Compliance and Attestation
The platform's compliance posture, the controls a reviewer looks for and where each is enforced, and the per-app CAIQ attestation filled from cited evidence on every deploy.
The platform holds a SOC 2 Type II report, runs a HIPAA program with Business Associate Agreements in place with its infrastructure and model providers, monitors its controls continuously, and gives every app built on it an attestation of its own: a CAIQ filled from the platform's cited facts and the app's own analysis. This chapter covers the posture, the controls a reviewer will look for, and the attestation product.
#Posture
| Program | State |
|---|---|
| SOC 2 Type II | Report available under NDA |
| HIPAA | Program in place: BAAs signed with AWS, Cloudflare, Google Cloud (Vertex, Document AI, Cloud Storage), Anthropic, Microsoft (Azure) and OpenAI; HIPAA-scoped controls monitored alongside SOC 2. A workspace becomes a HIPAA workspace by accepting a BAA with us |
| Continuous monitoring | Vanta, connected to AWS, GitHub, identity and endpoint systems, with every infrastructure resource tagged for ownership and data classification |
| Trust page | Subprocessor list rendered from the counterparty registry |
The deployment's agreements record, per counterparty, the agreements in force, and whether a surface is BAA-covered follows from them. That fact surfaces in the model picker, in the subprocessor list, and in the attestation, from one source. The HIPAA-configured Anthropic organization is an example: models served through it carry BAA coverage as a declared fact, and features the agreement excludes are stripped from requests rather than attempted.
A HIPAA workspace is one with a business associate agreement with us in force, accepted in the product by an owner or admin, or signed outside it and recorded by staff. The same record holds the workspace to the chain of agreements:
- The credential clearinghouse refuses the workspace any service on the deployment's credentials that no BAA covers, naming the company, and never reroutes to another (see Counterparties and Egress).
- Payload Logging is off for every app, and the idle timeout is at most thirty minutes.
- MindStudio is off.
- Its audit log shows staff access.
The workspace's own keys and connected accounts are its own relationship with the vendor, and what its apps' own code reaches is its own to cover. The one service the deployment decided needs no BAA is recorded beside the agreements with its reason: Twilio Verify's sign-in codes, which are templated codes that carry no health information.
Any workspace can hold its vendors to two more requirements of the same kind, drawn from the facts the registry records. No Training refuses a third-party service that trains on the content it is sent. Signed DPA refuses one that is sent personal data without a signed DPA and a lawful basis for EU transfers (see Counterparties and Egress). They answer the questions a privacy review asks of every subprocessor, and they hold for a vendor added after the review as much as for those it read.
#Controls a Reviewer Will Look For
Each of these is described in its own chapter; this is the index.
| Control | Where it lives | Chapter |
|---|---|---|
| Encryption at rest | Customer-managed KMS keys on the database, the secrets envelope, the registry and the backup vaults; server-side encryption on object storage; encryption at rest on the Valkey cache | Secrets, Platform Database |
| Encryption in transit | TLS at the edge, TLS 1.3 with a FIPS policy at the load balancer, TLS enforced at the database cluster, TLS required on every Valkey connection with the platform as the only admitted user (short-lived IAM tokens, unauthenticated connections refused), HTTPS-only bucket policies | The Trust Boundary |
| Tenant isolation | A microVM per app, a database file per app, an origin per app, a hook token per execution | Execution Islands, App Databases |
| Network isolation | Two VPCs, two clusters, a peering that admits three ports one way, sandbox egress to the internet only | The Trust Boundary |
| Identity and access | Cognito with SRP, MFA and federation; SSO; least-privilege workload identity | Identity and Auth |
| Secrets management | App secrets and every vendor key (the platform's own and each workspace's) sealed under a KMS key into the platform database, the platform's keys managed by staff with a recorded history, one code path to obtain any of them | Secrets |
| Audit logging | Security-relevant workspace and app changes (members, roles, keys, billing, settings, connections, secrets, deploys) and every human method invocation, with the credential used; WORM-stored for a year and kept after, in us-west-2; OCSF export | Audit Log |
| Vendor management | Every third party declared with data classes, retention, training policy, exposure and agreements | Counterparties and Egress |
| Malware and PII detection | Stored app files scanned with ClamAV and Presidio, within size and type limits | Content Inspection |
| Backup and recovery | Continuous backups, cross-region copies, weekly automated restore tests | Platform Database |
| Data residency | Four regions, data at rest pinned to the workspace's choice | Data Residency |
| Threat detection | GuardDuty with EKS runtime monitoring, CloudTrail with log-file validation, VPC flow logs, database audit logging, storage access logs | Observability |
| Change management | Tagged releases, migrations before rollout, CI validators, no-cancel production deploys | Build and Deploy |
| Web application security | Security headers, cookie flags, origin separation, embedding allowlists | Served-App Security |
#Attestation for Apps
A customer building on Remy is asked the same questions about their app that the platform is asked about itself. The platform answers most of them for them.
The assessment report. After every deploy, a static analysis of the release produces a structured report: the app's auth methods, its roles, its data handling, which models it calls and what it sends them, its external connections, and a data-flow diagram. A commit classifier decides whether a release changed enough to warrant regeneration, so an unchanged release reuses its prior report.
The CAIQ attestation. The Cloud Security Alliance's Consensus Assessments Initiative Questionnaire, the instrument enterprise buyers already use, is filled for the app from three layers:
- Platform facts, authored once and cited to a control, the component that enforces it, or an agreement. These are the same for every app: how sessions are handled, where data is encrypted, what the serving layer enforces.
- The app's own analysis, projected from the assessment report per release: which auth methods it enables, what data it stores, which models it uses.
- The builder's own controls, their governance, their people, their policies, left for the builder to complete, with guidance.
The rule that makes this trustworthy is that nothing is asserted without a citation, and nothing is guessed. A model maps and phrases known facts into the instrument's shape; it never decides whether a control is met. The result is served through the app's management API as the full control set with provenance on every answer, and the workspace's dashboard renders it as the questionnaire a reviewer expects, with inapplicable controls justified rather than omitted.