Dipp AI publishes its enterprise data boundary commitment

The commitment sets out how the Orcher data control gateway classifies payloads, enforces the enterprise's own boundary rules before dispatch, records every crossing and every refusal, and makes the boundary observable rather than asserted.

Every serious model provider now publishes some version of the same sentence: we will not train on your data. It is a reasonable commitment and an inadequate control. A commitment is enforced by contract, verified by attestation, and discovered in breach. A control is enforced by architecture, verified continuously, and observable by the party carrying the risk. Enterprises are being asked to accept the first where they need the second.

Dipp AI is publishing the commitment it holds itself to, in the form of the mechanism rather than the assurance. The Data Control Gateway is the component of Orcher responsible for it, and it operates before dispatch, not after logging.

The commitment, stated as mechanism

1. Every payload is classified before dispatch, against the enterprise's own data classes rather than a generic sensitivity scale.
2. Boundary rules are the enterprise's, expressed as policy, versioned, and enforced by the gateway rather than by the calling application.
3. A payload that would cross an unauthorised boundary is refused, and the refusal is a recorded event with a reason, not a silent drop.
4. Every crossing that is permitted is recorded: what class of data, to which provider, under which policy version, authorised by which role.
5. Redaction, tokenisation and local execution are routing outcomes, selected by policy, not manual pre-processing steps left to the developer.
6. The record is written to the Immutable Audit Ledger so the boundary can be evidenced, not merely described.

Why classification has to precede routing

Where classification happens determines whether the boundary is real. If a payload is classified after a provider has already received it, the enterprise has documented an exposure rather than prevented one. Placing classification ahead of the routing decision means data class becomes an input to model selection: a task carrying regulated records is eligible only for models and regions the policy permits, and if no eligible model exists, the task does not silently degrade to the nearest available option — it stops.

What this is not

This is not a claim about where data physically sits. Residency is a narrow property and, on its own, a weak control: data can sit in the correct region and still be absorbed into a training corpus, retained beyond its purpose, or reached by a party the enterprise never authorised. The commitment here is about the boundary itself — what may cross it, under whose authority, and with what evidence — and about ensuring that enterprise data does not become training material for anyone but the enterprise.

“Ask for the mechanism, not the sentence. If the answer to 'how would we know' is an attestation letter, it is not a control.”