AI映画産業発展白書2026

FIGURE 1 · From Fixed Files to Cultural Computing Systems

1.1 From generative supply to systems competition

The earliest market narrative focused on the falling price of producing a plausible image. That metric is incomplete. A professional project pays for intention, consistency, selection, revision, performance control, sound, rights clearance, documentation and release responsibility. As the number of candidate outputs grows, the cost of deciding what can be accepted—and proving why—may increase. Management should therefore track total cost per accepted finished second, the number of revisions required, control coverage and the proportion of assets that remain portable when a model is replaced.

1.2 Five structural migrations

Production moves from scarce shooting capacity to scarce acceptable output; organisation moves from fixed departmental hand-offs to orchestration across people, models and tools; distribution moves from content scarcity to trustworthy discovery; product design moves from a final master toward a maintained relationship; and the asset base moves from one completed title to a reusable family of worlds, characters, workflows and records. Each migration creates a different competitive advantage and a different failure mode.

1.3 The decision implication: build a replaceable capability stack

No studio should make its creative memory dependent on one provider's interface or one generation model. A resilient stack separates story and world state, asset storage, model routing, evaluation, rights, provenance and delivery. Models can then be substituted as performance, cost or policy changes. The durable advantage is the quality of the organisation's creative decisions and institutional memory, not temporary access to a single benchmark leader. *TABLE 1 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 2 · The Dual-Axis Map of AI-Native Cinema

2.1 The P axis: five dimensions of creative agency

P is evaluated across narrative, image, sound, character and interaction. Reviewers ask whether AI affects the structure of the work, not whether an AI tool touched a file. P0 denotes no generative role; P1, local assistance; P2, deep participation in a core stage; P3, substantive leadership across the weighted creative dimensions; and P4, a work whose artistic identity is itself a generative system. The 51%+ marker indicates entry into P3 when disclosed weights and evidence show that AI materially leads the relevant dimensions.

2.2 The R axis: runtime generation

R0 is fully pre-produced playback. R1 adapts before presentation; R2 combines pre-produced and generated components; R3 generates most experienced content in response to the session; and R4 is continuously generated, stateful and not exactly reproducible. Because runtime generation increases operational risk, R must be assessed independently from P. A high R score requires latency budgets, state controls, safety policies, logging, fallbacks and a clear operator of record.

2.3 The conductor principle and responsibility

Human authorship in an AI-native process often resembles conducting: defining intention, selecting systems, constructing constraints, deciding what to reject, arranging outputs and accepting responsibility. This does not make every prompt an authored work. It means creative control must be evidenced through choices and their consequences. The relevant record is a chain of decisions—not a screenshot of a prompt and not a claim that the model acted alone. *TABLE 2 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 3 · The G0–G5 Evolution of AI Audiovisual Media

3.1 What G0–G5 mean

G0 is conventional audiovisual production. G1 uses AI as an assistive tool. G2 generates substantial components. G3 is AI-native according to the creative-agency test. G4 introduces real-time, audience- responsive generation. G5 describes a persistent world with durable identity, multi-user state, social rules and an economy. The classification focuses on the ontology of the work: whether the output is a file, a generated work, an interactive experience or an operating world.

3.2 The 2026 position: visible capability, incomplete institutions

By 2026, strong evidence exists for G2 and G3 production workflows, while G4 has entered public product exploration through interactive world models and low-latency generation [R07–R10]. G5 remains a scenario horizon. Its individual components—identity, virtual economies, agents, spatial interfaces and persistent state—exist separately, but their combination into an open, durable and governable cultural world has not reached general maturity.

3.3 Stage gates and exit conditions

G2 to G3 requires evidence of integrated human–AI authorship rather than isolated generated shots. G3 to G4 requires runtime response, state persistence, bounded latency, safety and fallback. G4 to G5 requires multi-user consistency, identity, economic rules, dispute handling and long-term stewardship. A product should be described at the highest stage it can operate reliably, not the highest stage imagined in its roadmap. *TABLE 3 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 4 · Four Evidence Grades and Permitted Language

4.1 Four grades and four permitted languages

E1 supports language such as implemented, available or required—within a stated jurisdiction or product boundary. E2 supports demonstrated or reported, not universally mature. E3 supports research indicates a possible capability, not a commercial promise. E4 must be written conditionally: if defined assumptions hold, a future configuration becomes plausible. Company-reported performance remains E2 unless it has meaningful independent validation.

