Version: v0.2 Status: Review Owners: Foundation / FMIE Last Updated: 2026-03-25 Change Type: additive Migration Required: yes Depends On: Media Passport, External Cognition Gateway Supersedes: none Superseded By: none
Prevent raw or over-disclosing internal data from leaving Foundation when calling frontier models.
Foundation-V2 already has governed publishing, but its outbound frontier-model path can still send raw prompts and rich internal evidence directly to Anthropic. The Semantic Airgap is the mandatory boundary that converts private internal state into policy-compliant external task packets.
DisclosurePolicyFields:
- policy_id
- tenant_id
- classification
- allowed_modalities
- allowed_providers
- allowed_tasks
- default_mode
- retention_policy
- approval_profile
- effective_from
- effective_to
DisclosureTransformFields:
- transform_id
- source_passport_id
- source_asset_ids
- modality
- transform_kind
- redaction_rules
- surrogate_mode
- derived_fields
- quality_floor
- content_commitment
SurrogateRegistryMaps protected entities to stable session-scoped or task-scoped surrogate IDs.
TaskPacketFields:
- task_packet_id
- policy_id
- transform_id
- provider
- model
- task_type
- payload_class
- payload_commitment
- requested_output_schema
- expiration_at
ExternalInferenceReceiptFields:
- receipt_id
- task_packet_id
- provider
- model
- status
- request_commitment
- response_commitment
- started_at
- completed_at
- retention_class
- disclosure_mode
ReasoningImportFields:
- reasoning_import_id
- receipt_id
- trust_status
- trust_score
- contradiction_flags
- local_crosscheck_status
- surrogate_outputs
DisclosureExceptionFields:
- exception_id
- policy_id
- approver_id
- reason
- scope
- expires_at
- audit_ref
denypublic_research_onlyabstract_reasoning_onlyapproved_raw_exceptionDisclosurePolicyDisclosureTransformExternalInferenceReceiptReasoningImportabstract_reasoning_onlyapproved_raw_exceptionRequired counters and rates: - packet count by disclosure mode - denied export count - raw exception count - provider/model distribution - average trust score - contradiction rate on imported reasoning - transform failure count
This spec closes the gap between existing publication governance and missing outbound disclosure governance.
abstract_reasoning_only