# Semblance Likeness License — Specification v0.1.2 **Schema file:** `likeness-license.schema.json` (JSON Schema draft 2020-12) **`$id`:** `https://semblance.studio/schemas/likeness-license/v0.1.2.json` **`schema_version`:** `0.1.2` — the 0.1.1 revision closed the gaps found by pressure-testing v0.1.0 against three real deal shapes (AI ad, game NPC voice, VTuber sponsorship); the 0.1.2 revision applies every finding of two adversarial reviews (legal completeness and implementability). Summaries are in **"What 0.1.2 changed"** and **"What 0.1.1 changed"** below; deliberately deferred items are in **"Known limitations / v0.2 candidates."** > **Production domain.** `semblance.studio` is the registry's production domain; the `$id` above is canonical as of 2026-07-06. > secured, the `$id` (and the `verify_endpoint` conventions) change to the real domain and this > spec is updated. Instances written against the placeholder `$id` remain structurally valid; > only the identifier URI changes. This document is the human-readable, field-by-field specification of the canonical Likeness License schema — the merge of three independent drafts (legal-completeness, machine-executability, and deal-realism lenses). Resolution rules applied throughout: **legal** wins on consent/revocation/union/compliance structures; **agent** wins on types, enums, identifiers, and machine-evaluability; **commercial** wins on pricing and deal structures. Design principles the merge preserves: 1. **Illegal or talent-hostile terms are inexpressible, not discouraged.** Minors, irrevocable consent, waived AI disclosure, sublicensing, auto-renewal for humans, wildcard model IDs, below-scale union deals, machine-granted exclusivity, and edits to the prohibited-use baseline all fail schema validation outright. 2. **Fail-closed machine evaluation.** Anything not machine-decidable (free text, union coverage, exclusivity, personas, persistent uses, revenue share) routes to manual approval. Free text exists for humans and courts; it is never a matching input. 3. **No floats, no FX, no fuzzy matching.** Money is integer minor units + ISO 4217; percentages are integer basis points; territories are ISO 3166-1 sets; platforms/models are registry IDs; competitor blocks are NAICS prefixes. 4. **The consent record is bound to the exact terms** via SHA-256 over the JCS-canonicalized `/grant` object. 5. **What JSON Schema cannot express is still normative**: `x-state-machine` (lifecycle), `x-invariants` (numbered cross-field rules INV-001…INV-036 the platform validator MUST enforce), and `x-auto-approval-algorithm` (the ordered, deny-fast, 16-step evaluation). Validation status (0.1.2): compiles clean under ajv (draft 2020-12, formats enabled); the four example instances in `examples/` (auto-approved AI ad, manual game NPC voice, MCP-executed VTuber standing offer, manual estate archival documentary) validate with real RFC 8785 (JCS) SHA-256 `grant_hash` and `approval_hash` values; 30 negative/positive mutations of the 0.1.2 conditionals verified to behave as designed. --- ## Required at execution time vs. draftable A license record exists from first configuration (`draft`) through terminal states. Not everything can exist on day one: | Stage | Blocks required | |---|---| | Always (all statuses, incl. `draft`) | `schema_version`, `license_id`, `license_version`, `status`, `lifecycle`, `audit`, `parties`, `grant`, `approval`, `pricing`, `usage_reporting`, `union_rider`, `asset_handling`, `revocation`, `provenance`, `warranties` | | From `quoted` onward | `integration` (with `quote_id`) — every priced license traces to its immutable quote | | From `executed` onward (**execution-time blocks**) | `consent` (the signed consent instrument + grant hash), `execution` (signatures, certificate, payment, contract ref), and `integration.execution_idempotency_key` (INV-010) | | From `executed` onward, when `approval.mode` ≠ `manual` | `content_declaration` (the licensee's request-time content codes, INV-026), `consent.advance_authorization` (the pre-signed envelope binding + `licensor_terms_ref`, INV-023/INV-029), and **`campaign` with `product_or_service`** (0.1.2 — the record of what the likeness actually promoted, disclosed at execution) | | Optional at any stage | `amends_license_id`, `deliverables`, `governing_law`, `campaign` (for manual mode; top-level — outside the hashed grant), `grant.credit`, `x_notes` (annotations) | Rationale: a draft or quote describes *proposed* terms; consent is captured against the final grant at approval, and the grant hash (INV-007) is verified at the `executed` transition. The `declined` and `cancelled` terminal states never require consent/execution — they exist so `quoted` and `pending_approval` have machine-unambiguous exits. --- ## Top-level fields ### `schema_version` *(const `"0.1.2"`)* Which rule set applied at execution. Evaluators MUST refuse versions they do not implement (fail closed). ### `license_id`, `license_version`, `amends_license_id` `lic_` + Crockford-base32 ULID; immutable. `license_version` increments on any amendment to `/grant`, `/pricing`, or `/revocation` — and every increment requires a **fresh consent record** (INV-007) — with one express exemption (0.1.2): **`pricing.escrow.state`** is operational, platform-managed payment state, not a negotiated term; escrow transitions never bump `license_version` and never require re-consent. Renewals are **new documents** pointing back via `amends_license_id` (INV-021): this is how the schema makes the **NO FAKES Act draft 10-year cap** re-check itself per renewal — chained renewals can never quietly become perpetual paper. ### `status`, `lifecycle.status_history` Ten states (`draft`, `quoted`, `pending_approval`, `executed`, `active`, `expired`, `revoked`, `terminated`, `declined`, `cancelled`); the only legal transitions are in `x-state-machine`. 0.1.2 completes the machine: **`executed → expired`** (clock event — a license whose `end_date` passes before activation can no longer sit in `executed` forever with funds captured), **`pending_approval → declined` on `sla_timeout_declined`** (the system transition matching `booking_approval_sla.on_timeout: "auto_decline"`; actor `system` — silence still never *approves*), and **`expired → terminated`** (`post_term_breach_terminated` — the for-cause lever against abuse of a runoff window or an embedded-products survival election, so `expired` is now quasi-terminal with exactly one exit). Activation is defined for service deliverables too: `persona_endpoint_access` / `live_stream_session` licenses activate on the platform's endpoint-provisioning event, not on a watermarked file delivery that may never exist. `status_history` is the append-only audit trail (actor-typed, webhook-event-keyed) that must replay to `status` (INV-011). Why it exists: auditable authorization histories are precisely the evidence a NO FAKES-style notice-and-takedown regime and courts reward, and `actor.type` distinguishes human approval from `system`/`licensee_agent` action — critical in a later dispute about whether a human ever approved. ### `integration` **Two idempotency keys** (0.1.2): `booking_idempotency_key` for `create_booking` and `execution_idempotency_key` for `execute_license` — a retried execution can never be confused with the booking that preceded it. INV-010 scopes both to *(licensee_id, operation)* — a key collision across tenants can never return another account's record — with a normative ≥ 24-month retention window. Plus quote/booking lineage with **`quote_expires_at`** (the quote TTL made machine-decidable in-document until the v0.2 quote schema lands), **`reference_registry_snapshot`** (0.1.2, INV-036: the date-versioned snapshot of the ISO 3166-1 country set, EU/EEA membership, ISO 4217 exponents, and NAICS tables the document is evaluated against — two conformant implementations compute identical territory expansions and money math), and webhook subscriptions against a closed event-name set. ### `audit` Document-level created/updated timestamps and creating actor. Grant/pricing/revocation are immutable per version; only platform-managed fields mutate. --- ## `parties` Semblance is the verifying registry, broker of record, and escrow agent — **not** a party to the grant. ### `parties.licensor` - **`identity_type`** (`human` | `estate` | `character`) drives everything: verification block, term rules, and union posture. - **`adult_verified`** *(const `true`)* + **structured age attestation (0.1.2, INV-012)** — the minors exclusion is structural *and evidenced*: the constant is backed by a **`$defs.age_attestation`** (`date_of_birth_verified` const + hash-fixed `age_verification_ref` naming the KYC artifact that verified DOB ≥ 18 + `verified_at`) for **every verified natural person the document touches**: the human licensor (`licensor.age_attestation`, required iff `human`), the estate representative (`estate_authority.representative`), the character-owner signatory (`character_provenance.owner_signatory`), and any embodied performer (`character_provenance.embodied_performer`). A court reading the document traces 18+ to evidence, not to a schema-required word. - **`human_verification`** (required iff `human`): approved KYC vendor + reference + biometric liveness timestamp. References only — no biometric data ever lives in a license document (BIPA/GDPR hygiene). Nobody can list a face that isn't theirs. - **`estate_authority`** (required iff `estate`): decedent name, date of death, governing jurisdiction (post-mortem regimes differ: CA Civ. Code §3344.1 ≈ 70 years, NY CRL §50-f ≈ 40, TN potentially perpetual), authority type (executor / heir / foundation / licensed representative), hash-fixed authority documents, **`post_mortem_rights_expire_at`** — the recorded bound the term may never exceed (INV-002), **re-verified at every execution and renewal** (0.1.2 — renewal-contingent regimes like the NO FAKES draft's 10 + 5 can move it) — and the required **`representative`** (0.1.2): the named, age-attested natural person who signs for the estate. - **`character_provenance`** (required iff `character`): chain-of-title evidence (commission / work-for-hire / assignment / VTuber model-ownership proof), optional copyright & trademark registrations, **`no_third_party_elements_warranty`** *(const `true`)*, the required **`owner_signatory`** (0.1.2: named, age-attested signer for the owning entity), and the trap most drafts miss — **`embodied_human_performer`**. **0.1.2 promotes the embodied performer to a first-class sub-party**: when `true`, the **`embodied_performer`** block is required, with the performer's own `performer_id` (`per_` ULID), legal name (confidential — public anonymity stays governed by `credit.performer_anonymity_required`), `adult_verified` + age attestation, full KYC/liveness `verification`, own intended-use **`consent_ref`**, and own data-subject **`bipa_written_release_ref`** (§15(b) release executed by the performer, never by the owning entity). The performer's union coverage flows into `union_rider`, `payout_split` **must** carry an `embodied_performer` entry (INV-013), and the **INV-001 10-year cap applies to the whole grant** — a character wrapper can no longer launder a living performer's voice past any human-protective rule. - **`representation`** / **`payout_split`**: talent agents stay in the loop (approvals route to the agent of record) and licensor-side splits (agent commission, estate beneficiaries, union trust, **embodied performers** — 0.1.2) are first-class, in basis points summing to 10000 (INV-013, remainders per INV-032) — never off-platform side deals. ### `parties.licensee` - **`verification_tier`** (`tier_0_unverified` … `tier_4_enterprise_contracted`): named so **lexicographic order equals trust order** — min-tier checks are plain string comparisons. Higher-risk categories gate to tier 3+ (INV-006). - **`booked_via`** (`web` | `api` | `mcp_agent`): `mcp_agent` **requires** `agent_delegation` — the agent's identity, the **named principal who is legally bound** (an agent cannot be the counterparty), the delegation protocol (OAuth2 / ACP / UCP / AP2 / Visa TAP / Mastercard Agent Pay / x402), the mandate reference, and an optional licensee-side `human_confirmation_required_above` threshold (the agentic-payments human-in-the-loop pattern, from the buyer's side). **0.1.2 (INV-034): `booked_via` is platform-derived, never caller-asserted** — a credential registered as an autonomous agent is forced to `mcp_agent` with full delegation; presenting as `api` to skip HITL thresholds is rejected platform-side. - **`brand` / `is_agency_of_record` / `end_client_legal_name`**: agencies book, brands conflict. Exclusivity and competitor blocks are evaluated against the **end brand** (`brand.industry_naics`, NAICS 2022, prefix-matchable) — never against the media agency (INV-016). Agency flow is deliberately *not* sublicensing. --- ## `consent` — the AB 2602 engine *(execution-time block)* The consent instrument is structured so that the **"reasonably specific description of intended uses" required by Cal. Lab. Code §927 (AB 2602)** and its Illinois/New York digital-replica analogues is **forced by the schema**, never left to optional prose: - **`intended_use.project_or_campaign_name`** — a named project. Generic grants can't validate. - **`intended_use.narrative_description`** — **minimum 200 characters** of plain language: what the outputs depict, in what context, for what audience, what the licensor is shown saying or doing. Boilerplate ("any and all uses, in perpetuity, in all media") is unrepresentable. - **`intended_use.deliverables[]`** — at least one enumerated deliverable, each tying a concrete description (≥ 20 chars) to a `media_type` (which must be within `grant.media_types`, INV-019) and a quantity estimate. - **`intended_use.licensee_identity_disclosure`** *(0.1.1, replaces the const-true boolean)* — `disclosed_prior_to_consent` (forced for manual approvals by root conditional) or `disclosed_at_execution_per_advance_authorization` (advance modes: the licensor pre-consented to a *class of verified licensees* bounded by the envelope's tier floor and allow/blocklists, and is shown each executing licensee's verified identity at execution). Anonymous buyers cannot obtain a license under either value. - **`advance_authorization`** *(0.1.1 — the auto-approval consent story, defined; 0.1.2 — hardened)* — required from `executed` onward whenever `approval.mode` ≠ `manual` (and *forbidden* under `manual`): for machine execution, **the instrument executed is the licensor's pre-signed advance instrument** (`instrument_ref`), which reproduces the intended-use narrative **and the full `auto_approval_parameters` / `standing_offer` verbatim**; `approval_hash` (SHA-256 over the JCS-canonicalized `/approval` object) pins the instrument to the exact envelope; `captured_at` = `envelope_signed_at` = when the human signed — never stamped at machine execution (INV-023). **0.1.2 adds the required `licensor_terms_ref`** (hash-fixed reference to the licensor's published standing-terms document) **plus INV-029**: at every machine execution the platform verifies **byte-equality (after JCS) of every talent-fixed block** — `/grant/content_exclusions`, `/grant/platform_allowlist`, `/grant/ai_disclosure`, `/revocation`, `/asset_handling`, `/usage_reporting`, `payout_split` — against that document. `approval_hash` pins the envelope; `licensor_terms_ref` pins everything else the licensor fixed, so a buggy or compromised booking assembler cannot drop an excluded code, flip `at_will_permitted`, or release reference material the licensor never authorized. `execution.licensor_signature_ref` resolves to the advance instrument, so *who executed what, when* has one answer. Per-execution specifics (licensee identity, `/campaign` — **required with `product_or_service` for every non-manual executed license, 0.1.2** — and `/content_declaration`) are disclosed at execution via dashboard/webhook. For advance modes the consent narrative describes the authorized **class** with its hard bounds; the named "project" is the envelope/offer itself. - **`representation`** — records which §927 enforceability safe harbor applies (`legal_counsel` | `union_cba_coverage` | `unrepresented_specificity_relied_upon` | **0.1.1:** `entity_licensor_not_applicable` for character/estate entities licensing owned IP, where the human-performer safe-harbor analysis does not apply). Under §927 a non-specific provision is unenforceable *unless* the performer was represented; recording both prongs proves enforceability either way. **0.1.2 representation gate (schema conditional):** a `human` licensor under `auto_approve_within_parameters` / `standing_offer` **must** hold `legal_counsel` or `union_cba_coverage` — class-level advance consent by an *unrepresented* human is exactly the statutorily untested configuration (Known limitation #2), so until the fifty-state matrix lands it is unrepresentable rather than merely risky. Character licensors sign as `character_owner` (the authority doc may reference the onboarded chain-of-title summary — it binds the signatory to the verified entity, not a redundancy). - **`statutory_bases[]`** — closed list of the regimes the instrument was drafted to satisfy (CA §927/§3344/§3344.1, IL digital-replica & **BIPA**, NY CRL §50-f/§52-c, TN ELVIS Act, NO FAKES draft, GDPR Art. 9, EU AI Act Art. 50; **0.1.1 adds** Canadian regimes — `canada_quebec_civil_code_privacy`, `canada_provincial_privacy_acts`, `canada_common_law_personality` — `japan_publicity_right_caselaw`, and `contract_and_ip_only_no_publicity_right`, the honest basis for pure character/IP deals resting on copyright/trademark/contract) — what enforcement systems and takedown notices will cite. Cover the licensed *territory's* principal regimes, not only the forum's — **now validator-linted (0.1.2, INV-031)**: US-inclusive territory ⇒ at least one US