Hostwatch / Weavatrix Hetero integration field notes
Run recorded: 2026-10-07. Publication review: 2026-10-10.
Source: local Weavatrix Hetero BENCHMARKS.md section 3c and results/compute-network-20261007/production.txt.
Operator identities, credentials and network endpoints removed.

Qwen3-4B; context 2048; n-gram draft; fixed prompt; maximum 48 new tokens.
Mac: 6.00 GiB lent to Metal; 5.15 GiB offered for layers; layers [0,26), 5.09 GiB required.
Linux container: 2 vCPU, 80% power; 2 threads, 3.20 GiB lent; 2.70 GiB offered for layers; layers [26,36), 2.04 GiB required.
Recorded memory upper observations: Mac stage 5.20 GiB, Linux stage 2.08 GiB.
Distributed answer hash prefix: bef34a972079ffd9.
Whole-model Mac answer hash prefix with same generation flags: bef34a972079ffd9.
Recorded comparison: byte-identical answer.
Model ends: 0.72 GiB; approximate layer size with cache for context 2048: 0.196 GiB.
Estimated compute: 212.4 ms per verification pass (placement estimate, not measured end-to-end tokens/s).
A coordinator configuration reserving a 0.6B draft beside model ends was rejected before launch with only 32 of 36 layers fitting.

This is a fixed-case integration result, not a throughput comparison or a confidentiality guarantee against a worker administrator.