4.2 Technology scanning and forecasting

The research process separates capability, reliability, affordability, deployability and legitimacy. A benchmark improvement is a capability signal; it is not proof of production reliability. A public demo is a deployability signal; it is not proof of sustainable unit economics. A regulation is a legitimacy constraint; it may also create infrastructure demand. Forecasts are built from the interaction of these variables rather than a single exponential performance curve.

4.3 Annual revision and falsification

The white paper should maintain a prediction register with assumptions, leading indicators and disconfirming evidence. If a forecast fails, the edition should record why. This discipline turns publication into a cumulative public research programme and prevents institutional credibility from depending on certainty theatre. *TABLE 4 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 5 · The Video Foundation-Model Capability Stack

5.1 From pixel synthesis to audiovisual foundation models

Diffusion and transformer-based systems have improved motion, composition, prompt adherence and duration, while native audio-video approaches move sound from post-production attachment toward a shared temporal representation [R02–R06]. The technical objective is no longer a single coherent shot. It is a controllable representation of identity, action, camera, sound, timing and physical context that can be repeatedly edited without collapsing unrelated elements.

5.2 Controllability as the industrial divide

Professional workflows need selective change: alter a gesture without changing the face; extend a shot without shifting wardrobe; replace dialogue while preserving performance timing; move a camera while keeping geography stable. Control should therefore be measured as coverage across creative variables and as the size of the unintended blast radius after revision. A model that requires complete regeneration for every change transfers cost from shooting to iteration.

5.3 Native audio-video and a common timeline

Joint generation can improve lip synchronisation, environmental timing and the coupling of rhythm, speech and action. It also expands the rights and safety surface. A production record must identify which voices, music, effects and performance signals were generated, adapted or sourced, and how they were approved. Shared temporal generation should therefore be accompanied by shared temporal provenance. *TABLE 5 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 6 · The Feature-Length World-State Loop

6.1 A feature is a state database

Every scene depends on facts established earlier: who knows what, which object moved, how a relationship changed and what physical constraints still apply. These facts should be represented explicitly rather than left inside an opaque context window. A production-grade state system links the story bible, character state, spatial map, asset versions, approved shots and unresolved continuity risks. Generated output is then checked against the state before it becomes canonical.

6.2 Hierarchical planning: world, sequence, scene, shot and frame

Long-form control requires multiple planning horizons. World rules constrain the whole project; sequences carry dramatic movement; scenes update relationships and information; shots specify action, camera and sound; frames render the local result. Evaluation and regeneration should occur at the smallest level that can correct the problem without destabilising higher-level intent. This hierarchy also enables different models and human specialists to work on separate layers.

6.3 Metrics beyond visual quality

At shot level, metrics include action completion and audio-video synchronisation. At scene level, identity, wardrobe, prop and location continuity matter. At sequence level, causal and spatial coherence dominate. At whole-work level, emotional development, theme and pacing must remain intelligible. The most useful metric is often the cost and time required to restore the work after a detected inconsistency. *TABLE 6 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 7 · Minimum Interactive World-Model Loop

7.1 The minimum interactive loop

The smallest useful loop observes input, interprets intent, selects an action, updates state, generates a response and writes the resulting event into memory. If the system cannot distinguish observation from authoritative state, or cannot explain which action produced a change, it is a generative interface rather than an operational world model. For cinema, this distinction determines whether an audience action has durable narrative meaning.

7.2 The 2026 capability boundary

Public systems have demonstrated real-time interaction at 720p and 20–24 fps over minute-scale sessions, while research explores streaming generation and longer memory [R07–R10]. These are significant E2 and E3 signals. They do not yet establish open-ended multi-hour stability, unrestricted action adherence, broad- device affordability, multi-user synchronisation or event-level governance. Product claims should preserve that boundary.