basis (`tn_elvis_act` whenever a living person's voice is synthesized for US distribution; `il_bipa_740_ilcs_14` whenever biometric data is present); EU-inclusive ⇒ an EU basis; `worldwide` ⇒ both minimum sets. A floor, not conflict-of-laws analysis. - **`revocable`** *(const `true`)* — irrevocable consent cannot be expressed. - **`signatory`** — non-self signatories (estate rep, character owner, authorized agent) must attach hash-fixed authority documents. - **`language`** — BCP 47 tag of the instrument as executed (informed consent must be readable). - **`grant_hash`** — SHA-256 over the RFC 8785 (JCS) canonicalization of `/grant`, verified at the `executed` transition (INV-007). "The licensor consented to *these specific terms*" becomes machine-verifiable, and post-consent tampering is detectable. Fixed algorithm and canonicalization so independent implementations compute identical hashes. - **`instrument_ref`** — evidence-grade reference (id + SHA-256) to the signed writing a court reads; it must contain the intended-use narrative verbatim (INV-019). Consent is per-license and per-version. Renewals require a fresh record (INV-021). --- ## `grant` — the machine-matched core The exact object hashed into `consent.grant_hash`; immutable once executed. Everything not expressly granted is reserved to the licensor. Every dimension has a 1:1 mirror in `auto_approval_parameters`, so machine evaluation is dimension-wise subset/threshold checks. - **`use_categories`** — closed 7-value taxonomy (`advertising`, `film_tv`, `games`, `social_ugc`, `corporate_training`, `interactive_persona`, `music_audio`). - **`media_types`** — closed modality list; voice and visual likeness are separately enumerable because statutes (and NO FAKES) treat them as distinct rights; `music` (singing) is distinct from `audio_voice` (speech) — different union and moral-rights posture. **Normative pairing rule (0.1.1):** any output in which the likeness *speaks* with a synthesized voice requires `audio_voice` alongside the visual modality — a talking spokesperson video is `["video","audio_voice"]`; `video` alone grants silent/third-party-voiced visuals only. - **`channels`** — closed list with **no "all channels" value**: every channel is an affirmative enumerated grant (AB 2602 specificity by construction). Paid vs organic social are distinct (ad-platform provenance policies and union rates differ). **Normative boundary definitions (0.1.1, in `$defs.channel`)**: `social_paid` = paid placements bought through a social/video platform's *own* ad system, expressly including **YouTube pre-roll**; `digital_display_ads` = programmatic/display bought outside those systems; `streaming_ott` = content distribution on OTT services, never ad insertion; `in_game` = the playable product incl. patches/DLC — store pages, trailers (even in-engine captures), and marketing surfaces are *not* `in_game`, and end users' own gameplay streams are out of scope. An ad buy fits exactly one channel. - **`languages`** *(0.1.1, required)* — BCP 47 tags as a **grant dimension**: a third-language clip is now a machine-detectable breach, not a spec quibble. **Matching is RFC 4647 basic filtering at subtag boundaries (0.1.2)** — `en` covers `en-US`, but `es` never covers `est` and `zh` never covers `zho` (naive string-prefix matching is non-conformant); primary subtags are schema-required lowercase so case-sensitivity can never defeat matching. Mirrored by `allowed_languages`; deliverable-spec languages must be within it (INV-019). - **`distribution_platforms`** *(0.1.1, optional)* — hard allowlist of **distribution** platforms (`meta.instagram`, `google.youtube`, …), distinct from the *generation* `platform_allowlist`. "Meta, TikTok, YouTube only" is a machine term now; mirrored by `allowed_distribution_platform_ids`. - **`territory`** — `worldwide` | `include` | `exclude` + ISO 3166-1 alpha-2 sets. Evaluation is set expansion + subset check (INV-009). No "EMEA". Plus deal-realism flags: `geo_targeting_required` and `organic_spillover_permitted` (internet distribution is enforced through paid-media geo-targeting; organic spillover is addressed explicitly, not pretended away). - **`term`** — explicit `start_date`/`end_date` only; **`auto_renew` is const `false`**; no perpetual terms are representable in v0.1. **Living persons: ≤ 10 years (INV-001, NO FAKES draft cap)** — applying to human licensors **and to any grant with a living embodied performer** (0.1.2) — renewal only via a new document with fresh consent. **Cap measurement (0.1.2, normative and date-inclusive):** the encumbrance **anchor** is the *earliest* of `consent.captured_at`, `execution.executed_at`, and `start_date` (a person is encumbered from signing, so a postdated flight start can't stretch a 10-year cap into 15 years from signature), `end_date` is **inclusive** (exercisable through 23:59:59Z), and the check is `((end_date + 1 day) − anchor) ≤ P10Y` — a start 2026-07-01 grant must end by 2036-06-30; 2036-07-01 is 10 years + 1 day and fails. **Estates: term ≤ `post_mortem_rights_expire_at` (INV-002), re-verified at each execution/renewal.** Characters without an embodied performer: uncapped but still date-bounded (renewal is one API call; perpetual paper is a v0.2 question). `post_term_runoff_days` (sell-off for published outputs) and `persistent_service_wind_down_days` (active decommissioning of persona services) are standard deal mechanics, bounded. *JSON Schema cannot do date arithmetic — the caps are INV-001/002, mandatory validator rules.* Two 0.1.1 elections make term semantics explicit: - **`generation_may_precede_start`** — `start_date` bounds *distribution*; whether generation (and pre-flight delivery for review) may begin at execution is this explicit flag (the standard ad-deal shape), never an inference. - **`delivered_outputs_survival`** — the delivered-vs-new-generation split **at expiry**, mirroring `/revocation`'s split at revocation: `expires_with_term` (default) or `embedded_products_survive` — outputs embedded in products distributed during the term (shipped game builds) stay licensed for continued distribution after `end_date`; only *new* generation stops. `verify_license` reports surviving content as `licensed_post_term_survival` (INV-024), never as infringing. Survival is a major grant: always manual (INV-004(f)), and it never extends the INV-001/002 *generation* caps (it is continued distribution of lawfully made copies). - **`use_persistence`** — the single most consequential structural split: `one_time_outputs` (bounded deliverables) vs `persistent_synthetic_performance` (the likeness runs as a service). Drives economics (rev-share/periodic), wind-down obligations, session disclosure, and auto-approval gating (persistent always routes to manual, INV-004). `interactive_persona` forces persistent via conditional. - **`exclusivity`** — two instruments with distinct machine semantics (0.1.1, INV-016): 1. **Exclusivity modes** `none` | `category_exclusive` | `full_exclusive` (category exclusivity must enumerate categories) — a machine **can never grant an exclusivity mode**; any non-`none` mode routes to manual (INV-004(c)). 2. **NAICS-coded `competitor_blocks`** — expressly valid **alongside `mode: "none"`** ("non-exclusive with a competitor-category block", the standard ad-deal shape) and **machine-grantable** when the licensor opts in via `auto_approval_parameters.naics_competitor_blocks_permitted`, because the conflict check is deterministic: prefix match against the registry **block ledger** and the end-brand NAICS codes of the licensor's other active licenses; any indeterminacy routes to manual. Free-text `named_competitor_blocks` always route to manual; `scope` (exclusivity may be narrower than the grant); compensated `holdback_days_post_term`; `carve_outs` for pre-existing commitments. Both instruments are priced — pair with an `exclusivity_premium` component (which 0.1.1 makes auto-approvable *only* as consideration for auto-granted NAICS blocks) or state inclusion in a flat fee's `covers`. - **`output_caps`** — explicit `capped`/`uncapped` election (silence is impossible); output / impression / interactive-session ceilings; total-term or per-period scope; required `measurement_source` (whose numbers govern is the fight); and a required **`overage` policy** (`hard_block` | `bill_overage` | `requires_new_approval`) with pre-agreed overage rates — so success has a price instead of a breach claim. - **`platform_allowlist`** — how licenses stay **portable but controlled** (INV-008): reverse-DNS platform IDs from the published Semblance Platform Registry; per-platform `model_scope` (`all_models` — an explicit, disclosed election, never a wildcard default — or `listed_models` with exact IDs; `*` is unrepresentable); `model_version_policy` for point successors; `api_enforced` marks platforms that check certificates before generating (revocation technically enforced, not merely contractual); `unlisted_platforms_prohibited` *(const `true`)*; **`self_hosted_generation_permitted`** *(0.1.1: boolean, default `false`)* — a licensee-operated pipeline (e.g. a studio's internal voice pipeline fine-tuned on released reference audio) is now expressible: the pipeline must itself be **registered in the Platform Registry and enumerated in `platforms[]`** so allowlist/revocation/destruction semantics attach; usage reporting with `output_hashes` becomes mandatory and the license can never auto-approve (INV-025, INV-004(g)). This resolves the prior internal contradiction with `asset_handling` supporting voice-sample release and fine-tuning. - **`ai_disclosure`** — **`required` const `true`** (never waivable): closed method list (C2PA manifest + human-perceptible methods), optional exact `label_text` and placement rules, `eu_ai_act_art50_labeling` (forced `true` for worldwide grants by schema conditional; forced for any EU-member territory by INV-018), and persona session-start disclosure. - **`content_exclusions`** — two layers: 1. **`baseline_prohibited_uses`** *(const array, 13 entries)* — the platform-wide floor (NCII, minor sexualization, political deception, news simulation, medical/financial impersonation, fraud, hate/defamation, biometric spoofing, undisclosed synthetic media, training general-purpose models on reference data…), aligned with the TAKE IT DOWN Act, NO FAKES exclusions, and platform norms. Because it is a `const` array, **removing or altering any entry invalidates the instance** — the baseline cannot be negotiated away, and every executed license carries verbatim the full text it agreed to. (This resolves the agent draft's ToS-only approach in favor of the legal draft's evidence posture; evaluators still never branch on it — a const is free.) 2. **`excluded_codes`** — talent red lines from a closed **26-code** taxonomy (`energy_drinks_high_caffeine` graduated from custom usage in 0.1.1), machine-matched against the licensee's **`/content_declaration`** (0.1.1: the request-time declaration is now a first-class block — codes, timestamp, declarant; required for every auto/standing execution, INV-026; a false declaration is a for-cause breach). Deny rule: `declared ∩ (excluded − affirmative_permissions) ≠ ∅` (INV-017). Empty arrays are explicit elections — silence is impossible. **`affirmative_permissions`** (0.1.1) — codes the licensor *expressly permits*: a declared code listed here never denies, so "energy drinks explicitly welcome" is a recorded term resolving category-mapping ambiguity, not mere absence-of-exclusion. Disjoint from `excluded_codes`; can never whitelist the baseline. **`custom_exclusions`** (x-prefixed code + ≥10-char description) now carry a required **`enforcement`** election (0.1.1): `booking_match` customs constrain *which deals* may be booked and fail-close auto-approval (INV-004(a)); `deliverable_review` customs constrain *what the pixels contain* and are enforced at mandatory human output review — they coexist with auto-approved bookings (a "no before/after imagery" red line no longer bricks a talent's auto-approval profile). **0.1.2: the review surface must exist under any mode** — a schema conditional forces `deliverable_approval.required: true` with `generated_outputs` in scope whenever any `deliverable_review` custom (or any `prohibited_phrasings` entry) exists, so an exclusion's sole enforcement surface can never silently be absent. Recurring customs graduate into the taxonomy in later versions — the standards-body function of publishing the schema. The 26-code taxonomy itself is hoisted into **`$defs.content_code`** (0.1.2) — one source of truth for exclusions, permissions, and declarations. **Derived-code rule (0.1.2, INV-017):** first-person endorsement/testimonial framing **must** declare `first_person_endorsement_phrasing`; evaluators auto-derive it from spokesperson posture, and its omission is an incomplete declaration (a for-cause ground). Plus `endorsement_restrictions` ("appears in the ad" vs "personally recommends the product" is a real price/consent/FTC line — `prohibited_phrasings` are deliverable-review-enforced, consistent with `deliverable_review` customs) and `script_approval_required`. - **`style_guide_refs`** *(0.1.1, optional)* — hash-fixed references to character bibles / style guides the outputs must adhere to, bound into `grant_hash`: "consistent with Character Bible v3.2" becomes a hash-pinned machine term, enforced at deliverable review, and therefore compatible with agent-executed standing offers. - **`interactive_persona_terms`** (required when `interactive_persona` granted) — session disclosure *(const `true`)*, max concurrent deployments, transcript retention regime, runtime content-filter requirement. - **`credit`** — `credit_required` | `credit_optional` | **`anonymity_required`** (VTuber operators and body doubles need the *opposite* of attribution), plus orthogonal **`performer_anonymity_required`** (0.1.1): credit the character **and** never identify the embodied human performer — the standard VTuber shape, previously inexpressible. - **`sublicensing_allowed`** *(const `false` in v0.1)* — rights are personal to the verified licensee; the consent chain stays one hop long and NO FAKES-defensible. Agency flow is `is_agency_of_record`, not sublicensing. Structured programs (fan content, creator whitelisting, localization) are the first v0.2 candidate. --- ## `approval` — two surfaces, machine envelope, HITL **Booking approval (the deal)** vs **deliverable approval (the pixels)** are separate surfaces because agents and producers fight about both. - **`mode`**: `manual` | `auto_approve_within_parameters` | `standing_offer`. Approval latency kills conversion — hence SLAs as deal terms. - **`booking_approval_sla`**: response hours + `on_timeout` ∈ {`auto_decline`, `quote_expires`, `escalate_to_agent_of_record`}. There is **deliberately no auto-approve-on-silence** — consent must be affirmative. - **`auto_approval_parameters`** (required for non-manual modes) mirrors the grant 1:1: category/media/channel subsets, **language envelope (`allowed_languages`, prefix-matched) and distribution-platform envelope (`allowed_distribution_platform_ids`) — both 0.1.1**, territory envelope, ISO-duration term ceiling, **fee floor (`min_total_fee`, blocks underpriced agentic spam) and ceiling (`max_total_fee`, big deals deserve human eyes)**, output/impression ceilings, platform allowlist, `allowed_fee_structures` (revenue share is *not in the enum* — rev-share auto-approval is inexpressible; 0.1.1 adds `exclusivity_premium`, lawful only as the price of auto-granted NAICS blocks), **`exclusivity_permitted` const `false`** (a machine can never grant an exclusivity *mode*) with the separate 0.1.1 opt-in **`naics_competitor_blocks_permitted`** (deterministic block-ledger check, INV-016(2)), minimum verification tier, licensee allowlist/**blocklist (blocklist hits are rejected with no human retry channel; an allowlist *miss* routes to manual — the buyer simply isn't pre-authorized, 0.1.2)**, a **`rate_limit` circuit breaker** against runaway agents, `require_escrow`, **`reference_release_permitted` + `reference_release_allowed_asset_classes` (0.1.2, INV-004(m))** — releasing raw reference material or authorizing fine-tuning is never an implicit side effect of machine execution; without this per-asset-class opt-in, any such request routes to manual — **`max_interactive_sessions`** (the step-11 envelope mirror; note personas themselves never auto-approve), and **`deliverable_review_retained`** (auto-approve the deal while keeping human review of the content — recommended default, and the mechanism that lets deliverable-review-enforced red lines coexist with auto-approval). `interactive_persona` is absent from the allowed-category enum entirely. - **`human_confirmation_threshold`**: above it, even in-envelope requests need a licensee-side natural person to confirm before capture (ACP/AP2/Visa-TAP HITL pattern); the licensor side stays auto. - **`standing_offer`**: published, pre-consented offer with `valid_until` and `max_total_executions` (an execution budget checked atomically) — the narrowest, most commoditized shape. - **`deliverable_approval`**: scope (scripts/prompts/outputs/final cut), included rounds, per-round SLA, **`deemed_approved_on_timeout` default `false`** (talent-protective; brands may negotiate `true`, and the trade is visible in the paper), and `post_approval_edit_rules` (cutdowns of an approved master OK; new creative not). This is where SAG-style creative approval lives. The full evaluation order is normative in `x-auto-approval-algorithm` (**16 deny-fast steps**, 0.1.2). Outcome semantics are explicit: `route_manual` preserves the human retry channel; `reject` is terminal and reserved for blocklist hits (step 4), content-exclusion denials (step 7 — the licensor excluded it; there is nothing for a human to reconsider), and execution-integrity failures (step 16). Step 11 defines absent-ceiling