The persona package

Every persona is a signed archive with four mandatory components:

Manifest — identity, version, compatibility, creator.
Consent manifest — what uses the voice owner approved.
Voice model — the actual acoustic data.
Personality config — the style prompts and refusal patterns.

Manifest format

{
  "schema_version": "1.0",
  "persona_id": "per_01HX3A...",
  "name": "Warm British Narrator",
  "version": "2.1.0",
  "creator": { "id": "creator_abc", "name": "Jamie Lee" },
  "compatibility": ["gera.agent.v2+", "robot.gera.v1+"],
  "locales": ["en-GB", "en-US"],
  "voice_tier": "studio",
  "content_ratings": ["PG"],
  "subscription_model": "per-install-monthly"
}

Consent manifest

The consent manifest is the legal-technical backbone of the platform. It records:

{
  "consenting_party": {
    "identity": "verified_creator:creator_abc",
    "relationship": "self",
    "consent_date": "2026-03-10",
    "consent_method": "signed-contract"
  },
  "allowed_use_cases": [
    "personal_household",
    "narrator",
    "meditation",
    "audiobook"
  ],
  "disallowed_use_cases": [
    "political_advocacy",
    "explicit_content",
    "financial_instruction",
    "impersonation_to_third_party_without_disclosure"
  ],
  "invocation_allow_list": "public",
  "revocation_contacts": ["creator:creator_abc"],
  "estate_beneficiaries": [
    { "party": "heir_xyz", "share_percent": 100 }
  ]
}

The consent manifest is the gate at runtime. Any invocation that falls outside allowed_use_cases is refused with the reason logged.

Voice model tiers

Studio tier — full acoustic model recorded in studio conditions; larger package, highest quality.
Home-recording tier — compact model trained on user-uploaded recordings; smaller package, good quality with warnings about volume / naturalness.
Heritage tier — models trained from archival recordings (home videos, interviews); lowest fidelity, highest emotional value.

Personality config

The personality config shapes what the persona says as well as how. It contains:

Style prompts (“speak slowly, use traditional storytelling cadences”).
Preferred vocabulary / avoided vocabulary.
Refusal patterns (“when asked about politics, gracefully decline and offer to read a poem instead”).
Disclosure patterns (“when invoked outside the declared-family allow-list, state that this is a voice skin of X”).

Install flow

User browses catalogue; selects persona; clicks install.
Platform verifies the host device is compatible (checks compatibility vs device capability descriptor).
Platform verifies the invoker identity is in theinvocation_allow_list.
Platform downloads the package and verifies signature against creator’s public key.
Package is decrypted, the voice model loads, the personality config applies.
Subscription billing activates; residual payments begin accruing to creator.

Runtime behaviour

Every response from a persona-skinned agent passes through three gates before being spoken:

Consent-scope gate: is this use-case inallowed_use_cases? If not, refuse with reason.
Personality-config gate: does the proposed response fit the style / vocabulary / refusal patterns?
Disclosure gate: does the context require a “this is a voice skin” prelude? If so, prepend.

Revocation and uninstall

Revocation: creators can revoke a persona with 90 days notice. During the notice period, installed users see warnings. After the period, the package stops activating on new sessions; already-active sessions complete normally.

Uninstall: users can uninstall any time. Residuals stop at the end of the current billing cycle. A 30-day reinstall grace period applies.

Integration with GeraNexus and GeraWitness

Financial transactions initiated with a persona voice route through GeraNexus for consent tokens. High-value or sensitive persona uses are reviewable via GeraWitness. The persona is the voice; the transaction is the protocol.

Persona API Spec — Voice Package Format, Install Flow, Runtime