Identity
People, services, roles, and source authority.
ATTRIBUTABLE EVENTCorporate GPT applies one authority chain from identity and source access through evidence, model policy, tool scope, human approval, audit, evaluation, versioned release, monitoring, and rollback.
THE AUTHORITY CHAIN
People, services, roles, and source authority.
ATTRIBUTABLE EVENTPermitted sources, provenance, and correct abstention.
ATTRIBUTABLE EVENTModel routing, tool scope, data paths, and action limits.
ATTRIBUTABLE EVENTNamed approval for consequential action and promotion.
ATTRIBUTABLE EVENTAudit, evaluation, signed version, monitoring, and rollback.
ATTRIBUTABLE EVENTControl implementation is configuration-specific. Certification status, penetration testing, recovery objectives, formal mappings, support access, and exceptions belong in the security pack.
CAPABILITY ATLAS
Enterprise identities, service identities, permissions, and source ownership resolve before retrieval or action.
Material claims, model routing, tool scope, and permitted data paths remain explicit for the task.
Consequential actions and capability promotion pause for the named decision owner.
Events are attributable and candidates are compared against agreed baseline and held-out work.
Signed versions, operational monitoring, revocation, and rollback close the chain.
Least privilege for people, services, data, and tools.
Idempotency, retries, failure handling, and recovery paths are designed per workflow.
Human gates, kill switches, revocation, and replay remain available to accountable operators.
Tool calls, evidence, approvals, outcomes, and failures can be reconstructed.
In the air-gapped configuration, local models, customer infrastructure, and signed offline update bundles allow operation without a required external path. Support, updates, integrations, and telemetry are documented for the exact deployment.
SOC 2, HIPAA, GDPR, and EU AI Act control families can be mapped in the security pack. This describes readiness and configuration-specific controls—not certification, legal advice, or universal applicability.
No ‘zero hallucinations.’ Grounded-answer mode attributes evidence and can refuse, with quality measured on the agreed evaluation set. No universal zero-egress claim across connected configurations. No compliance seals we have not earned.