semantics (a `capped` request omitting a ceiling the envelope bounds routes to manual — never a vacuous pass); step 16 executes atomically and **serialized per licensor**, committing the INV-016 block ledger (which spans `executed`+`active` licenses plus live holdback windows) in the same transaction as escrow capture, grant-hash (INV-007) and standing-terms (INV-029) verification, under the execution idempotency key. --- ## `pricing` — composable, integer, choreographed - **`currency`**: one ISO 4217 code per license (INV-003). No FX inside evaluation; mismatch ⇒ manual. FX risk sits with the payer. - **`fee_components[]`** — a **discriminated array (on `model`), not a single price-model enum**, because real deals are hybrids (flat setup + rev-share + minimum guarantee is one license): - `flat_fee` — fixed amount + plain-language `covers` (prevents scope disputes). - `per_output` — integer rate per typed unit (video, image, audio minute, voice line, render, stream hour, interactive session), committed prepaid units, billing mode. **0.1.1 semantics (INV-027):** `committed_units` ≤ `output_caps.max_outputs`; units beyond the commitment but within the cap bill at the component rate per `billing`; units beyond the cap are governed exclusively by `output_caps.overage` — prepayment and ceiling are distinct numbers. - `per_impression` — rate **per 1,000 impressions** in minor units (no fractional cents) + **required `measurement_source`** (whose numbers govern is the fight, not the CPM; `licensee_attested` requires audit rights, INV-020) + committed impressions (multiples of 1,000). - `revenue_share` — **integer basis points**, basis class, and an always-required **`basis_definition` (≥ 20 chars)** — "net" without a cost definition is how talent gets zero; the undefined-royalty-base dispute is precluded at the schema layer. Optional minimum guarantee (recoupable or stacking — no agent accepts naked rev-share) and payout cadence. Required for interactive personas (INV-006); never auto-approvable. - `exclusivity_premium` — **separately stated** (fixed fee and/or bps uplift) so renewals can reprice exclusivity without reopening the deal; 0.1.1 extends it to price NAICS competitor blocks under `mode: "none"`, so a category block's consideration is never implicit. - **`election`** *(0.1.1, any component)* — marks a component that entered the deal **by election at execution** (`licensee_elected` / `licensor_elected` / `mutual_election` + timestamp) rather than as a mandatory term: the schema-native encoding of elective pricing tiers (e.g. an optional rev-share tier taken at signing). The executed document records the elected outcome; the untaken alternative lives in the immutable quote. Full elective menus are a v0.2 candidate. - **`payment_schedule`** — milestone ladder over fixed fees (bps summing to 10000, INV-014; the classic 50/50 is `5000/5000`), triggers from execution through net-days, late-payment interest in bps. - **`escrow`** — the money choreography that lets strangers transact: `provider` (Stripe Connect in v0.1; agentic rails land as new enum values), `capture_at` const `execution` (atomic with the transition), and the **state machine `unfunded → funded → partially_released → (released | refunded | disputed)`**. A license cannot go `executed → active` until funded (INV-015); `disputed` freezes releases. **`escrow.state` is expressly exempt from the amendment rule (0.1.2)**: it is operational payment state living inside `/pricing` for single-source-of-truth reasons; its transitions never increment `license_version` or trigger re-consent. - **`kill_fee`** — staged ladder by deal stage (typical 25/50/100/100 ⇒ 2500/5000/10000 bps): talent's calendar and exclusivity were committed even if no pixel shipped. Plus `exclusivity_kill_addon` for blocked-category opportunity cost. Waived on licensor at-will revocation. - **`options[]`** — pre-negotiated renewals/expansions with exercise windows keyed to expiry, pricing methods (fixed / bps uplift / rate card / match-offer), an `option_fee` paid now, and the crucial **`licensor_can_decline`** flag (true option vs negotiation right). Options are how 12-month deals become 5-year relationships **without violating the 10-year human cap** — each exercise papers a new document with fresh consent (INV-021). - `total_guaranteed_fee`, `platform_fee_bps` — transparency and threshold bases. **0.1.2 money-math rules:** `total_guaranteed_fee` is no longer free-standing — it MUST equal the machine-computed guaranteed sum over the components (INV-033); and every basis-point computation in the document (splits, milestones, platform fee, rev-share settlement, kill fees, uplifts, interest) follows one normative rounding rule (INV-032): **integer floor per line, residual remainder to the last entry of the ordered list** (or retained by the payer for single-payee computations) — the odd cent of a 50/50 split on $0.01 has one answer. The `exclusivity_premium.uplift_bps` base is defined as the guaranteed total of the *other* components, excluding the premium itself and uncommitted metered/share upside. `payment_schedule` milestones with trigger `net_days_after_invoice` must state `net_days` (schema conditional, 0.1.2); billable overage policies must state the rate matching each set cap (`overage_rate_per_output` when `max_outputs` is capped, per-1000 rate when `max_impressions` is — schema conditionals, 0.1.2). --- ## `deliverables` — the statement of work *(optional block)* Typed line items (`video_asset` … `persona_endpoint_access`, `live_stream_session` — service enablements, not files), quantities, technical spec (durations, resolutions, aspect ratios, **languages as BCP 47 — localized voice in a new language is a grant question, not a technicality**, formats), due dates, and an acceptance window (spec failures only; silence past the window is acceptance). Mirrors, in production detail, the consent record's enumerated deliverables (INV-019). --- ## `usage_reporting` Anything metered, capped, shared, or exclusive needs numbers flowing back. `required` is forced `true` for per-impression/rev-share economics (INV-005); interval + grace window as ISO durations; closed metric list (including `output_hashes`, which feed the enforcement index); `delivery_method` const `report_usage_api` (machine-ingestible, not spreadsheets); machine-enforceable `late_report_policy` (suspension is an access control, not a status transition); audit rights with a cost-shift threshold (mandatory when measurement is licensee-attested); and a records-retention period that **does not extend biometric retention** (that is exclusively `asset_handling`'s job). --- ## `union_rider` — SAG-AFTRA Digital Replica interface `covered` is **always required** — `false` is an explicit onboarding assertion, never a default, so evaluators never guess. When `true` (which disables auto-approval, INV-004): - **`agreement`** — the governing CBA: 2023 TV/Theatrical, **2025 Commercials Digital Replica Rider**, or the **2025 Interactive Media Agreement**, whose third-party-authorized-platform carve-out for independently created replicas is the provision under which Semblance-brokered union interactive work proceeds. ACTRA/Equity UK enumerated for non-US talent. - **`replica_classification`** — `employment_based_digital_replica` vs `independently_created_digital_replica`: the 2025 IMA distinction that selects which CBA consent and compensation rules apply. Required, because mis-classification is the compliance failure mode. - **`rider_ref`** — hash-fixed reference to the generated, executed digital-replica rider that travels with the contract (producers can hire union talent through the platform without breaching their CBAs — compliance-as-a-product). - **`union_notice_given`**, **`scale_minimums_compliance`** *(const `true` — a below-scale union license is schema-invalid)* with 0.1.1's **`scale_rate_evidence_ref`** (hash-fixed rate worksheet turning the bare assertion into an auditable record), **`contributions_handled_by`** (0.1.1 clarifies: producer-paid P&H is paid **on top of** `/pricing` fees directly to the plans and must *not* be encoded as a `payout_split` carve-out; the `union_trust` split role is only for platform-remitted, gross-compensation-computed contributions), plus 0.1.1's **`ph_contribution_rate_bps`** and optional **`ph_contribution_amount`** making the P&H economics machine-readable, `producer_signatory_id`. Semblance never induces breach of a CBA — by construction. --- ## `asset_handling` — BIPA-grade biometric & derived-model regime Two distinct instruments by design: **`written_release_ref`** is the **BIPA §15(b) written release covering the *data*** (collection, storage, this specific disclosure), separate from the consent record covering the *use*. Conflating them is the classic BIPA pleading error, and BIPA's private right of action is existential-risk-grade. - **`biometric_data_present`** *(0.1.1, required)* — whether the reference