7.3 Collaboration with game-engine infrastructure

A practical near-term architecture combines explicit geometry, physics, navigation and deterministic rules with generative rendering, dialogue, performance and variation. The engine controls what must be reproducible; the model supplies what benefits from probability and expressive breadth. Hybrid design also makes safety and debugging more tractable because critical state can remain inspectable even when the surface is generated. *TABLE 7 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 8 · Eight-Layer Real-Time AI Cinema Architecture

8.1 The eight-layer stack

Sensing converts voice, gaze, gesture, controller input and world events into bounded signals. Intent interpretation separates a user request from incidental observation. Narrative policy constrains what may happen; state and memory preserve what has happened. Model routing assigns specialised models; generation and rendering produce the audiovisual result; character and voice control maintain performance; delivery coordinates device, edge and cloud; provenance and safety record the event and enforce release policy.

8.2 Latency budgets and graceful degradation

Not every operation belongs on the same deadline. Perceptual feedback may require very low latency, while a complex narrative response can use staged output. Systems should cache safe assets, predict likely actions, use lighter local models, stream progressive quality and fall back to approved pre-generated material. When policy checks or model calls fail, the experience should become simpler—not unsafe, incoherent or unavailable without explanation.

8.3 Dynamic provenance

File-level provenance is insufficient for a continuously generated session. The event record should identify the submitted intent, active policy, model and version, referenced assets, generated result, safety decision, operator and timestamp. Public disclosure can remain a concise summary while sensitive logs stay access- controlled. The goal is not surveillance of the audience; it is accountability for consequential system behaviour. *TABLE 8 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 9 · Six Components of an Operational Character

9.1 An agent is more than a character surface

A production-ready agent requires a stable identity model, an explicit set of goals, bounded memory, permitted capabilities, behavioural policy and a performance layer. These components should be separately versioned. A new voice model should not silently alter character ethics; an expanded memory should not automatically expand consent; and a change in action capability should trigger review even when appearance remains unchanged.

9.2 Performance control and digital replicas

Face, voice, motion, personality and autonomous behaviour are distinct permission domains. Consent should specify media, context, audience, territory, duration, remuneration, withdrawal and high-risk exclusions. Converting a passive likeness into an autonomous agent is a material change, not routine reuse. The system must also distinguish a fictional character owned by a production from the protected identity of a performer [R15–R16].

9.3 Multi-agent production and operations

Agents can support breakdown, continuity review, asset search, localisation, policy checking and world operations. Their work should be organised as proposals that enter a deterministic approval pipeline. Important decisions—canonical story changes, high-value transactions, sensitive representations and public release—require an identifiable approving authority. Agent collaboration is useful only when responsibility does not dissolve across a chain of automated actions. *TABLE 9 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 10 · Four Audiovisual Device Migrations

10.1 From collective sessions to contextual events

The theatrical unit is the scheduled session; the television unit is the programme; the smartphone unit is the feed, short video or live stream. AI glasses introduce a possible unit based on place, relationship and event. A scene can begin because a person enters a location, meets another participant or asks a character to reveal a layer of the world. This shifts commissioning from minutes of content toward rules for when and why an experience appears.

10.2 Device–edge–cloud collaboration

Local processing is essential for sensitive sensing, immediate feedback and basic safety. Edge infrastructure can reduce latency and support regional state, while cloud systems provide large-model inference and persistent world services. A well-designed product decides which information never leaves the device, which events require shared state and which outputs can be generated from anonymised or minimised signals.

10.3 The real adoption gates

Weight, heat, battery life, field of view, display legibility, prescription support, price and style all matter. So do visible recording status, bystander expectations, social etiquette and the user's ability to silence the system. Public prototypes and platform work show momentum [R24–R27], but the market will scale only when ordinary people can wear the device without technical, social or psychological friction. *TABLE 10 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 11 · The Contextual Content Loop for AI Glasses

11.1 Location narrative and environmental characters

A street, museum, theatre district or transit journey can become an active story space. Characters may understand broad context, refer to persistent world events and reveal content that depends on location or time. The strongest experiences will not cover reality with constant graphics. They will use sparse intervention, legible spatial anchors and clear transitions between the factual environment and the fictional layer.

11.2 Companion characters and attention design

A companion can support continuity across sessions, but persistent relationship design can also create dependence or manipulate vulnerability. Products need intervention limits, quiet modes, memory controls, age-appropriate defaults and clear commercial disclosure. Emotional engagement should not be used to obtain consent, raise prices or bypass a user's stated attention boundaries.

