Historical report imported 2026-10-10 from docs/experiments/benchmarks/alpha-phase2-delivery-run-01.md. Original SHA-256: c11c6f427e9c5ab1b3bc2ae4e6f00047c200cdb8fe8bc20ceded938c12abb90f. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Phase-2 delivery with ten real nodes, run 01

Status: all six registered trials passed on 2026-09-28 (plan step 2.6). One host; this is not the fleet tier gate.

Method

As registered in the protocol: ten fabric-node run processes, each offered 2 lines/s of 512-byte bodies and sampling host metrics every 15 s, one fabric-server, TLS on loopback, 15 s warmup plus 120 s measured, three seeds per commit mode. The binaries and harness were frozen from commit 71fef99e51695f3db594ac29ba26425783f121ab; their SHA-256 values are in hashes.txt (local evidence omitted from stage). The runner is the frozen phase-0 runner (SHA-256 2425e950…). A first attempt at this run did not start any trial: the runner refused an output path under a copied runner location, exit 2 for all six, and nothing was measured. For each mode M and N=1,2,3:

python3 -B tools/alpha/runner.py --out target/alpha-p2-delivery-M-seedN --duration-s 300 --disk-bytes 1073741824 --max-output-bytes 1048576 -- python3 -B target/alpha-p2-frozen/tools/delivery_tier.py --seed 0xA11FA00N --mode M --bin-dir target/alpha-p2-frozen

All six runner invocations exited 0 with passed=true. Compact summaries and runner results are in data/alpha-phase2/tier10 (local evidence omitted from stage).

Results

Mode	Seed	ACK p50 ms	ACK p99 ms	Max backlog	Server frames	Server CPU s	Server VmHWM KiB	Max node VmHWM KiB
grouped	1	57.7	64.3	0	136	0.33	9044	5420
grouped	2	55.6	60.5	0	136	0.32	8680	5424
grouped	3	57.5	63.4	0	136	0.34	8824	5428
individual	1	40.7	79.0	0	1360	0.81	8632	5404
individual	2	41.1	78.2	0	1360	0.81	8600	5420
individual	3	40.7	77.8	0	1360	0.81	8456	5428

Every trial recovered 1,360 batches, had 1,200 acknowledged attempts in the measured window, passed the delivery oracle, drained all offered lines, and every process exited 0.

Interpretation

Both modes meet the 1 s ACK p99 gate by more than an order of magnitude, with no backlog at this load. Grouping put about ten batches in each frame, so the server wrote one tenth of the frames and used about 40% of the CPU. The cost is the 50 ms grouping window, visible as a higher median. Individual commits had the lower median but a higher p99, because ten nodes polling on the same second queue behind one another's syncs. This is one host at about ten batches per second; it does not predict behavior at the fleet tiers.

Limits

All processes share one host and one disk, so network latency is loopback only. Node CPU includes each node's TLS client and one-second log polling.