Service Level

Downtime is the period during which the API endpoints (`/api/predict`, `/api/dual_predict`, `/api/calibrate`, `/api/warranty_floor`) return 5xx errors or fail to respond within 30 seconds for non-stress-test workloads.

Excluded: scheduled maintenance announced ≥48 hours in advance; outages caused by upstream provider incidents (Railway, Stripe, customer’s own network); customer- caused issues (rate limit, expired keys, malformed input).

For confirmed P1 incidents (full API outage or data-integrity issue):

• Initial acknowledgment within target response window for your tier
• Status updates every 60 minutes during the incident
• Post-incident report within 5 business days, including root cause and mitigation
• Security incidents: 72-hour breach notification per the Data Processing Agreement

Status page: https://cozy-hope-production-cb41.up.railway.app/api/health always returns the current model version and endpoint inventory. A dedicated status page is on the roadmap.

Each prediction response carries model_version and output_schema_version fields. Customers can pin to a specific model version by passing ?model_version=18.3 (or any prior published version) as a query parameter.

Version retention commitment:

• Starter / Free: 24-month retention from version deprecation announcement.
• Professional: 36-month retention from deprecation announcement, with documented sunset schedule and migration guide.
• Enterprise: 10-year retention contractually, with model weights + Docker container image archived at a third-party escrow (Iron Mountain or equivalent) accessible by the customer's legal team in the event of a subpoena, regulator audit, or vendor wind-down. Required for warranty-reserve, aerospace qualification, marine class-society (DNV-RP-0043), and insurance underwriting use cases.

Deprecation procedure: a model version moves to "deprecated" status only after (a) at least 12 months of advance notice via API response headers and the /docs changelog, and (b) a published migration guide demonstrating numerical reproducibility on the 8 calibrated reference datasets. Deprecated versions remain callable for the full retention window of the highest-tier customer using them; no version is removed while an active contract requires it.

If we miss the contractual uptime target in a billing period, the customer is eligible for service credits:

• Below 99.9% but at or above 99.5%: 10% credit on the affected period
• Below 99.5% but at or above 99.0%: 25% credit
• Below 99.0%: 50% credit
• Customer must request credits within 30 days of the affected period via support@scaleprognostics.com.

Honest disclosure of where the SLA does not extend:

• Free trial and Starter tiers carry no contractual uptime commitment. Best-effort means we operate the service in good faith but do not refund time loss.
• We do not provide phone support at any tier below Enterprise.
• Mirroring / on-prem deployment is Enterprise-only; air-gapped operation requires custom contract.
• Compliance certifications (SOC 2, ISO 27001) are on the roadmap but not yet held. Customers requiring these should engage Enterprise tier and discuss the timeline.

• Dedicated status page (target: Q3 2026)
• SOC 2 Type 1 attestation (target: Q4 2026)
• Multi-region API deployment for sub-100ms latency outside US (target: Q1 2027)
• SDK clients for Python (alpha published) and TypeScript (in design)

SLA discussions and Enterprise contracts: jason@scaleprognostics.com. Operational support: support@scaleprognostics.com.