materials include biometric identifiers of a living human (an embodied performer's voice/performance data counts; a pure Live2D character model does not). The §15(b) `written_release_ref` is required **iff** this is true — the block is no longer BIPA-framed for pure character-IP releases, while the destruction regime applies to any released reference material regardless. - **Release-before-collection, structurally (0.1.2, INV-030):** when biometric data is present, the document also requires **`reference_materials_collected_at`** (first collection timestamp for the data subject) and **`written_release_signatory`** (the natural person who signed, in capacity `licensor_self` or `embodied_performer` + `performer_id` — a §15(b) release is personal to the data subject; owner/representative capacities are deliberately unrepresentable). INV-030 orders them: `written_release_ref.captured_at` ≤ every recorded collection timestamp, including liveness checks. A document recording collect-first-paper-later — the textbook per-violation BIPA pleading pattern — is non-conformant. - **`raw_biometric_reference_released`** — default posture `false`: raw reference material (face captures, voice samples, motion data, trained models, LoRAs, model files) never leaves the Rights Vault; licensees receive outputs and usage rights. If `true`, the schema **requires** `released_asset_classes` and the full **`destruction_obligations`** regime: - `destroy_on` *(const)*: earliest of purpose-satisfied / expiry / revocation / termination — §15(a) retention logic made bilateral; 0.1.1 adds **`purpose_satisfied_definition`** (who declares the purpose satisfied, and when — content-complete vs full patch/DLC term is now a written term; absent = no later than final delivery acceptance or term end); - **`max_retention_days` ≤ 30** for operational systems (longer is inexpressible); **`cold_backup_purge_days` ≤ 90** (0.1.2, down from 365 — BIPA §15(a) recognizes no backup exception, so the window is exposure, not entitlement) for immutable cold backups / VCS history where in-place deletion is infeasible — access-restricted in the interim, restoration prohibited; **any value over 30 requires `backup_purge_infeasibility_ref`** (documented technical infeasibility + counsel sign-off, schema conditional); - NIST-grade `destruction_method`; - **`destruction_certificate_required` const `true`** (signed certificate identifying every purged system) with 0.1.1's **`applies_to_recipients`** *(const `licensee_and_each_receiving_platform`)*: certificates are owed severally by the licensee **and every platform/pipeline that received the material** (INV-028) — no more ambiguity about who signs when reference data went to HeyGen or a studio pipeline; - `audit_right`. - **`no_sale_of_biometric_data`** *(const `true`)* — §15(c): licensing outputs is not selling biometric data, and actual sale is unrepresentable. - **`derived_models`** — closes the fine-tune loophole: any LoRA/voice clone/checkpoint created to fulfill this license is scoped **const `this_license_only`** (never other clients, never general model improvement — the baseline separately bans training general-purpose models on reference data), with required post-term disposition (**certified destruction or return to the Rights Vault**). **0.1.1 classification rule:** a *derived model* is an asset capable of generating **new** likeness content; rendered/compiled assets embedded in delivered outputs (a game's baked face rig, dialogue portraits, final media) are **delivered outputs** governed by delivered-outputs treatments and any survival election — never by this destruction regime. A *surviving* licensed derived model — a vastly bigger grant — is deliberately not expressible in v0.1. --- ## `revocation` — the crisp two-effect split Drafted for judicial application *and* machine execution (PLAN §3.5's distinction, made structural): 1. **`new_generations`** *(const `"cease_immediately_all_platforms"`)* — at the effective time, all new generation, re-rendering, and derivative production stops on every allowlisted platform, propagated by `license_id` via certificate checks and vault key revocation. A hard-coded platform guarantee, not a term. 2. **`delivered_outputs_treatment`** — a mandatory three-way **election at execution**: `licensee_retains_through_term` (brand-safe default; campaigns in flight are safe) | `wind_down_then_cease` (+ required `wind_down_days`, 1–180) | `immediate_cessation` (premium-priced). A court applies the election — no "reasonableness" litigation. Priced honestly (commercial lens): **`at_will_permitted`** is an explicit boolean — an at-will-revocable license is worth less, and buyers pay for certainty; both shapes are expressible. `notice_days` (≤ 90) and **`refund_treatment`** (`pro_rata_refund_of_unearned_fees` | `no_refund` | `per_component_schedule`) are (0.1.1) **required only when `at_will_permitted` is true**; when false they are optional and, if present, govern statutorily mandated consent-withdrawal only — no more undefined semantics on for-cause-only deals. `licensee_kill_fee_waived_on_at_will` (it should be) and an optional `revocation_fee` keep revocation *real* while its cost is a number, not a lawsuit. Guardrails (legal lens, all consts): `for_cause_overrides_to_immediate` (INV-022 — for-cause always permits immediate takedown, no refund), `biometric_destruction_on_revocation` (revocation always triggers the destruction regime), `statutory_rights_non_waivable` (state digital-replica law, BIPA, GDPR, or an enacted NO FAKES Act control if broader), and required enumerated `surviving_obligations`. **`for_cause_grounds` is required with a non-negotiable minimum set (0.1.2)** — `content_exclusion_breach`, `use_outside_granted_scope`, `unapproved_platform_or_model` must always be enumerated, so INV-022's immediate-takedown override always has defined contractual triggers. `persistent_service_wind_down_days` is capped at 30 — for personas, the ongoing use *is* the harm surface. **verify_license outcome vocabulary (0.1.2, INV-008)** — closed and normative, computed from *(status × delivered-outputs treatment × window)*: `licensed_active`, `licensed_pre_activation` (delivered under `generation_may_precede_start` while `executed`), `licensed_runoff`, `licensed_post_term_survival`, `licensed_revoked_retention` (not-for-cause revocation with `licensee_retains_through_term`, inside the original term), `licensed_wind_down_until` (with the deadline), `licensed_terminated_retention` (licensee cancellation with kill fee paid), and `unlicensed` (everything else — expressly including all delivered content of a *for-cause* revocation). Reporting lawfully retained post-revocation/post-termination content as `unlicensed` is itself a conformance failure: wrongful-takedown protection is part of the spec. --- ## `provenance`, `warranties`, `governing_law` **`provenance`** — platform floors, all const where it matters: every delivered asset carries **C2PA Content Credentials** and an invisible **watermark keyed to `license_id`** (wild-asset attribution is a single lookup — or the asset exposes itself as unlicensed); **`verify_endpoint`** is public, free, and unauthenticated by design (platform adoption of this check is the long-term moat); optional registration of final published-version hashes; delivery only via signed URLs or platform-direct-with-certificate-check (off-platform delivery breaks provenance and is not offered). The schema is public — competitors adopting it inherit the ethics floor. **0.1.1, product-embedded pipelines:** where outputs compile into formats that cannot carry credentials (engine assets, runtime-rendered media), the C2PA obligation attaches to the delivered asset *package* and the licensee files an **`embedded_use_manifest`** per product release (build/patch/DLC), keyed to `license_id` — how embedded/runtime uses stay traceable and how `verify_license` resolves shipped-game content (INV-024). Deeper build-level provenance is a v0.2 candidate. **`warranties`** — the indemnification triangle as three consts: licensor warrants rights ownership (characters: complete chain of title, no unlicensed third-party elements); the platform warrants the chain of consent; the licensee warrants use compliance. Optional E&O reference. **`governing_law`** — jurisdiction/venue/dispute path (platform human-review process first by default). Non-waivable statutory rights apply regardless of this election. --- ## `execution` *(execution-time block)* Created at the `executed` transition: `executed_at`, the **`certificate_id`** (the signed, publicly verifiable certificate embedded in C2PA manifests and resolvable at the verify endpoint), both **signature refs** (for agentic execution, the mandate-backed acceptance record), the hash-fixed **`contract_document_ref`** (the generated human-readable contract with consent instrument and any union rider — *the machine document governs evaluation; on conflict, the signed