Skip to content

Federated Clean Room ​

Availability

Federated Clean Rooms require an Enterprise subscription. The feature is gated by the clean_room feature flag.

The Federated Clean Room enables multiple organizations to run joint computations on combined data without exposing raw inputs to each other. Each party attests to the exact computation environment, results are held in encrypted escrow until all parties approve release, and independent provenance records are generated per organization.

Use Cases ​

  • M&A due diligence -- model revenue overlap and customer concentration risk without exposing either party's raw financial data
  • Joint research -- combine clinical datasets from multiple institutions under HIPAA-compliant governance
  • Supply chain analysis -- benchmark cross-org logistics without sharing proprietary cost data
  • Regulatory collaboration -- share aggregated findings while keeping individual submissions confidential

Lifecycle ​

Clean rooms follow a strict state machine:

1. Create ​

The initiating organization creates a clean room with a name, description, and optional computation template.

2. Invite and Accept ​

The initiator invites partner organizations by org ID. Each invited org must explicitly accept the invitation to become a participant.

3. Set Computation Template ​

The computation template defines the exact code that will run. When the template is set, an attestation bundle is generated containing SHA-256 hashes of:

  • Image hash -- the PhD sandbox Docker image
  • Code hash -- the computation template
  • Config hash -- sandbox configuration and governance policies
  • Combined hash -- a single hash of all three for convenience

The bundle moves the clean room to attesting status.

4. Mutual Attestation ​

Each participating organization reviews the template, image, and configuration, then submits their attestation with the bundle hash. This confirms that General Counsel or the designated reviewer has approved the exact code and environment.

When all participants have attested, the clean room moves to ready.

5. Execution ​

Computation runs in a gVisor-isolated PhD sandbox. Each organization's data enters through its own governance channel (Dymium connection with per-org column masking). The clean room enforces the intersection of all governance policies -- the most restrictive rules from any party apply.

Results are encrypted with AES-256-GCM and placed in escrow.

6. Release or Reject ​

Each organization's designated approver reviews a summary of the results (without seeing the full payload). All parties must approve for results to be released. If any party rejects, encrypted results are destroyed and provenance records note the rejection.

7. Provenance ​

Each organization receives independent provenance records proving:

  • What computation code ran (SHA-256 verified)
  • What sandbox environment executed it (image hash verified)
  • What their side contributed (input lineage -- the other party's inputs are not visible)
  • What the output was (shared, identical hash)
  • When each party attested and approved

Security Model ​

ControlImplementation
Data isolationEach org's data enters through its own Dymium connection with org-specific masking
Policy intersectionMost restrictive governance rules from all parties apply
AttestationCryptographic hash verification of image, code, and config
EscrowAES-256-GCM encrypted results, dual-approval release gates
ProvenancePer-org audit records with row-level security in PostgreSQL
SandboxgVisor-isolated PhD container with no network, read-only filesystem
TTLEscrowed results auto-expire after configurable period (default 7 days)

Configuration ​

VariableDefaultDescription
CLEAN_ROOM_ENABLEDfalseMaster toggle
CLEAN_ROOM_MAX_PARTICIPANTS10Max organizations per clean room
CLEAN_ROOM_RESULT_TTL_HOURS168Escrow expiry (7 days)
CLEAN_ROOM_REQUIRE_ATTESTATIONtrueRequire all orgs to attest before execution

Clean Room also requires PHD_MODE_ENABLED=true and SANDBOX_ENABLED=true.

SOC 2 Type II Certified | BAA Available | HIPAA-Ready