for sustained runs
gap between waves
Verified operational truths
Verified execution capacity model
audited 2026-07-22Maximum simultaneous real on-device inference executions observed. Only node_e7e7eb0e has ever executed real inference (latency_ms > 0).
17 nodes registered historically (3 online, 14 offline). 8 appear in task logs, but 7 of those are re-dispatch cancellations (latency_ms: 0) — not real inference.
10 sequential requests: 9 executed on-device via WebGPU, 1 cloud fallback when the node was occupied. 100% success rate, verified against persisted RoutingLog.
GPU-bound, not router-bound. Avg latency 8,745ms across 10 requests. Router overhead is negligible; the floor is WebGPU cold-start + generation time.
Dispatch layer verified functional: capability-based routing (only nodes with the requested model in models_available are eligible), heartbeat freshness gating (60s staleness threshold), and honest cloud fallback (triggers when no capable node completes within the 2-stage deadline).
Scaling constraint — physical, not software: benchmarks at 10, 50, or 100 concurrent nodes cannot be generated without provisioning real GPU-equipped devices. Each node requires a browser tab with a WebGPU-capable GPU, the model loaded into VRAM, and an active heartbeat. Simulated nodes are prohibited — a simulated node does not execute inference; it would relabel cloud fallback as "mesh," falsifying the very claim being measured. Scaling to 10–100 nodes requires provisioning real hardware, not software changes.