Built to measure waste — not to become it.

Your prompts, outputs, and ingested traces are used only to run the service for your workspace. We do not train models on your data and we do not sell it. Every database read is scoped to your organization ID; the only cross-tenant surface is the opt-in, k-anonymized industry benchmark, which publishes aggregate percentiles only once enough distinct organizations contribute that no single one can be re-identified.

Trace ingestion stores metadata only — model, token counts, cost, and latency — which is all the savings engine needs. Raw prompt/response text is stored only when you explicitly opt in per event, is redacted after a short window (30 days by default), and whole trace records age out on a configurable retention schedule.

Evaluations run on your own provider keys, encrypted at rest with AES-256-GCM using version-prefixed ciphertexts (v1) so encryption keys can be rotated without downtime. ReasonRank can never overspend them: every run has a pre-flight cost estimate, per-run caps, a monthly budget, an output-token ceiling, live spend tracking, and a kill switch. The optional gateway is self-hosted and single-tenant only — we never proxy your provider traffic through shared infrastructure.

You are not locked into ReasonRank's prompt runner. Agents can run in provider mode (we call models with your BYOK), HTTP endpoint mode (HMAC-signed POST to your HTTPS agent), or external mode (your CI/VPC worker submits cell outputs via API). All three feed the same score → recommend → CI gate → verify loop. Integration contract: BYO keys & harness.

We never repoint a production model on a hunch. A recommendation is only surfaced when a statistical non-inferiority test on real score samples shows the cheaper model is not meaningfully worse — with the confidence interval and sample size shown. After you apply one, a verification loop measures realized cost and quality from live traffic and auto-flags regressions with one-click rollback.

Every workspace is isolated by construction: access is scoped on every query, role-based permissions govern who can do what, sensitive actions are audit-logged (viewable and exportable from Settings → Activity log), and secrets are encrypted. Web sessions idle out after 24 hours with a 7-day absolute lifetime; API tokens are scoped, expire after a year by default, and rotate in one click. Access to production systems is restricted and audited.

Data is stored in Postgres with point-in-time recovery. Webhooks are idempotent, runs are reclaimed via heartbeats if a worker dies, and background jobs (verification, drift detection, retention) run on a scheduled, monitored cron. Current status: /status.

We follow SOC 2 aligned practices today (audit logging, RBAC, encryption at rest, least-privilege access, incident response) and are on a path to formal SOC 2 Type II. Download the security packet above for our CAIQ-lite answers, DPA template, and subprocessor register — or contact sales for a custom questionnaire review.

Export or delete your data at any time from the app; a workspace owner can permanently delete the whole workspace. For data-subject requests, email privacy@reasonrank.ai and we will confirm completion in writing. See our Privacy Policy for details.