DoE: K under network latency -- the capsule repository pulling from Nexus, latency dialled in #150

New Issue

2026-11-07T23:16:09Z

rajames commented

2026-11-07 23:16:09 +00:00

Ruled

Captain Bob, 2026-11-07: "Networking latency will be a super good K test."

Why

Everything that has tested K≡1.0 so far comes from inside the machine -- words, messages, blocks, heat. A network is the first source of delay the system does not control: late, uneven, sometimes lost. K holding while the engine waits on the world is a stronger result than K holding on a self-paced workload.

The experiment

Factors, dialled in on the host side of the QEMU guest's network device with tc netem (repeatable to the millisecond): delay (0, 10, 100, 1000 ms), jitter, loss, bandwidth
Workload: the capsule repository (#149) resolving and pulling capsules from Nexus -- real work across the network
Response: K, and the heat and scheduling measures the DoE records, at each setting, on amd64, aarch64 and riscv64 (parity is the other half of the claim)

Before it

Networking, and #149's pull from Nexus
#142 (three DoE columns always zero) and #141 (empty DoE CSVs among the evidence), so the campaign's record has no holes and stands out

Done when

The campaign designed (levels, reps, run order), run, and K reported against latency per architecture, with the CSVs committed

🤖 Filed with Claude Code

### Ruled Captain Bob, 2026-11-07: "Networking latency will be a super good K test." ### Why Everything that has tested K≡1.0 so far comes from inside the machine -- words, messages, blocks, heat. A network is the first source of delay the system does not control: late, uneven, sometimes lost. K holding while the engine waits on the world is a stronger result than K holding on a self-paced workload. ### The experiment - **Factors**, dialled in on the host side of the QEMU guest's network device with `tc netem` (repeatable to the millisecond): delay (0, 10, 100, 1000 ms), jitter, loss, bandwidth - **Workload:** the capsule repository (#149) resolving and pulling capsules from Nexus -- real work across the network - **Response:** K, and the heat and scheduling measures the DoE records, at each setting, on amd64, aarch64 and riscv64 (parity is the other half of the claim) ### Before it - Networking, and #149's pull from Nexus - #142 (three DoE columns always zero) and #141 (empty DoE CSVs among the evidence), so the campaign's record has no holes and stands out ### Done when - [ ] The campaign designed (levels, reps, run order), run, and K reported against latency per architecture, with the CSVs committed 🤖 Filed with [Claude Code](https://claude.com/claude-code)

rajames added this to the v4.0.0-alpha.5: SDK milestone 2026-11-07 23:16:09 +00:00

rajames added the

labels 2026-11-07 23:16:10 +00:00

rajames referenced this issue from a commit

2026-11-07 23:19:31 +00:00

FABRIC-3 XLVIII: the review, the protocol, the SDK, the demo's limits