Responsible AI

The Nebbos platform commits to a defined set of responsible AI principles and refuses a defined set of operations at the substrate. Enforcement is a property of the platform architecture rather than a policy document; every commitment and every prohibition maps to a substrate mechanism that produces auditable evidence of adherence.

Commitments

01 · Attested human approval at the tier appropriate to the action

Guest-tier reads clear on device-native biometric authentication. Host-tier writes require the Cradle physically present and produce a Cradle-signed delegation. Architect-tier administrative operations require an additional enclave-signed approval token. The substrate verifies each signature before executing the operation.

02 · Citation provenance on every recommendation

Every Pearl recommendation is accompanied by references to the memory registers, prior decisions, and source documents consulted in its production. Recommendations without accompanying provenance are not returned to the caller.

03 · Operator data remains within operator scope

Nebbos does not train foundation models. Nebbos does not aggregate content from one operator’s scope into another operator’s Pearl. Nebbos does not use operator content to improve shared models. Enforcement is at the substrate; see Row-level isolation.

04 · Portability verified on every production deploy

The export path executes as part of every production deploy against scoped test tenants. A failure in the export path fails the deploy. See Portability & export.

Prohibited operations

• Recommendations issued without citation provenance.
• Writes to another operator’s scope, regardless of the calling identity’s tier or authorisation elsewhere.
• Operations that bypass the tier gate corresponding to the operation’s consequence class.
• Reads of a Pearl’s memory register from outside the department that owns it.
• Recommendations concerning protected-category decisions (hiring, dismissal, promotion, credit approval, medical treatment) without a named human decision-maker in the approval chain.
• Training operations against operator data. Nebbos orchestrates access to third-party model providers rather than producing models.

Enforcement path

Each commitment above is mapped to a substrate policy. Each prohibited operation is mapped to a substrate refusal path. Every refusal produces an entry in the audit chain with the specific policy that produced the refusal. Adherence to responsible AI commitments is verifiable from the chain.

This page is the public summary. The complete policy list, enforcement rule set, and drift instrumentation are part of the EU AI Act · Annex IV pack.