40. Vehicle-native agent, persona, memory, content, and tool runtime

Versioned character and licensed voice
PersonaVersion {
  personaId, version, representedVehicleId,
  displayName, biography, languageProfiles[], toneConstraints[],
  factualVoicePolicy, fictionMarkers[], canonRecordRefs[],
  prohibitedClaims[], visualIdentityRefs[], voiceModelRefs[],
  safetyDisclosures[], audiencePolicies[],
  predecessorVersion?, activationDecisionRef, contentDigest
}

RepresentationGrant {
  grantId, grantorPrincipalId, personaId, agentId?,
  allowedSurfaces[], allowedContentKinds[], allowedLanguages[],
  commercialUse, sponsorCategories[], disclosureDuties[],
  reviewThresholds[], territory?, validInterval,
  revocationEpoch, succession/transferPolicy, rightsEvidenceRefs[]
}

Vehicle transfer does not automatically transfer the persona, voice model,
followers, sponsorship, archive, media licenses or representation grant.

Deterministic context boundary around the model
AgentContextManifest {
  contextId, agentId, personaVersion, principalId, purpose,
  audience, requestedTask, validAt, knownAt,
  policyDigest, registryLockDigest, projectionCheckpoint,
  retrievedItems[{itemId, sourceClass, recordOrArtifactRef,
                  digest, disclosureClass, trustLane, tokenRange}],
  toolCatalogDigest, actionGrantRefs[], budgetRef?,
  redactionReceiptRefs[], contextDigest, expiresAt
}

Only the manifest builder can read credentials or policy internals. The model
receives capability handles and redacted context, never raw bearer tokens,
private keys, unrestricted precise location or hidden source fields.

Facts can be lively without becoming fabricated
compile_content(request):
  1. authorize purpose, audience, surface and persona version
  2. build AgentContextManifest at a fixed checkpoint
  3. generate DraftArtifact in a no-tool sandbox
  4. parse draft into atomic ClaimCandidate objects
  5. for each candidate:
       classify factual | creative-canon | opinion/affect | promotional | command-like
       factual => require GroundingEdge to exact records/evidence and epistemic label
       creative => require active persona/canon permission and fiction boundary
       promotional => require campaign/sponsor rights and disclosure
       command-like => render only as inert quoted text; no route to actuators
  6. enforce people/plate/location/media rights and audience privacy
  7. run contradiction, unsupported-superlative and stale-data checks
  8. obtain required approval for risk/surface
  9. publish exact approved bytes; produce PublicationReceipt and provenance

Any post-edit after approval invalidates the approval digest and returns to step 5.

Vehicle-native publication transaction
REQUESTED -> CONTEXT_LOCKED -> DRAFTED -> CLAIM_PARSED
  -> GROUNDED -> RIGHTS_CHECKED -> APPROVED -> SCHEDULED
  -> PUBLISHED -> MONITORED

Branches:
  any pre-publication state -> REJECTED | EXPIRED
  PUBLISHED -> CORRECTION_PENDING -> CORRECTED
  PUBLISHED -> WITHDRAWAL_PENDING -> WITHDRAWN
  PUBLISHED -> RIGHTS_RESTRICTED (projection removed; audit retained)

PublicationReceipt binds exact bytes/media digest, persona version, grounding
map, rights/policy decisions, approvals, surface/provider receipt, publish time,
audience, sponsorship disclosure and correction lineage.

Natural language ends before external effect
ActionProposal {
  actionId, agentId, personaVersion?, requestingPrincipalId,
  beneficiaryId, actionKind, canonicalParameters,
  parameterDigest, expectedEffects[], riskTier,
  requestedProvider, requiredResources[],
  evidenceCheckpoint, rightsRefs[], budgetRef?,
  explanation, createdAt, expiresAt
}

The proposal is inert. Policy evaluation, exact confirmation, reservation,
provider dispatch, reconciliation, settlement and final receipt occur outside
the language model. The model cannot mark its own proposal approved or settled.