instrument controls in court, and says so*), the payment reference, **`human_confirmation`** (0.1.2, INV-035: `confirmed_by` / `confirmed_at` / `method_ref` — required whenever the fee crossed `human_confirmation_threshold` or the principal's `human_confirmation_required_above`; the dispute-grade artifact of the agentic HITL step), and the **delivery `asset_manifest`** (per-asset SHA-256 + watermark id — what `verify_license` matches wild content hashes against, INV-008). Escrow disposition lives solely in `pricing.escrow.state` (single source of truth; the agent draft's duplicate escrow field was removed in the merge). --- ## Shared definitions (`$defs`) | Def | Type discipline | |---|---| | `money` | integer `amount_minor` + ISO 4217 `currency`; no floats anywhere | | `basis_points` | integer 0–10000 for every percentage (splits, shares, uplifts, interest); rounding per INV-032 | | `territory` | `worldwide`/`include`/`exclude` + ISO 3166-1 alpha-2 sets (pinned to the registry snapshot, INV-036), plus geo-targeting flags | | `iso_duration` | ISO 8601 durations; week designators are exclusive (`P2W` valid, `P1W2D` rejected — 0.1.2) | | `utc_datetime` | (0.1.2) every timestamp: RFC 3339 date-time with **required trailing `Z`** — non-UTC offsets are unrepresentable | | `evidence_ref` | ref id + **required SHA-256** + URI: every compliance instrument is hash-fixed for authentication in a dispute | | `age_attestation` | (0.1.2) `date_of_birth_verified` const + hash-fixed `age_verification_ref` + `verified_at` — per natural person (INV-012) | | `content_code` | (0.1.2) the 26-code content taxonomy, hoisted to a single source of truth for exclusions/permissions/declarations | | `verification_tier` | 5 tiers, lexicographic order == trust order; single-digit namespace — expansion past `tier_9` requires zero-padded renaming in a breaking version | | `use_category`, `media_type`, `channel`, `status` | closed taxonomies; no wildcard values | | `licensor`, `licensee`, `consent` | party and consent structures described above | Identifier scheme: `lic_`/`tal_`/`lce_`/`cst_`/`cert_`/`quo_`/`bkg_`/`ast_`/`off_`/`evt_`/`per_` + 26-char Crockford-base32 ULID. --- ## Key merge resolutions (and why) | Conflict | Resolution | |---|---| | Prohibited-use baseline: per-license const array (legal) vs ToS-only (agent) | **Const array kept** — every executed license carries the text it agreed to (evidence posture); consts cost evaluators nothing | | Term: `duration_months` + renewal enum (legal/commercial) vs dates-only + `auto_renew:false` (agent) | **Dates-only** — redundant duration invites inconsistency; the 10-yr cap is INV-001 (all three drafts conceded date math is validator-level); auto-renewal for anyone is out in v0.1, characters renew via one-call new documents | | Perpetual character terms (commercial) | **Not representable in v0.1**; open question below | | Model allowlist: every model enumerated, no `*` (legal) vs explicit `all_models\|listed_models` (agent) | **Agent's enum** — `all_models` is an explicit, disclosed consent election, not a wildcard ambiguity; `*` remains unrepresentable | | Sublicensing: prohibited const (legal, agent) vs structured programs (commercial) | **Const `false`**; fan-content/whitelisting programs are the flagged v0.2 candidate; agency flow via `is_agency_of_record` | | Exclusivity in auto-approval: `max_exclusivity_level` (commercial) vs const false (legal) | **Const `false`** — machines never grant exclusivity *modes*; 0.1.1 carves out deterministic NAICS competitor blocks under the separate `naics_competitor_blocks_permitted` opt-in (INV-016(2)) | | Money/percentages: major-unit floats (legal/commercial) vs integer minor units/bps (agent) | **Integer everywhere** | | Escrow state: agent's `execution.payment.escrow_state` vs commercial's richer machine | **Commercial's 6-state machine**, single-sourced under `pricing.escrow` | | Derived models: `persistent_licensed` fine-tuning (commercial) vs `this_license_only` const (legal) | **Legal wins** — surviving LoRAs are a bigger grant than v0.1 prices; open question | | Verification tiers: 3 naming schemes | **Agent's lexicographic `tier_N_*`** scheme | | Revocation naming/elections | **Legal's three-way delivered-outputs election + const cease-immediately**, plus commercial's at-will pricing flags and agent's optional `revocation_fee` | ## Known limits (application-level, normative via x-invariants) Date arithmetic (INV-001/002, anchored, date-inclusive), payout/milestone sums (INV-013/014, rounding per INV-032), escrow gating (INV-015), cross-license exclusivity/block-ledger conflicts (INV-016, executed+active+holdback ledger with serialized commit), EU-member detection for country lists (INV-018), consent-instrument verbatim match (INV-019, RFC 4647 language matching), grant-hash verification (INV-007), territory set expansion (INV-009), approval-hash verification and advance-instrument timing (INV-023), standing-terms byte-equality (INV-029), release-before-collection ordering (INV-030), statutory-bases territory lint (INV-031), the `total_guaranteed_fee` equality (INV-033), platform-side agent-channel enforcement (INV-034), HITL confirmation artifacts (INV-035), reference-registry snapshot pinning (INV-036), the verify_license outcome table (INV-008), post-term survival reporting (INV-024), self-hosted pipeline registration and reporting (INV-025), content-declaration matching, derived endorsement codes, and affirmative-permission disjointness (INV-017/026), committed-units-vs-cap ordering (INV-027), and per-recipient destruction certificates (INV-028) cannot be expressed in JSON Schema and MUST be enforced by the platform validator and its conformance suite. ## What 0.1.2 changed (adversarial-review revision) Applies every finding of the two adversarial reviews (legal completeness; implementability vs. the auto-approval evaluator): 1. **Embodied performers are first-class sub-parties** (`character_provenance. embodied_performer`): own `per_` id, age attestation, KYC/liveness, intended-use consent, data-subject BIPA release, mandatory `payout_split` role — and **INV-001's 10-year cap now covers any grant with a living embodied performer**, closing the character-wrapper evasion. 2. **Structured age attestation** (`$defs.age_attestation`) for every verified natural person: human licensors, estate representatives (new required `estate_authority.representative`), character-owner signatories (new required `owner_signatory`), embodied performers (INV-012). 3. **Term-cap math defined normatively**: anchor = earliest of consent capture / execution / start_date; end_date inclusive; check `((end_date + 1 day) − anchor) ≤ cap` (INV-001, step 10) — kills both the postdated-start extension and the off-by-one-day pass. 4. **Advance-mode representation gate**: unrepresented human licensors cannot publish auto/standing envelopes (root conditional; §927 safe-harbor posture). 5. **BIPA ordering made structural** (INV-030): `reference_materials_collected_at` + `written_release_signatory` (data-subject-only capacities), release-before-collection invariant; `cold_backup_purge_days` ceiling cut to 90 with counsel-signed infeasibility documentation required over 30. 6. **`campaign` (with `product_or_service`) required for every non-manual executed license**; both auto examples now declare `first_person_endorsement_phrasing`; INV-017 gains the derived-endorsement-code rule. 7. **Standing-terms hash binding** (`advance_authorization.licensor_terms_ref` + INV-029 byte-equality over all talent-fixed blocks) — the booking assembler is no longer trusted. 8. **Closed verify_license outcome vocabulary** (INV-008): eight statuses over (status × treatment × window), including revoked-retention, wind-down, pre-activation, and terminated-retention — lawfully retained outputs can never report as unlicensed. 9. **State machine completed**: `executed → expired` clock transition, `pending_approval → declined` on `sla_timeout_declined` (actor `system`), `expired → terminated` for runoff/survival abuse (INV-024 survival revocation); activation defined for service-enablement deliverables. 10. **INV-004(m) + envelope opt-in** (`reference_release_permitted` / `reference_release_allowed_asset_classes`): raw-reference release and fine-tuning route to manual unless the licensor pre-authorized the exact asset classes. 11. **Money math pinned**: INV-032 rounding/remainder rule (floor per line, remainder to the last entry; uplift base defined), INV-033 `total_guaranteed_fee` equality, overage-rate and `net_days` conditionals. 12. **Algorithm split and renumbered (16 steps)**: blocklist reject (4) vs allowlist-miss route-manual (5); category-subset manual route (6) vs exclusion-intersection deny (7); absent-ceiling semantics + `max_interactive_sessions` mirror (11); serialized ledger commit (16). INV-016 ledger scope: executed+active plus live holdbacks. 