11.3 A new shot grammar

The frame is no longer fully controlled by a director because the user can look elsewhere and move through space. Composition therefore shifts toward fields of relevance, acoustic guidance, timed reveals, spatial blocking and event priority. Creators design what can be noticed, how attention is invited and what happens if a cue is missed. The result is closer to choreographing perception than editing a fixed sequence. *TABLE 11 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 12 · Hybrid Spatial Content Pipeline

12.1 Three routes to spatial content

Pre-produced 3D offers deterministic quality but high asset cost. Reconstruction and neural representations can capture real environments quickly but complicate editing, identity and rights. World generation offers variation and responsiveness but must manage consistency and safety. Professional products are likely to combine the three: deterministic hero assets, reconstructed real locations and generated variation within governed boundaries.

12.2 Remote generation and foveated streaming

Rendering can be split across device, edge and cloud according to latency, privacy and quality. Foveated techniques allocate quality where the user is looking, reducing transport and rendering demand [R26]. Yet gaze is a sensitive signal. Systems should use it for delivery efficiency without silently converting it into behavioural profiling, purchase intent or emotional inference.

12.3 Interoperability and asset portability

OpenXR and related spatial standards provide common interfaces for devices and entities [R27], but narrative portability also requires semantic and behavioural conventions. A door must be recognisable as an object with rules, not only a mesh. A character needs identity and rights metadata, not only animation. Interoperability should therefore cover state, events, permission and provenance as well as geometry. *TABLE 12 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 13 · The Audience Rights Ladder

13.1 Four levels of audience rights

Watching includes starting, pausing, stopping and accessing the work. Choosing means navigating declared branches. Acting means changing persistent state through speech, movement or other input. Co-governance allows participants to create assets, propose rules or share in value. Products should state which level they offer and which decisions remain under editorial or operational control.

13.2 Default viewing and choice fatigue

Cinema can remain immersive because viewers are not required to manage every moment. AI-native design should preserve a strong default experience and invite action only when it adds meaning. The system can infer that a person is disengaged, but it should not convert passive signals into irreversible story changes or commercial decisions. A user must be able to return to a coherent experience after doing nothing.

13.3 Rebuilding collective viewing

Shared experiences require rules for conflicting intentions. A group may vote, delegate, take turns or follow roles. The design should disclose how influence is allocated and protect minorities from harassment or systematic exclusion. Collective agency is not merely a larger input channel; it is a small governance system inside the work. *TABLE 13 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 14 · Multimodal Interaction and Authorisation

14.1 A multimodal interaction map

Voice is effective for explicit intent but should not be retained by default. Gaze is useful for reference and rendering priority but not for consent. Gesture supports spatial manipulation but high-risk gestures need confirmation. Location enables context but should not expose a person's movements. Physiological data should be limited to necessary accessibility or safety functions unless a separate, revocable purpose is clearly accepted.

14.2 Natural-language direction and constraint language

Users may ask a world to become quieter, reveal another perspective or let a character explain a past event. The system should translate open language into a bounded action plan governed by story, safety and rights. It should also explain when a request conflicts with the work's rules. Natural language is a convenient interface; it is not an unlimited instruction channel.

14.3 Feedback, latency and trust

An interactive system should acknowledge input quickly, distinguish processing from completion and show when an action is unavailable. If latency is high, it can provide a safe preliminary response rather than silently ignoring the user. Trust declines when the world changes without a visible cause, when the system pretends to understand or when an inferred intention cannot be corrected. *TABLE 14 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 15 · Value and Risk in Personalisation

15.1 Four objects of personalisation

Presentation includes subtitles, volume, language, pacing and accessibility. Content alters scenes, explanations or difficulty. Relationship adapts character memory and interaction across time. Value personalisation changes persuasion, ideology, price or the emotional pressure applied to a user. Defaults may be appropriate at the presentation layer, while deeper layers require explanation, consent and accessible controls.

