Defensible, not aspirational.

Tamper-evident signatures

Every signed approval includes a SHA-256 hash of the serialized record at signing time. Any subsequent edit detaches the signature from the snapshot and is visible in the audit trail.

Password reverification on every signature

The caller is reauthenticated against their current password at the moment of signing - not the session. Failed password returns 401 without leaking whether record or credential was at fault.

Append-only audit trail

Approvals, revisions, and follow-through are written to an audit log in the same transaction as the signed record. Export as CSV at any time.

Role-based approval chains

Chains are defined per record type and enforced at the database layer. The signing user must hold the required role at the time of signature, not when the record was created.

Per-org isolation

Every record carries an organization ID. Every query, every API route, every background job filters by it. 119 database models, no cross-tenant leakage path.

Postgres with pgvector

Managed Postgres for both structured records and semantic embeddings. Single database, single connection layer, no separate vector store to keep in sync.

Rate limiting at the edge

Upstash Redis enforces per-user and per-org request budgets on write-heavy routes. Prevents brute force on the signature endpoint.

Observability scoped to your tenant

LLM calls, tool invocations, and latency traced via Langfuse. Your prompts and outputs stay visible to you, not other organizations.

Clerk-backed authentication

Identity, sessions, and password policy managed by Clerk. No custom password hashing, no homegrown session tokens.

Multi-factor authentication

TOTP and email verification supported today. Can be required at the organization level on pilot request.

SAML SSO

Available on enterprise pilots. Talk to us during scoping; we configure per-org.

Session policy

Idle timeout and forced reauthentication on sensitive actions, including every Part 11 signature.

Your evidence does not train a third-party model.

Model calls use provider APIs with training opt-out in effect. No fine-tuning on tenant data.

Providers are named, not hidden.

Anthropic and OpenAI for reasoning. Voyage for embeddings. You see what leaves your tenant.

Prompts, outputs, and tool calls are traced.

Every agent step is logged with the exact inputs, outputs, and token counts. Review any claim the system made.

No third-party data retention beyond the API call.

Providers operate under zero-retention agreements where available; otherwise standard API retention applies. Posture stated explicitly during onboarding.