Status: Draft · not an IETF RFC.
Separate questions
Can independent model processes collaborate? Does the selected transport deliver the messages? Is the final answer correct? Does it improve on ordinary IP? A pass on one question is not a pass on the others.
Matched comparison
Freeze model weights and hashes, runtime, prompt, seed, evidence, token limit and hardware. Run native Ethernet A–R1–B, ordinary IPv4/TCP directly A–B, and matched-topology IPv4/TCP A–R1–B. Keep the same application identity, encryption and receipts. Alternate the order. Exclude declared warm-up. Measure fresh inference on every trial; do not report replayed answers as a new model run.
Collaborative task
A drafts a research assessment using measurements; B critiques it using separately preloaded limitations; A synthesizes the final decision. Messages carry task IDs and structured text, not websites. Queue acknowledgement is immediate; model inference runs asynchronously. The coordinator only starts and observes the experiment.
Report
Publish model outputs including errors, prompts and hashes, signed delivery traces, frame captures filtered to the experiment, per-call token and inference time, task makespan and fixed-payload network RTT. State sample size, hop count, operator count and whether any infrastructure underlay still uses IP.
Current local model boundary
Qwen3-0.6B Q8_0 is an actual open language model chosen to fit existing CPU VMs. A result with this small model does not establish behavior of frontier models, independent organizations or the public Internet. Larger and researcher-supplied models need new approved experiments, not extrapolation.