Historical report imported 2026-10-10 from docs/experiments/benchmarks/stress-run-01.md. Original SHA-256: a3eeb5036a7677b4c8495b9a67415b0abe8cb9f04dea4a1042f4d6e0d239a7ef. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Burst, rejection and concurrent management, run 01

Status: the three trials registered in revision 2 of the stress protocol ran on 2026-09-28, and each passed every gate. The state is Measured and passing under revision 2 on a four-CPU host. This is not a target-profile qualification: revision 1 on the target host has not run.

Method

The run followed revision 2 as registered.

• Workload. 1,000 identities ran the frozen workload for 180 s, with five times the log rate from second 60 to second 79.
• Rejection probes. 200 rounds ran concurrently, each sending a revoked token, an unknown token and a malformed body, all from fresh node identities.
• Management. Every 0.5 s, the harness sent an inventory request, a configuration change and a 30 s log query.
• Placement. The server ran on CPUs 0-1 and the simulator on CPUs 2-3.
• Artifacts. Binaries, examples and harness were frozen from 2b5c939 into target/alpha-p5r2-frozen, with 28 SHA-256 values (local evidence omitted from stage); the binaries are byte-identical to the outage and history sets. sha256sum -c after the soak found 0 mismatches.
• Host. The host was otherwise idle.

For each seed N:

python3 -B tools/qualification/runner.py --out target/alpha-p5r2-stress-seedN --duration-s 1500 \
  --disk-bytes 5368709120 --max-output-bytes 1048576 -- \
  python3 -B $PWD/target/alpha-p5r2-frozen/tools/stress_tier.py --seed 0xA11FA00N \
  --bin-dir $PWD/target/alpha-p5r2-frozen --server-cpus 0-1 --sim-cpus 2-3

All three runner invocations exited 0 with passed=true, no stop reason and cleanup confirmed; each took 204 to 206 s. The summaries, runner results and simulator summaries are in data/delivery-recovery/stress (local evidence omitted from stage).

Results

Seed	Batches created / acknowledged	Backlog at 55 s, 79 s, 100 s, 175 s	Rejections: revoked, unknown, malformed	Rejected batches committed	Management ok / failed	Concurrent query p99 ms	Server VmHWM KiB
1	180,000 / 180,000	0, 0, 0, 0	200× 401, 200× 401, 200× 400	0	265 / 0	369	652,668
2	180,000 / 180,000	0, 0, 0, 0	200× 401, 200× 401, 200× 400	0	268 / 0	324	656,244
3	180,000 / 180,000	0, 8, 0, 0	200× 401, 200× 401, 200× 400	0	268 / 0	329	663,356

Gate	Result
The delivery oracle passes over the simulator transcript and server_dump	passed in all three trials, with 0 violations
Revoked and unknown tokens always get 401; malformed bodies get 400 or 413	passed: every answer was 401 or 400
No batch from a rejected identity is committed	passed: 0 committed
The backlog at 175 s is at most 1,000 batches above the backlog at 55 s	passed: every backlog at 175 s was 0
Server VmHWM is at most 2 GiB	passed: the largest was 648 MiB
Every management call succeeds	passed
Both processes exit 0	passed

Interpretation

On two server CPUs the server absorbed the burst with essentially no queueing. The largest sampled backlog was 8 batches, at the end of the burst in seed 3. Rejections were clean and consistent: 401 before any sequencing for revoked and unknown credentials, and 400 for bodies that are not a Batch. No rejected identity reached the journal.

Deviation from the registered report

The protocol says latency during the burst is reported, not gated. The registered harness computes no burst-window latency: its summary gives only the backlog at four instants and the concurrent query p99. That metric is therefore not reported. Adding it now would change the harness after the run. The backlog at the end of the burst (79 s) stands in for it. The harness has since been amended to report the metric in later runs.

Limits

• The run used one host, loopback TLS and simulated identities.
• A full queue (503) is covered by unit behavior and is not forced here.
• Management ran about 1.5 times per second, not twice, because each round waits for its three calls.