The committed evidence is in results/: 1,000-event pressure smoke, durable pipeline worker scaling, decoder/pack/normalizer microbenchmarks, and the 10,000-event all-format pressure run. These were captured on an Apple M2 with 16 GB RAM running Darwin arm64 on 2026-09-24. The 1,000-event smoke run reported 4,623.1 events/s, p50 166 µs, p99 1.542 ms; the concurrent 10,000-event all-format run reported 10,000 normalized, zero quarantined/pending, 22.096 s (452.6 events/s), and queue-inclusive p99 21.691 s. These measure different workloads; neither proves sustained or multi-node capacity.
To reproduce on your machine, use Go 1.25+ from a checkout, keep your machine otherwise idle, record git rev-parse HEAD, git status --short, go version, OS/CPU/RAM/storage, active packs/sinks, start/end UTC, and the exact commands. Save new logs under ignored data/benchmarks/ so the published evidence is not overwritten. The benchmark harness builds its own temporary store; live methods must use a fresh -db path for each run. Report accepted, normalized, quarantined, pending, producer acknowledgements, output counts, and errors together with elapsed time. accepted = normalized + quarantined + pending is lifecycle accounting, not exactly-once external delivery.
This Go test submits seven format families from eight producers to a durable SQLite pipeline, drains it, and records queue-inclusive latency. It does not include a network listener.
PowerShell:
New-Item -ItemType Directory -Force data/benchmarks | Out-Null
$env:WORM_RIGOROUS_BENCH = '1'
$env:WORM_BENCH_EVENTS = '10000'
go test -count=1 -timeout=10m -v -run '^TestRigorousPressureMatrix$' ./test/pressure 2>&1 | Tee-Object data/benchmarks/rigorous-10k.txt
Remove-Item Env:WORM_RIGOROUS_BENCH, Env:WORM_BENCH_EVENTS
Bash:
mkdir -p data/benchmarks
WORM_RIGOROUS_BENCH=1 WORM_BENCH_EVENTS=10000 go test -count=1 -timeout=10m -v -run '^TestRigorousPressureMatrix$' ./test/pressure 2>&1 | tee data/benchmarks/rigorous-10k.txt
Pass if events=10000, normalized=10000, quarantined=0, pending=0, and PASS appear. The throughput is a saturated single-store run, not an EPS promise.
This compares worker configurations and allocation cost. -benchtime=10000x runs 10,000 iterations per sub-benchmark, not one shared 10,000-event run. The committed scaling result used Go’s normal benchmark calibration, so its exact ns/op need not match a fixed-work rerun.
go test -count=1 -run '^$' -bench '^BenchmarkPipeline_E2E$' -benchtime=10000x -benchmem ./test/pressure
For the decoder/pack measurements in results/microbenchmarks.txt:
go test -run '^$' -bench . -benchmem ./internal/decode ./internal/packs ./internal/normalize
This includes the simulator, OS pipe, CLI framing, SQLite, and NDJSON output. It is a single-producer correctness run, not maximum throughput. In PowerShell from the repository root:
New-Item -ItemType Directory -Force bin | Out-Null
go build -o bin/worm.exe ./cmd/worm
go build -o bin/logsim.exe ./cmd/logsim
$run = "data/benchmarks/stdin-$(Get-Date -Format 'yyyyMMdd-HHmmss')"
New-Item -ItemType Directory -Force "$run/sources", "$run/sinks" | Out-Null
$elapsed = Measure-Command {
.\bin\logsim.exe --source app --scenario normal --format stdout --rate 0 --count 10000 |
.\bin\worm.exe -stdin -stdin-framing line -db "$run/worm.db" -packs-dir packs -sources-dir "$run/sources" -sinks-dir "$run/sinks" -workers 4 -syslog-udp none -syslog-tcp none -syslog-tls none -http none -ui none -inbox none -output-file "$run/normalized.ndjson" -stdout=false 2>&1 |
Tee-Object "$run/worm.txt" | Out-Host
}
"elapsed_seconds=$($elapsed.TotalSeconds)" | Set-Content "$run/duration.txt"
Select-String -Path "$run/worm.txt" -Pattern 'Accepted:|Normalized:|Quarantined:|Pending:|Loss Audit:'
Check Accepted: 10000, Normalized: 10000, Quarantined: 0, Pending: 0, and a passing loss audit. This method has no committed 10k result; do not attribute the Method A measurement to it.
This exercises a real listener and HTTP acknowledgements. The simulator sends sequential app JSON requests; acked=10000 is the producer’s count of successful responses, not proof that downstream sinks committed 10,000 records. First build bin/worm.exe and bin/logsim.exe as shown in Method C. Use two PowerShell terminals and a new run directory each time.
Terminal 1:
$run = 'data/benchmarks/http-10k-01'
New-Item -ItemType Directory -Force "$run/sources", "$run/sinks" | Out-Null
.\bin\worm.exe -db "$run/worm.db" -packs-dir packs -sources-dir "$run/sources" -sinks-dir "$run/sinks" -workers 4 -syslog-udp none -syslog-tcp none -syslog-tls none -http 127.0.0.1:18080 -ui 127.0.0.1:19090 -inbox none -output-file "$run/normalized.ndjson" -stdout=false 2>&1 | Tee-Object "$run/worm.txt"
Terminal 2:
$run = 'data/benchmarks/http-10k-01'
$elapsed = Measure-Command {
.\bin\logsim.exe --source app --scenario normal --format http-post --target http://127.0.0.1:18080/api/v1/ingest --rate 0 --count 10000 2>&1 |
Tee-Object "$run/sender.txt" | Out-Host
}
"sender_elapsed_seconds=$($elapsed.TotalSeconds)" | Set-Content "$run/duration.txt"
do {
$stats = Invoke-RestMethod http://127.0.0.1:19090/api/v1/stats
if ($stats.pending -ne 0) { Start-Sleep -Seconds 1 }
} while ($stats.pending -ne 0)
$stats | Select-Object accepted,normalized,quarantined,pending,delivered,delivery_pending,delivery_failed
Select-String -Path "$run/sender.txt" -Pattern 'Finished logsim:'
Pass only if the sender reports emitted=10000 acked=10000 errors=0 and the final API reports accepted=10000 normalized=10000 quarantined=0 pending=0. Stop WORM with Ctrl+C after retaining the stats. The sender’s elapsed time excludes subsequent drain; publish both sender time and final-drain time if you report complete-path latency. No Method D 10k result is committed today.
For Linux/macOS, build bin/worm and bin/logsim without .exe, use shell variables and tee, and query the stats endpoint with curl. Keep Method A–D results separate: they cross different boundaries and cannot be ranked by their EPS alone. A 10,000-event burst is not a sustained one-billion-events/day validation; that requires a measured multi-node, real-broker/sink soak with bounded backlog and recovery evidence.