Transparency statement: what we publish and how the figures are produced
This page states the disclosure practice of the smaapi Gateway (www.smaapi.com, Slime Mould Tech): which data we publish, how those figures are produced, and what we explicitly do not promise. The hardest thing to verify during vendor selection is rarely the feature list — it is whether the numbers can be trusted. Our approach is to publish the methodology, sample counts, as-of date, and generation path alongside every figure, so readers can judge for themselves instead of taking the conclusion.
What data do we publish?
The measured dataset we currently publish is the model access index: probe availability and sample counts for each connected model x channel, listed row by row with an as-of date. This period contains 8 rows.
How are those figures produced?
- Measure: the production gateway sends a short probe to every connected model across every task domain, recording success and elapsed time.
- Export: results are aggregated into a per-model availability weighted by probe rounds, then mapped through de-identifying labels into public terms.
- Gate: models with fewer than 30 samples are excluded; rows missing the primary metric or carrying invalid field types are dropped as malformed; if nothing qualifies, the page falls back to a placeholder and publishes no figures.
- Inject: the page is generated by the build pipeline from the corresponding export rows — no human transcription step.
The chain has exactly one design goal: every public number must trace back to a real measurement. If the measurement was not taken, or the sample is too thin, the result is a blank on the page — not a rephrasing that makes the claim survive.
What do we explicitly not promise?
- Availability is not an SLA. Published availability is a probe-side measurement; service-level commitments are contractual.
- No absolute claims. We give no unconditional guarantee on service continuity or account standing — no party can honour that, and putting it in writing only backfires when something does go wrong.
- No claimed official partnerships. Vendor and model names are used nominatively; channel and authorization status are stated per page and fixed in contract.
- No substitute for your own compliance assessment. Cross-border data assessments and algorithm filings are obligations of the service provider under law; the gateway supplies technical means, not legal conclusions.
Which regulatory scope applies?
Article 2 of China's Interim Measures for the Management of Generative AI Services draws an explicit line on scope. Its third paragraph reads, verbatim:
"行业组织、企业、教育和科研机构、公共文化机构、有关专业机构等研发、应用生成式人工智能技术，未向境内公众提供生成式人工智能服务的，不适用本办法的规定。"
That distinction is what separates internal enterprise use from public-facing service: using LLMs inside a company's own operations and offering generative AI services to the public in mainland China fall under different regulatory treatment. SMA can route public-facing traffic to registered domestic models while overseas models serve internal scenarios. Whether a given workload counts as "providing services to the public" is a determination each enterprise must make with its own counsel — we do not make that call on your behalf.
How to report a problem
If any figure here conflicts with reality, or any wording reads as misleading, tell us — corrections to public figures are published together with the methodology note explaining what changed. Every page carries a last-updated date and the data page carries an as-of date, so currency can be checked directly.
Frequently asked questions
Can the published availability be treated as an SLA?
No. Published availability is a measured statistic from active probing — it reflects connectivity from the probe's vantage point and is not a service-level commitment. SLA terms are fixed in the enterprise contract. We publish the methodology, sample counts, and as-of date alongside the numbers precisely so readers can judge what the figures do and do not cover.
Why are some metrics left blank?
Because there is no real measurement to publish yet. Latency percentiles (P50/P95) require per-request latency records; the current probe only has means, and presenting a mean as a percentile would be false. So the column stays blank. The rule is simple: missing data is left missing, never filled with estimates or industry averages.
Could the figures on these pages have been typed in by hand?
No. The data page and every measured figure quoted elsewhere are injected by the build pipeline from the gateway export file, so any number traces back to a specific row of that export. The export side is fail-closed as well: rows below the sample threshold, missing the primary metric, or carrying wrong field types are dropped, and if nothing qualifies the page publishes no figures at all.
What is your relationship with the model vendors?
We integrate and govern access to model capabilities; we are not a model vendor. Vendor and model names on these pages are used nominatively to identify what is being integrated, and do not imply endorsement by, or an official partnership with, those vendors. Channel and authorization status are stated on the relevant pages and fixed in contract.
References
- Interim Measures for the Management of Generative AI Services (Cyberspace Administration of China) — first-hand source on scope and obligations
- Anthropic API documentation — authoritative reference for Claude access
- AWS Bedrock documentation — the commercial cloud platform path for model access
Trademarks: Claude belongs to Anthropic, AWS and Amazon Bedrock to Amazon, Vertex AI to Google. Names are used nominatively on this page and do not imply endorsement by, or an official partnership with, those holders. Last updated: 2026-07-25.
About the name: smaapi (the SMA gateway) is the enterprise AI gateway built by Slime Mould Tech — SMA stands for Slime Mould Architecture. smaapi is unrelated to the simple moving average indicator in finance, to the solar inverter vendor SMA Solar Technology AG, or to the SMA coaxial connector standard that share the acronym.