13. **Integration hardening**: separate `booking_idempotency_key` / `execution_idempotency_key` scoped per (licensee, operation) with a retention window (INV-010); `quote_expires_at`; `reference_registry_snapshot` + INV-036 table pinning. 14. **Type discipline**: `$defs.utc_datetime` (required trailing Z) for every timestamp; ISO-duration week-mixing rejected; RFC 4647 basic filtering + lowercase primary subtags for all language tags; `$defs.content_code` hoisted; zero-padded tier naming rule documented. 15. **Revocation floor**: `for_cause_grounds` required with a mandatory minimum trio; deliverable-review customs and `prohibited_phrasings` force the review surface to exist under any mode. 16. **`pricing.escrow.state` expressly exempted** from the license_version amendment rule (documented, not moved — single source of truth preserved). 17. **Agent-channel integrity** (INV-034) and the **HITL confirmation artifact** (`execution.human_confirmation`, INV-035). 18. **INV-031 statutory-bases territory lint**; all four examples cite territory-complete bases. 19. **Examples**: new `estate-archival.json` (CA post-mortem estate, manual, INV-002 exercised); AI ad re-cut counsel-represented with corrected BIPA timeline and complete declarations; game NPC re-priced well above scale with union-counsel note, term corrected for inclusive cap math, backups shortened; VTuber carries the full performer sub-party and performer payout split. All hashes are real JCS SHA-256 values; everything validates under ajv. ## What 0.1.1 changed (pressure-test revision) Closed from the three gap reports (AI ad / game NPC voice / VTuber standing offer): 1. **Non-exclusive + NAICS competitor block** is now a coherent, auto-approvable encoding (`mode:"none"` + `competitor_blocks` + `naics_competitor_blocks_permitted`; INV-004(e), INV-016(2)); the block is priced via `exclusivity_premium`. 2. **Custom exclusions elect an enforcement surface** (`booking_match` vs `deliverable_review`); only booking-match customs fail-close auto-approval. 3. **`grant.distribution_platforms`** — distribution allowlist as a machine term, mirrored in the auto-approval envelope. 4. **Advance authorization** (`consent.advance_authorization`, INV-023) defines the consent story for machine execution: pre-signed instrument, `approval_hash` over JCS(/approval), `captured_at` = signing time, `licensor_signature_ref` → the advance instrument; `licensee_identity_disclosure` replaces the const-true boolean; `/campaign` moved out of the hashed grant so per-execution metadata never breaks INV-007. 5. **Channel boundary definitions** (YouTube pre-roll = `social_paid`; `in_game` excludes store pages/trailers) and the **video+voice pairing rule** for talking outputs. 6. **`grant.languages`** as a required grant dimension (prefix-matched, enveloped, INV-019). 7. **Post-term survival of embedded outputs** (`delivered_outputs_survival`, INV-024) — shipped game builds stop reading as infringing on day one of year eleven. 8. **Self-hosted generation** expressible as a registered-pipeline election (INV-025). 9. **Elective fee components** via `election` markers; committed-units vs caps disambiguated (INV-027). 10. **Derived-model vs delivered-output classification rule**; backup/VCS purge window (`cold_backup_purge_days`); `purpose_satisfied_definition`; per-recipient destruction certificates (INV-028). 11. **Union economics**: `ph_contribution_rate_bps` / `ph_contribution_amount`, `scale_rate_evidence_ref`, producer-paid-P&H encoding rule. 12. **`content_declaration`** block + **`affirmative_permissions`** (+ taxonomy code `energy_drinks_high_caffeine`), INV-017/026. 13. **Entity licensors**: `entity_licensor_not_applicable` representation, Canadian/Japanese/ contract-only `statutory_bases`, character-owner signatory guidance. 14. **`credit.performer_anonymity_required`**; **`embedded_use_manifest`**; `biometric_data_present` gating the BIPA release; at-will-conditional `notice_days`/`refund_treatment`; `generation_may_precede_start`; `x_notes` annotations; auto-execution actor rule (initiating `licensee`/`licensee_agent`, never `system`). ## Known limitations / v0.2 candidates Gaps surfaced by the pressure tests that are **deliberately deferred**, with the 0.1.1 stopgap noted: 1. **Elective/menu pricing at quote time.** 0.1.1 records only the *elected outcome* (`fee_components[].election`); a first-class offer menu (alternative component sets, tier elections, contingent fees) belongs in the quote/booking schema split (item 7 below). 0.1.2's `integration.quote_expires_at` is the minimal in-document TTL until that lands. 2. **Statutory adequacy of class-level advance consent.** INV-023 defines the *mechanics* (pre-signed instrument, envelope hash, standing-terms binding, execution-time disclosure); whether class-level description satisfies each jurisdiction's "reasonably specific description" standard for every deal shape needs the fifty-state matrix and counsel review. **0.1.2 stopgap:** human licensors cannot publish auto/standing envelopes without `legal_counsel` or `union_cba_coverage` representation (schema conditional), so the unrepresented-human + class-consent configuration — the one with no §927 fallback — is unrepresentable until the matrix resolves it. Conservative jurisdictions may still require manual mode even with counsel. 3. **Deep game-pipeline provenance.** `embedded_use_manifest` bridges builds/patches; per-asset watermark survival through engine compilation, wild in-game audio matching, and runtime-rendered C2PA remain open engineering/spec work. 4. **Full union P&H remittance modeling** (contribution ledgers, plan remittance evidence, per-session-vs-buyout bases). 0.1.1 records rate/amount/payer only. 5. **Distribution-platform registry.** `distribution_platforms` ids assume a published Semblance Distribution Registry (like the Platform Registry); publish both with the schema. 6. **`in_game` sub-surfaces** (store pages, in-engine promotional capture, UGC/streamer programs) as closed channel or exclusion codes — currently handled by the normative channel definition plus custom exclusions. 7. **Marketing/promotional-use taxonomy.** "No marketing use" is structurally out-of-scope when `advertising` isn't granted, but a closed promotional-use code set would remove the remaining custom-exclusion reliance. 8. **Language dialect/locale matrices** (prefix matching is deliberate; per-locale union rates and moral-rights nuances may need structure). 9. **Post-term survival for non-embedded persistent artifacts** (e.g. printed matter, OOH archives) — 0.1.1 scopes survival strictly to product-embedded copies. 10. **Reference Registry snapshots as published artifacts.** 0.1.2 pins evaluation to `integration.reference_registry_snapshot` (INV-036), but the snapshot files themselves (ISO 3166-1 set, EU/EEA membership, ISO 4217 exponents, NAICS 2022 tables) must be published and versioned alongside the Platform Registry — publication is platform work, not schema work. 11. **Union-counsel review of live union deals.** The game-NPC example is re-priced well above the computed scale floor with a documented classification memo (its `x_notes`), but real union bookings still need per-deal union-counsel review of the worksheet math and replica-classification reading; `scale_minimums_compliance` remains an assertion whose evidence lives in `scale_rate_evidence_ref`, not a computation the schema can check. 12. **Estate-regime renewal mechanics.** INV-002 now requires re-verification of `post_mortem_rights_expire_at` at each execution/renewal, but renewal-contingent regimes (NO FAKES 10 + 5) may eventually deserve first-class renewal-state fields rather than a re-verified date. ## Open questions for v0.2 1. Structured sublicense programs (fan content, creator whitelisting, localization, platform-catalog integration) with sublicense registration into the enforcement index. 2. Perpetual or auto-renewing terms for `character` licensors (pure IP licensing) without weakening the human/estate posture. 3. Persistent licensed derived models (`persistent_licensed` fine-tuning) as a priced, vetted grant. 4. Exclusivity ledger: registry-level cross-license state so category exclusivity can be machine-checked (until then, exclusivity ⇒ manual). 5. Multi-currency talent pricing (per-currency parameter sets vs platform FX at quote time). 6. ISO 3166-2 subdivision territories (US states matter for right-of-publicity nuance). 7. Splitting profile-terms / booking-request / executed-license into three schemas sharing `$defs` (this document doubles as terms-and-license in v0.1). 8. Jurisdiction-specific consent formality variants from the fifty-state matrix; moral-rights handling for non-US territories; multi-licensor ensemble grants. 9. Self-hosted generation pricing/audit depth (0.1.1 made it *expressible* as a registered pipeline with mandatory reporting, INV-025; premium pricing norms and audit tooling remain). 10. Additional escrow/payment rails (ACP, UCP, AP2, x402) as `escrow.provider` values.