BonVoyage benchmarks.
With the context attached.

CAMPAIGN 03 / V1.0.4 / NOVEMBER 1-4, 2026

At 600K, the difference was tail stability.

The post-release replay bound the public v1.0.4 package to its exact binary, matched the TLS, payload, connection intent, upstreams, transformations, client lanes and 24-CPU application allowance, and rotated run order. At the supplemental 600,000 offered-RPS point, both systems returned every measured request successfully. BonVoyage stayed inside the 500 ms p99 gate in all three repetitions; NGINX stayed inside it in one of three.

How the comparison was constructed.

Four separate shared AWS c7a.8xlarge roles isolated traffic generation, the active gateway, deterministic upstreams, and evidence collection. Only one gateway ran on the target at a time. Application work used logical CPUs 8-31, while housekeeping and IRQ work remained on 0-7. The product order rotated between repetitions.

How to read the CPU number.

Both products received the same 24-CPU allowance: 24 BonVoyage reactor shards or 24 NGINX workers. App CPU is accumulated application CPU-seconds divided by wall time. It reports average compute consumed, not how many workers ran, how many CPUs participated, or how much useful work each CPU completed.

At 480K, both products met the traffic and tail gates, so NGINX consuming less measured application CPU is a valid advantage for that exact operating point. At 600K, NGINX still consumed less application CPU, but two of its three repetitions missed the p99 gate. The lower number therefore cannot establish greater efficiency at that point. This campaign did not profile NGINX scheduler wait, locks, internal queues or per-worker saturation, so it does not claim a root cause for the tail widening.

Why these load points were used.

Fixed offered rates show how latency and resources change as pressure rises without selecting a convenient peak after the fact. The frozen campaign used 20%, 40%, 60% and 80% of a 600,000-RPS shared-VM envelope. C03 executed 20%, 40% and 80%; its unused 60% point is reported rather than silently filled in. The separate 100% run was added only as a tail diagnostic.

At 480K, both systems were stable and the trade-off was different.

The frozen 80% point is important context for the high-load result. NGINX retained slightly lower primary latency and consumed less measured application CPU. BonVoyage retained materially lower resident memory and fewer descriptors. A fair report needs both observations.

The retained runs, replayed through the observability cockpit.

Each view was generated from checksum-verified per-run summaries labeled shared-vm-directional. They make repetition and gate outcomes easier to inspect without replacing the tables or retained evidence authority.

These are local Grafana visualizations of retained AWS result summaries, not live AWS telemetry or raw time-series samples from the original load windows.

What did the actual combined policy chain cost?

D03 measured BonVoyage against its payload-matched control at 24,000 offered RPS. The control established the cost of the same transport and payload without the composed policy path. The policy profile then exercised twelve components together, including RS256 JWT, source-IP admission, CORS, content guards, local admission controls, header transformations, rewrites, retry budget and passive endpoint health. It did not estimate the result by adding separate policy tests.

Expected result: positive traffic had to reach the upstream with the exact transformations, while missing JWT, denied source, invalid content type and preflight cases had to return their exact bounded outcomes. Why it was accepted: All three control and all three combined-policy measurements reconciled, every warmup and measured request succeeded, all six positive and negative semantic cases passed, primary tails stayed inside the gate, and no gateway restart occurred.

What outcome can an operator take from this?

Evidence class: Shared-EC2 directional evidence. Shared EC2 does not prove dedicated-host, NUMA or physical-NIC behavior. The 600K point is a supplemental diagnostic outside the frozen 20/40/60/80% curve. p99.9 remains diagnostic; p95 and p99 control the tail gate. The public package is a sanitized derived summary, not the raw evidence authority.

HISTORICAL EVIDENCE / SEPTEMBER 18, 2026

The earlier campaigns remain available, but separate.

The September tables belong to older BonVoyage builds and different test questions. They remain public for transparency and historical comparison, but they are no longer mixed into the v1.0.4 performance story.

What we are not claiming.