15.2 Personal memory and data minimisation

Long-term character memory can make a world meaningful, but the system should store only what serves a declared purpose. Users need to inspect, correct, delete and export important memories. Session memory, cross-work memory and commercial profiling should be separate. A fictional character should not retain sensitive real-world information merely because it makes dialogue more convincing.

15.3 Preventing emotional price discrimination and hidden persuasion

A system that infers urgency, loneliness or susceptibility should not use that inference to raise a price, increase advertising pressure or make withdrawal harder. Personalisation models must not become invisible negotiation agents acting against the user. High-risk adaptation requires policy, logging, independent review and a non-personalised path. *TABLE 15 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 16 · From Story Constitution to Each Experience

16.1 From script to story constitution

A story constitution preserves what must remain true while allowing the experience to vary. It defines theme, canon, character commitments, causal boundaries, sensitive content rules and the range of valid endings. Generation occurs inside this authored space. The constitution does not replace scenes or dialogue; it provides a higher-order structure that keeps thousands of possible scenes recognisably part of the same work.

16.2 New artistic units

The unit of composition may become a rule, a relationship, an encounter, a memory or a designed transition between world states. Rhythm can emerge across sessions rather than minutes. A character performance may be evaluated through consistency under unplanned input. Critics and juries therefore need methods that consider the realised experience and the system of possibility behind it.

16.3 Preserving works that cannot be exactly replayed

Preservation should capture the released model and version, asset set, story constitution, policy, important states, representative sessions and documentation of the runtime environment. A recording alone shows one performance but not the work's possibility space. Preservation institutions will need emulation, controlled replay and event archives alongside conventional masters. *TABLE 16 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 17 · One World, Multiple Media

17.1 From shared IP to shared state

Traditional adaptations share names, imagery and lore while rebuilding production assets and logic. A shared-world approach defines identity, geography, chronology, abilities, relationships and provenance in machine-readable layers. Film can interpret a character dramatically, a game can execute abilities, and live media can respond to events while remaining connected to a common version history. Reuse becomes semantic and operational, not merely visual.

17.2 The world studio

A world studio combines a story and art department with state architecture, asset systems, continuity engineering, live operations and governance. It distinguishes canonical state from temporary session state and community-created branches. The operating model resembles both a studio and a long-lived service organisation: it releases works, maintains the world, resolves conflicts and plans how value returns to contributors.

17.3 Convergence risks

A shared world can become creatively rigid if every medium must obey a central database. It can also produce surveillance, exploitative labour, inaccessible canon or indefinite dependency on one platform. Governance should preserve medium-specific interpretation, forkable versions, contributor attribution and clear limits on what a commercial operator may change retroactively. *TABLE 17 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 18 · Generative Live-Operations Loop

18.1 Three real-time product families

The first family is a scheduled generated programme with controlled participation. The second is an event layer that responds to sport, news, performance or community activity. The third is a persistent world that continues between major events. Each requires different latency, moderation and continuity policies. A scheduled show may use extensive pre-approved fallback; a persistent world must manage durable consequences and participant identity.

18.2 Audience input and governance latency

Safety review cannot always occur after generation because harmful output can be experienced immediately. The system should classify input, constrain the action space, select approved models and monitor the generated result. Governance latency is part of the end-to-end budget. When confidence is low, the world should reduce autonomy, delay publication or switch to a known-safe response.

18.3 Commercial and operational metrics

Key metrics include end-to-end latency, fallback rate, unexplained interruption, state-recovery time, incident frequency, cost per session minute, participant retention and world-state integrity. Growth can hide structural losses if every additional participant increases generation cost faster than revenue. Real-time products need service-level economics, not only audience metrics. *TABLE 18 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 19 · The New AI Cinema Value Chain

19.1 Six layers of the emerging value chain

Foundation infrastructure provides models, data and compute. World and workflow systems organise state and tools. Production creates works and reusable assets. Distribution connects audiences, devices and contexts. Transaction systems manage permission and value flow. Operations maintain identity, safety, continuity and community. Organisations may occupy several layers, but conflicts of interest and responsibility should remain visible.

19.2 New professional combinations

An AI-native producer combines creative, budget and system responsibility. A world architect models rules and state. A continuity engineer manages identity and causality across generations. A provenance engineer designs records and metadata. A generative safety director defines refusal, degradation and recovery. These roles are not fixed job titles; they identify capability gaps that conventional departmental boundaries often leave unowned.

