Security & Data Handling
How EvalGate protects your data and keeps AI release decisions auditable throughout the evaluation lifecycle.
Release evidence
Gate results, baseline diffs, judge provenance, and reviewer actions stay attached to release decisions.
Data boundaries
Org isolation, scoped API keys, PII policy, and retention controls define what data can move and for how long.
Control plane
Provider allowlists, model policy, cost caps, and audit logs run before external judge execution is trusted.
Release Evidence Checklist
EvalGate treats security and governance as artifacts reviewers can inspect at merge time, not as a separate spreadsheet after the fact.
Baseline change
- Evidence
- Diff against the run or baseline used for the gate
- Reviewer question
- Did quality move, and was the movement intentional?
Judge configuration
- Evidence
- Model, rubric, parser, threshold, cost, latency, and retries
- Reviewer question
- Can this score be reproduced and trusted?
Promotion workflow
- Evidence
- Candidate case status, approver, timestamp, and failure mode
- Reviewer question
- Was production signal reviewed before becoming coverage?
Org policy
- Evidence
- Provider allowlist, PII setting, retention state, and audit entry
- Reviewer question
- Did the gate follow the org's release policy?
| Control | Evidence | Reviewer question |
|---|---|---|
| Baseline change | Diff against the run or baseline used for the gate | Did quality move, and was the movement intentional? |
| Judge configuration | Model, rubric, parser, threshold, cost, latency, and retries | Can this score be reproduced and trusted? |
| Promotion workflow | Candidate case status, approver, timestamp, and failure mode | Was production signal reviewed before becoming coverage? |
| Org policy | Provider allowlist, PII setting, retention state, and audit entry | Did the gate follow the org's release policy? |
Encryption
- All data encrypted in transit via TLS
- Secrets encrypted at rest using provider-managed encryption
- API keys stored as SHA-256 hashes — never in plaintext
- Provider keys encrypted using AES-256 via application-managed encryption keys
- Model-backed workflows are BYOK; the connected provider bills inference usage directly
Data Retention
- Account data retained for 90 days after deletion
- Trace data: 90 days (free tier), 1 year (paid tiers)
- Evaluation results: duration of subscription + 90 days
- Share links: configurable expiry with a 7-day default
PII Protection
- Automatic PII redaction in shared exports — emails, SSNs, phone numbers, credit cards, and API keys
- Redaction enabled by default; opt-out available for org admins
- Export size limits (10 MB) to prevent bulk data extraction
Audit Trail
- Full audit trail for critical mutations
- Immutable audit log entries
- Admin-only read access to audit logs
Access Control
- Multi-tenant architecture with organization-scoped data isolation
- Role-based access control: viewer, member, admin, owner
- Scope-based API key authorization with wildcard rejection
Compliance Framework Support
- Governance policy templates available: BASIC, SOC2, GDPR, HIPAA, FINRA_4511, PCI_DSS
These are policy templates for evaluation governance, not certifications.
Runs on infrastructure providers that support SOC 2 reports (available on request).
Rate Limiting
- Tier-based rate limiting: 1,000,000 requests/hour for anonymous traffic and up to 5,000,000,000 requests/day for authenticated plans
- Redis-backed sliding window algorithm