Last updated: 2026-08-22
This policy covers the SAGG Managed Inference API (the "Gateway"), accessed via /v1/chat/completions and related endpoints, including when reached through a third-party aggregator (e.g. OpenRouter).
Operated as an individual/sole proprietorship based in British Columbia, Canada (see CONCEPT.md sec 12 for the payout/legal-entity decision, and section 6 below for jurisdiction).
The Gateway logs, per request: a request ID, timestamp, client identifier, the upstream provider used, token counts (prompt and completion), and HTTP status. This is implemented in cmd/gateway/main.go (logJSON calls) and is limited to operational metadata needed for billing, rate limiting, and failure diagnosis.
We do not retain, persist, or use prompt/completion content for model training. Request and response bodies pass through system memory only for the duration of a single request (or, for streaming responses, are relayed line-by-line as received — see cmd/gateway/billing.go) and are not written to persistent storage, logs, or any database by this system.
Requests are forwarded to one of two upstream inference providers: the Gonka network (via an access broker) as Primary, and DeepInfra as Backup. Their data handling guarantees are materially different and are stated separately below rather than as one blanket claim.
DeepInfra (Backup): DeepInfra's published data-privacy documentation (docs.deepinfra.com/account/data-privacy) states they do not use API-submitted data to train their models, except for Google or Anthropic models routed through their platform (not applicable to us — we only route DeepSeek through DeepInfra). DeepInfra additionally states a zero-retention policy and SOC 2 / ISO 27001 certification.
Gonka network structural fact (applies to any broker we use, including our Primary): Gonka's own official developer documentation states explicitly that brokers are independent third parties, not part of the core protocol, and that data handling is determined by each broker individually. Whatever a broker itself states, that layer sits in front of a decentralized network of independent, pseudonymous Hosts worldwide who physically run the inference — no broker is positioned to make or enforce a no-training guarantee reaching that layer, because no broker controls it. Encryption status of data in transit to an individual Host is not addressed in Gonka's public architecture documentation (verified 2026-08-22) — we found no statement either confirming or ruling out end-to-end encryption to the execution layer, and do not represent this either way pending clarification from the network itself.
Our Primary broker (proxy.gonka.gg, operated by "Gonka Labs" — verified 2026-08-22): Gonka Labs identifies itself by name (site footer: "Product by Gonka Labs," a stated team of engineers, contact email published) and states on its own interface that request/response payload text is not logged by default ("Private unless you say otherwise... disable privacy in Settings to review request text"). We found no dedicated Privacy Policy or Terms of Service document on proxy.gonka.gg or Gonka Labs' own site (gonkalabs.com) — /privacy and /terms both resolve to the homepage, not a distinct legal page — so this is a marketing-page statement, not a binding policy with stated governing law, data-subject rights, or a formal retention period. It also does not, and cannot, address the structural gap noted above: it describes Gonka Labs' own proxy/billing layer, not the pseudonymous Host network that actually executes the inference, which Gonka Labs does not control. No statement of any kind (formal or informal) about model-training use was found for either layer.
Compliance & enterprise suitability: SAGG optimizes routing for request cost and availability. While our Primary broker states that requests are not logged, the underlying infrastructure relies on a decentralized node execution network that Gonka Labs does not control, and for which the encryption status in transit to individual execution nodes is not publicly documented. Consequently, SAGG does not issue a unified, enterprise-grade Zero Data Retention (ZDR) or No-Training SLA for traffic routed through the Primary network — this is reflected honestly as compliance.zdr: false in our /models listing. Clients with strict compliance mandates (e.g. SOC 2, HIPAA, mandatory contractual ZDR) should evaluate workload applicability against this architecture, or use OpenRouter's own ZDR routing controls (account/guardrail/per-request) to restrict their traffic to certified-ZDR endpoints. SAGG's Backup path (DeepInfra) carries DeepInfra's own certified guarantees as stated above.
kept indefinitely by design. This is a running numeric balance (how much quota a client has left), not a per-transaction record — there is no way to reconstruct individual past requests from it.
used to cross-check invoices from partners such as OpenRouter — see cmd/gateway/openrouter.go) — 90 day TTL, reset to zero at the start of each billing month.
token counts, latency, timing/error flags — see section 1; never prompt/completion content) are maintained on a rolling buffer bounded by storage volume (500MB max, Docker's json-file driver, 50MB x 10 rotated files). Depending on actual API throughput, this bounds metadata log retention to an estimated range of roughly one to eight months: measured live traffic (2026-08-21) put usage at approximately 445 bytes/request, which at OpenRouter's own documented minimum traffic floor (100 requests/30 min sustained) yields ~245 days; materially higher post-launch volume than that floor — likely once real customer traffic arrives — will shorten this proportionally. Older entries are automatically overwritten once capacity is reached. We will publish a tighter figure once steady-state post-launch volume is observed.
We (SAGG) do not use request or response content to train, fine-tune, or otherwise improve any model — content is not retained past the request, per section 2. This is a claim about SAGG itself; it does not extend to upstream providers, whose own training-use policies differ and are stated separately in section 3.
This service is operated from Canada, British Columbia. Disputes and applicable data-protection obligations are governed by Canadian federal law (PIPEDA) and the law of British Columbia, including its Personal Information Protection Act (PIPA) where applicable to this service.
International users (GDPR): For users accessing the Service from the European Economic Area or UK, operational metadata (section 1) is processed on the basis of legitimate interests (Art. 6(1)(f) GDPR) — specifically, billing reconciliation, rate limiting, and abuse prevention. If you are an EEA/UK resident, you have rights under GDPR (access, rectification, erasure of metadata we hold) — contact us at the address in section 8. Note the limitation stated in section 3: SAGG cannot enforce erasure requests for content already dispatched to and processed by decentralized upstream execution nodes, as no persistent record or traceable identity of that dispatch exists at the Gateway level once the request completes.
This policy will be updated if our logging or retention practices change materially. Material changes will be reflected in this file's "Last updated" date and in the project's version history.
For privacy-related questions, data subject requests, or concerns about this policy, contact: dorofeevov17@gmail.com