# Capslane native transport benchmark

Published September 6, 2026. Observed production run: September 3, 2026.

Download these files into the same directory:

- [Corpus: 200 public video IDs](https://capslane.com/reliability/corpus-2026-09-03.json)
- [Versioned Node.js runner](https://capslane.com/reliability/benchmark-2026-09-03.mjs)
- [Historical observations](https://capslane.com/reliability/results-2026-09-03.json)

The runner matches the historical implementation except for the relative corpus file path. It records HTTP status and time until headers for successful requests; error bodies may add time on failures. It does not validate successful response bodies or measure transcript accuracy. There is no per-request client timeout in this historical runner. A stalled request can require interrupting the process. Do not use it as a production application client.

## Reproduce a small sample

Requires Node.js 22 or newer and a Capslane API key supplied through the CAPSLANE_API_KEY environment variable. Do not paste a real key into a committed file. This shell example checks five fixtures:

```sh
RELIABILITY_FIXTURES=native-public-001,native-public-002,native-public-003,native-public-004,native-public-005 \
RELIABILITY_OUTPUT=sample.json \
node benchmark-2026-09-03.mjs
```

PowerShell:

```powershell
$env:RELIABILITY_FIXTURES = 'native-public-001,native-public-002,native-public-003,native-public-004,native-public-005'
$env:RELIABILITY_OUTPUT = 'sample.json'
node benchmark-2026-09-03.mjs
```

Omitting RELIABILITY_FIXTURES runs all 200 fixtures and requires enough remaining transcript requests. Accepted cached requests also count. This native-only corpus starts no audio generation. Default concurrency is three with a 1.6-second delay between requests per worker. Do not increase concurrency to evade plan limits.

The default minimum success rate is 98 percent. A lower observed rate makes the process exit with code 1 after writing the complete report. Preserve failed fixtures in any publication. Historical results are not predictions: video availability, upstream behavior, region, cache state and deployed code can change.

The September 3 run used the production API with the shared cache enabled. It was not a cold-cache test and does not compare providers. Median and p95 summarize positive recorded latencies using nearest-rank percentiles; network errors are excluded from latency percentiles but retained as failures. Results include no transcript text.

The corpus lists public videos with caption metadata at selection time. It is biased toward videos expected to have captions; it does not measure arbitrary-video coverage. Audio generation and worker isolation checks on the reliability page are separate tests, not part of this corpus.

The public files contain video identifiers and transport measurements, not video media or transcript text.