Security & Governance

Policy-checked. Evidence-recorded. Restore-protected.

The 7 security guarantees below are the contract between Sidian Workforce and the CISO. The 6 restore hard gates are hard-coded. The 7 audit-policy thresholds are operator-editable. The trust-lifecycle signs, expires, and revokes keys.

Local-first Provider-agnostic Evidence on every action Restore on every mutation Open department packs SOC 2 in progress ISO 27001 in progress HIPAA-bounded

The 7 security guarantees

Every guarantee is auditable. Every guarantee is enforced by the runtime. Every guarantee is a property of the platform, not a feature.

1
No raw prompts to the renderer
Only bounded, redacted, JSON-serializable safe DTOs cross the IPC boundary. Resume cursors stay internal. No raw provider output reaches the UI.
2
No raw memory or evidence in context
The cognition engine uses only bounded title + summary. Raw memory and evidence payloads never enter the model context window.
3
No destructive Git from the renderer
6 hard gates are hard-coded. The renderer cannot bypass them. The Restore Console is the only path to git reset --hard.
4
No secrets in any DTO
14-pattern secret scanner + 16-forbidden-key redaction layer. Backups cannot leak secrets. Credentials stay in safeStorage or env vars.
5
No autonomous action in finance, tax, legal, HR, payroll, or money movement
Gated by approval routes. CEO is the default fallback. The AI drafts; the human signs. Always.
6
No silent agent decisions
Every autonomous action emits a bounded, audit-logged event with source, severity, fingerprint, and a human-readable message. The chief chat runtime detects loops and stops.
7
No provider lock-in
Adapter pattern. Switch from OpenAI to Anthropic to a private LLM without rebuilding the workforce. Resume cursors are opaque.
Restore governance

6 hard gates. 7 audit-policy rules. 4 trust-lifecycle states.

Autonomous engineering work mutates files and opens PRs. Sidian takes this seriously.

1Request
2Dry Run
36 Hard Gates
47 Policy Rules
5Typed Phrase
6Approval
7Execute + Safety Ref

The 6 hard gates (never configurable)

  • Plan not expired — 5-minute TTL from built to expires
  • Typed phrase exact-match — STRICT case-sensitive, whitespace-exact, no normalization
  • Request is approved — via the existing approval route store + ApprovalEngine
  • Working tree is clean — re-checked at click time via getWorkingTreeCleanStatus
  • Git is a repo — verified via isGitRepo
  • Target commit is in history — verified via getHeadSha + isAncestorCommit

The 7 audit-policy rules (operator-editable)

  • Cooldown — default 1 hour
  • Restore loop — 24h window, threshold 3
  • High-risk target — requires delegated approval
  • Failed-retry — requires delegated approval
  • Missing safety ref — denied
  • Stale dry-run — 5-minute TTL
  • Executor-approver separation — approver cannot be the executor

The 4 trust-lifecycle states

  • Signing — Ed25519 signatures on every restore policy export
  • Expiration — keys expire; rotation is operator-driven
  • Revocation — CRL maintained; revoked keys never verify
  • Encrypted backup — PBKDF2 + AES-256-GCM with PBKDF2-derived key
Governance primitives

Approval routes. Evidence. Action log. Event store.

Approval routes

  • 12 built-in approval kinds (customer email, tax filing, code push, payroll, bank transfer, etc.)
  • Per-route, per-department, per-agent, per-work-run policy
  • Concrete human reviewers + company admin fallback
  • Priority-sorted; first-match wins

Evidence ledger

  • 34 typed evidence kinds
  • Bounded payloads, secret-redacted
  • Always cited from a work run
  • Always inspectable from the work run panel

Action log + event store

  • 88 action log types
  • 74 event types
  • Replayable event store
  • Exportable to JSON
Connector diagnostics

9 safe checkers. Never sends test data.

The 9 connector families

  • Git / GitHub — credentials, env vars, command availability, schema validity
  • Jira — credential presence and shape
  • Email (SendGrid) — credential presence and shape
  • Slack — credential presence and channel format
  • Web search — credential presence and dependency availability
  • Provider config — credential presence and provider health
  • MCP servers — command availability, transport reachability, schema validity
  • Local shell — command availability, permission scope, dependency
  • Ticket sync — credential presence and dependency

Diagnostics are read-only by design. They never call a third-party API, never read a credential value, never execute a shell command, never send a test email, and never push a PR. They check shape, presence, and reachability — and tell you, in one place, which connectors are configured and how to fix the ones that aren't.

What Sidian is not (yet)

Sidian is not FedRAMP authorized today. Sidian is not SOC 2 Type II audited today. Sidian is not a substitute for a HIPAA Business Associate Agreement with a covered entity. Sidian is a workforce platform; your security team is responsible for deploying it within your own compliance envelope. The Trust Center lists the current security posture and the roadmap.

Talk to security.

We'll show you the Trust Center, the 7 guarantees, the Restore Console, the audit ledger, and a deployment on your data with your model.

Talk to us See the CISO use case