SLA & Compliance
1) Service Levels (SLA)
1.1 Availability Targets (Gateway/API)
Monthly SLO
Standard: 99.9% / month
Enterprise (contracted): 99.95% / month
Computation: Minute-level (UTC), excluding announced maintenance windows and issues not attributable to Sight (e.g., user-side networking, upstream-enforced blocks).
Service credits (illustrative):
< 99.9% and ≥ 99.0% → 5% of that month’s fees
< 99.0% → 10% of that month’s fees
1.2 Performance & Success Rate (SLO)
TTFB (first byte): For “small requests” (prompt ≤ 2 KB, expected output ≤ 1K tokens), p95 ≤ 2s (excluding cross-region network effects and upstream queuing).
Request success rate: For standard routing, ≥ 99% with automatic fallback/retry included.
Measurement: Based on Sight’s edge/gateway distributed tracing and health probes; an aggregated monthly snapshot is provided.
1.3 Rate Limits & Fair Use
Default quotas and rate limits are shown in Console and can be raised on request. In cases of strategic bursts (e.g., abuse-like retry storms), we may throttle to protect overall stability and will notify you to adjust.
1.4 Incident Response & Support
P0 (global/core outage): acknowledge within 30 minutes, continuous updates
P1 (regional/single-model severe impact): acknowledge within 2 hours
P2 (functional issues/perf degradation): respond within 1 business day
P3 (general inquiries/feature requests): reply within 2–3 business days
2) Privacy & Data Processing (Compliance)
2.1 Data Classes & Defaults
Business data: prompts/outputs, function-call params, file metadata, etc.
Technical metadata: timestamps, durations, usage, routes, error codes, etc.
Not used for training: Sight does not use your API traffic to train models. Upstream model retention/training follows each provider’s policy (see §2.4).
Zero-Data Retention (ZDR) option: You may restrict routing to ZDR upstreams (providers that commit to no request-body storage), consistent with OpenRouter’s ZDR concept.
Minimal logs: Gateway stores only usage and error summaries needed for billing/audit. With ZDR-only routing, we do not persist content bodies in the forwarding plane.
Optional extended retention: For audit/reconciliation, configure longer retention in Console (necessary fields only). You may shorten or purge at any time.
Prompt/response plaintext: Not persisted by default. If you enable “debug collection” or “evaluation pipelines,” we label it clearly and offer one-click purge.
2.3 Transport & Storage Security
In transit: TLS end-to-end; mutual TLS available for Enterprise.
At rest: Billing/audit data is encrypted; BYOK / Share-Your-Key uses client-side encryption + short-lived decryption at the executor with no disk writes (see Provider guide).
Secrets: Upstream keys are never exposed in plaintext to the platform; executors decrypt in a trusted environment (e.g., TEE, where applicable) and then zeroize.
2.4 Upstream Provider Policy Mapping (summary)
OpenAI API: By default, API data is not used for training; short-term retention typically for abuse/fraud controls (per provider docs).
Anthropic (consumer products): Opt-in sharing can enable training with retention up to multiple years; if not opted in, defaults to short-term retention. Enterprise/API channels via Bedrock/Vertex may have different terms.
Your Sight routing can avoid channels that don’t meet your compliance bar.
2.5 Data Residency & Cross-Border
Region-aware routing on a best-effort basis. When using regional upstreams (e.g., Azure OpenAI), we follow the upstream’s residency guarantees.
2.6 Rights & Agreements
DPA/SCC: Available for Enterprise.
Data subject rights: Access/export/delete of your account data and audit records (within legal constraints).
Subprocessor transparency: We maintain a list of upstreams and critical infra subprocessors and notify upon material changes.
3) Audit & Verifiable Settlement
Periodic snapshots: Weekly/monthly packages include the price table + detailed records + Merkle Root + signatures. Any party can recompute and verify offline (see the Settlement & Audit page).
Reconciliation deltas: Negative adjustments appear in the current/next snapshot with reasons.
This verifiable audit mechanism goes beyond simple invoice export common to many forwarding platforms.
4) Acceptable Use (AUP) & Content Restrictions
You must comply with the terms and policies of each upstream model/platform (e.g., OpenAI usage policies). Requests violating upstream policies may be blocked or removed from routing.
5) Security & Disclosure
Vulnerability disclosure: We welcome responsible disclosure via our security contact (see Console). We aim to provide a remediation plan or explanation within 90 days.
Penetration testing: Annual penetration testing and remediation validation can be appended to Enterprise contracts.
Key rotation: Periodic rotation and TTL supported; anomaly detection can automatically disable suspicious channels.
6) Changes & Notifications
This page may evolve with regulations, upstream policy changes, and platform capabilities. Material changes are announced via Console and email.
You can toggle ZDR-only, regional constraints, and upstream allow/deny lists in routing settings at any time to align with current compliance requirements.
Last updated