Data handling and encryption

The residency rule is simple because the deployment model does the work: everything in the inventory above lives inside your deployment. Run traces stay in your deployment in every mode. In a VPC, on-premises, or air-gapped architecture, that deployment sits inside your own perimeter — the control plane, execution environments, context filesystem, and run history included, not just inference.

What varies by mode is where the perimeter sits and which flows cross it, and that is documented during implementation. The deployment models page carries the mode-by-mode comparison, including trace residency for each architecture.

Context does not use one customer's traces, corrections, or institutional context to train models for other customers. That commitment is the platform's own; the second half of the question concerns the model providers a workflow calls.

The platform is model-agnostic — Claude, GPT, Gemini, Kimi, or open weights — and which endpoints are reachable depends on the architecture you select. Each provider's retention and training settings are documented and verified with you during implementation, endpoint by endpoint.

Deployments that cannot tolerate provider egress have structural answers: VPC deployments can use private endpoints so inference traffic does not leave your network, and air-gapped deployments serve open-weight models entirely inside the boundary.

Customer personal data is encrypted in transit using TLS and at rest using industry-standard encryption. Encryption is one item in the written information security program committed to in the DPA's Annex 2. The program also includes:

• Logical access controls: role-based access, least-privilege provisioning, multi-factor authentication for administrative access, and prompt revocation when roles change.
• Network and infrastructure controls: segmentation, firewalling, and hardened configurations with leading cloud providers.
• Vulnerability management: periodic scanning, patching, and third-party penetration testing.
• Logging and monitoring designed to detect unauthorized access.
• Personnel measures: background checks where permitted by law, confidentiality obligations, and security awareness training.
• Business continuity and disaster recovery, with backups and periodic testing.
• A documented incident response plan covering identification, containment, investigation, remediation, and notification.

The full annex is public in the Data Processing Addendum, and the combined security policies behind it are available in the document library.

Operational retention — how long traces and recordings stay queryable, what gets exported to your systems — is configured for your deployment, because audit and records requirements differ by customer and regulator.

Contractual deletion is defined in the DPA: when the agreement ends, Context deletes or returns customer personal data at your election made within thirty days, except where law requires retention or the data sits in routine backups — backup copies remain protected under the DPA and are deleted on the standard deletion schedule.

The DPA also covers the operational edges of data handling: assistance with data subject requests, and notification without undue delay if a personal data breach affects customer personal data, with the information needed to meet your own notification obligations.