Testing and Measurement Policy

This policy separates what a browser observation records from what an editorial claim would require. Measurement contract http-rtt-v1 describes new results with supporting evidence. It is not a certification of a service, device or connection.

1. Describe the endpoint honestly

The service uses Cloudflare Pages and Pages Functions for published test routes. That hosting arrangement does not establish independent physical test nodes, multi-CDN probing, a neutral peering location, resistance to traffic prioritization, or coverage of the route used by a particular game or call.

2. Name the actual statistic

• Throughput: an HTTP transfer-rate estimate for the configured download or upload phase, not a guaranteed access-line capacity.
• Idle latency: the worker's minimum HTTP completion-time estimate after warm-up, with the documented sub-millisecond substitution/floor; not a median or physical-distance measurement.
• Jitter: asymmetric smoothing of absolute differences between consecutive RTT inputs; not standard deviation, P90 or one-way delay variation.
• Loaded latency: a rounded arithmetic-mean download-phase delta from five accepted probes and a valid idle baseline. It is not a separate upload-loaded or bidirectional measurement.
• Grades: no new letter grade is assigned. Legacy grades do not establish the quality of an ISP or locate a router fault.

The methodology gives the exact calculation, warm-up treatment, timing basis and evidence rules. The version identifies behavior; it is not an independent accuracy audit.

3. Keep failed and missing observations visible

Only valid HTTP 2xx latency completions are admitted. Errors, timeouts, aborts, invalid timing and phase crossings cannot become successful probes. The engineering budgets are 2,000 ms per request, 6,000 ms for loaded sampling and 20,000 ms for the idle sequence. They bound resource use and may exclude slow connections; they are not service-quality standards.

Attempts and exclusion reasons remain in new snapshots. An incomplete loaded plan or invalid baseline produces no loaded delta. Missing or invalid core values must not be relabeled as genuine zero. User cancellation, offline events and stale callbacks must not create a result from another run.

4. Keep comparisons controlled

Record the device, browser, connection method, plan, time, endpoint if available, and relevant background activity. Compare a wired baseline with the problem Wi-Fi location as separate conditions. Repeat at different times without discarding failures or choosing only the best run.

Leave required work and security controls in place; record their presence rather than bypassing them. Tests transfer data, so consider metered allowances. A time-of-day pattern can justify investigation, but it does not identify congestion, throttling or a provider-wide defect by itself.

5. Separate diagnostic and editorial evidence

A device review requires an identifiable model, firmware, setup, dates and retained observations before it can be described as hands-on testing. A measured ISP comparison requires a defined population, collection period, eligibility rules and reproducible calculations. A browser observation or policy statement does not supply that missing evidence.

Commercial relationships are not measurement evidence. See the editorial policy and funding disclosure for publication standards.

6. Preserve the audit trail

New local history and exports retain the contract and available measurement evidence. Shared URLs can be edited by their sender, and local storage is not authenticated. Old records without supporting samples must remain identifiable as legacy; they must not receive fabricated evidence or a replacement date.

Method changes require review of code, explanations, result presentation and exports together. Material corrections should say what changed and why. Report a suspected error through the contact page with the affected URL, result context and a reproducible description, without posting private account information.

Editorial update: Earlier measurement claims were revised; unsupported claims remain withdrawn. See editorial corrections.