Real-network latency — RAIN rng-node on a remote host, pipelining OFF vs ON (measured 2026-09-14)

Closes review round 1 item 3 ("pipelining measured on a real network, not loopback"). Raw data: results.json (60 ms gap, the headline run), results-gap0.json (no animation gap — worst case), results-gap1000.json (realistic 1 s reel animation), loadtest-*.txt (console output), network-baseline.txt (DNS/TLS/TCP probes), RUN.txt (hosts, command lines), deploy-manifest.txt (sha256 of every vendored dist file on the server = commit 411e04e3 + this leg's tools), server.mjs + deploy-package.json (exactly what runs on the server).

1. Set-up

2. Network baseline (what the wire costs before RAIN does anything)

So the application RTT client ↔ RAIN node ≈ 95–100 ms (Virginia → Cloudflare → Oregon → back; a transcontinental path, i.e. a pessimistic placement compared with an operator who co-locates within one region). Tail spikes to 250–300 ms on the baseline are edge/origin scheduling noise on a starter instance and appear in every series.

3. Results — 1,000 spins per run, 60 ms animation gap (results.json)

Reading it. With pipelining OFF the spin costs exactly one WAN round trip (p50 88–91 ms ≈ the measured 95–100 ms baseline minus TLS/keep-alive savings), with very low jitter (WS IQR 0.6 ms). With pipelining ON and only a 60 ms gap, the prefetch has ≈ 60 ms of the ≈ 90 ms round trip covered, so the critical path is the remaining ≈ 30 ms (p50 36 ms HTTP / 23 ms WS) — i.e. critical path ≈ max(0, RTT − animation), precisely the model in whitepaper §5.2. p95 drops from 102 → 49 ms (HTTP) and 100 → 40 ms (WS); p99 from 119 → 68 ms and 102 → 58 ms. The max values (210–330 ms) are single edge/origin hiccups that hit both modes alike (also visible in the baseline).

3.1 Sensitivity to the animation length (results-gap0.json, results-gap1000.json)

(The gap-0 and gap-1000 runs are appended to this file by the runner when they finish; if a cell above still says "see …" the text file holds the numbers.)

4. What this shows, and what it does not

• Shows: on a real transcontinental path (≈ 95 ms application RTT, Cloudflare edge + Render origin), the unmodified 2.1 ceremony settles 1,000 consecutive rounds per run without a single error over both HTTP and WS, and pipelining removes exactly the overlapped portion of the WAN round trip from the player-visible critical path; with a realistic ≥ 1 s animation the network disappears from the critical path entirely. The whitepaper claim "the ceremony hides inside the reel animation as long as the animation is longer than the RTT" is confirmed on a real network — and the 60 ms case shows the honest boundary when it is not.
• Does not show: concurrent-player throughput (one sequential client), Postgres-backed persistence latency (memory store on the server), a same-region deployment (the ≈ 95 ms RTT is a pessimistic coast-to-coast placement; an operator would co-locate and see ≈ 10–30 ms), or browser → operator → node chains (the reference deployments' 92–104 ms p50 figures in the whitepaper cover that leg separately).
• Cost / teardown: one Render starter web service (~$7/month, billed by AgentBase credits). Left running for re-runs by the reviewer until the review closes; see RUN.txt for the URL — to be deleted afterwards (recorded in REVIEW-RESPONSE-R1.md).

5. Reproduce

# server: deploy packages/{rng-core,rng-session,rng-node}/dist with server.mjs (this directory) as a Node web service listening on $PORT,
#         env RAIN_STORE=memory RAIN_OPERATOR_ID=rain-lab-latency RAIN_CHAIN_LEN=4096  (or: docker build -f packages/rng-node/docker/Dockerfile .)
# client:
node examples/slot-spin-endpoint/loadtest-remote.mjs https://<your-node> 1000 60   --json results.json
node examples/slot-spin-endpoint/loadtest-remote.mjs https://<your-node> 1000 0    --json results-gap0.json
node examples/slot-spin-endpoint/loadtest-remote.mjs https://<your-node> 300  1000 --json results-gap1000.json