How it works

1. 01 · DECLARE

   Describe the vehicle and the question

   Zones, nodes, links and slices in YAML, optionally generated from a SysML v2 model. The manifest is the only source of truth; there is no database.

   vxperiments/experiments/exp-001-ws/main.yaml

   kind: experiment
   id: exp-001
   network:
     slices:
       - id: adas
         mode: vlan       # s-nssai: real 5G
         vlan: 101
         qos: { min_rate_mbps: 100,
                max_rate_mbps: 200, priority: 1 }
       - id: infotainment
         vlan: 103
         qos: { max_rate_mbps: 400, priority: 5 }
   method: { duration_s: 60, repetitions: 5 }

2. 02 · RUN

   Provision a virtual vehicle, idempotently

   A Rust control plane converges the host to the manifest: gateway, zones as containers, microVMs or full VMs, a 5G core, services and probes. Running it twice is a no-op.

   terminal

   # validate everything first
   $ vxedge experiment validate
   $ vxedge doctor --tier 1
   
   # provision, then drive load
   $ vxedge experiment up
   $ vxedge traffic up --manifest exp-001
   
   # or the whole campaign, ~35 min
   $ sudo -E ./scripts/run.sh --repro

3. 03 · PROVE

   Every run archives its own evidence

   Raw scrapes, structured samples, run environment, the manifest's SHA-256 and image digests. Analysis reads archives only, so anyone can regenerate the figure.

   excerpt of exp-001-summary.md

   | load | p99 µs (IQR) | loss % |
   |    0 |  89.6 (6.4)   |  0.0   |
   |  600 |  81.9 (5.0)   |  0.0   |
   
   adas p99 at max load: 81.9 µs
     = 0.91× baseline (threshold 2.0×)
   H1 HOLDS

Engineers and agents drive this loop through the same API. Neither gets a shortcut around the checks.

# the experiment manifest
vehicles: [vx-h1-vehicle]
vehicles: [vx-h-harness-vehicle]

$ vxedge experiment validate

Both vehicles are real manifests in vxperiments: exp-001 runs on X, exp-014 on Y. The diff is the one line that binds an experiment to its vehicle.