19.3 Labour transition and fair distribution

Automation can remove repetitive work while concentrating leverage in models and platforms. Transition programmes should track income, bargaining power, credit, reskilling and access to tools—not only the percentage of automated tasks. Contributors need clear attribution and remuneration when assets, performances and workflows continue to create value across products and years. *TABLE 19 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 20 · Persistent-World Revenue Portfolio

20.1 Expanding the revenue unit

A world subscription sells maintained access and state. A character subscription sells continuity of relationship. An interactive event sells a time-bound experience. Enterprise licensing sells controlled brand or training environments. Asset services sell identity, workflow or world components. Agent requests support machine-initiated licensing or generation. Each unit requires different consent, service levels, accounting and refund logic.

20.2 From production budget to service unit economics

Traditional finance concentrates cost before release. Persistent worlds combine production expenditure with inference, storage, moderation, safety, customer support and state migration. The financial model should classify fixed world investment, variable session cost, acquisition cost, retention, liability reserves and end-of- life obligations. A product is not sustainable if success increases losses or makes withdrawal technically impossible.

20.3 Pricing and value distribution

Pricing should correspond to legible value rather than inferred emotional vulnerability. Revenue-sharing rules should address the continued use of performers, character creators, world designers, community contributors and workflow developers. Machine-readable contracts can improve settlement, but exceptions and disputes still need human process. The system should show which participant receives value when an asset is reused. *TABLE 20 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 21 · A Multi-Polar Global AI Cinema Network

21.1 Competition is shaped by different endowments

Model capability is only one advantage. Access to production talent, local stories, game and animation expertise, data governance, capital, distribution, language markets and device supply chains all influence the field. A region can lead in world assets or governance even without training the largest model. The market will reward institutions that translate between these endowments.

21.2 Hong Kong's structural position

Hong Kong can contribute as a multilingual, cross-border market and standards connector with deep film culture, financial services and international networks. Its opportunity is not to imitate a hyperscale model centre. It is to make assets, projects, investment, festivals, evaluation and distribution interoperable across jurisdictions, while preserving a credible distinction between industry infrastructure and promotional claims.

21.3 A minimum language for international collaboration

Cross-border work needs shared definitions of work type, AI role, runtime behaviour, asset identity, permission scope, provenance status and responsibility. These fields should support local legal differences rather than pretending to replace them. A minimum common language allows creators and institutions to compare projects without forcing one jurisdiction's legal categories onto all participants. *TABLE 21 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 22 · The AI Audiovisual Asset Family

22.1 Six asset classes

A work is the released audiovisual experience. A workflow records tools, nodes and decisions. A character combines identity, model, performance and permission. A world contains semantics, rules and state. A behaviour defines an executable capability or interaction. A record provides provenance, event and lifecycle evidence. The same project may contain hundreds of related assets, each with different access and commercial terms.

22.2 Three conditions for usable assets

An asset must be identifiable: users and machines can distinguish it and its version. It must be interpretable: key technical, semantic and governance fields are readable. It must be usable within a known permission envelope. A visually impressive asset that lacks provenance, version clarity or permission scope may be valuable as inspiration but unusable in professional collaboration.

22.3 Evaluating asset value

Value may depend on audience performance, reliability, editability, relationship strength, reuse, provenance completeness and licence quality. One metric cannot compare all asset classes. A character can create durable value through identity stability; a workflow through acceptance rate; a world through extensibility; and a record through its ability to reduce transaction and dispute cost. *TABLE 22 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 23 · The Asset Lifecycle

23.1 From creation to retirement

Creation produces an asset and initial evidence. Declaration identifies the responsible submitter. Identification assigns a persistent ID and parent relationships. Verification checks statements or signatures. Evaluation adds independent observations. Discovery exposes permitted metadata. Permission enables use; operation creates events; settlement records value flow. Change, archive and retirement preserve history and determine what remains accessible.

23.2 The minimum verifiable record

The record should include asset ID, type, version, parent asset, responsible contact, provenance summary, model and tool declaration, rights summary, governance status, technical format, integrity reference and dispute or withdrawal state. Not every field should be public. The minimum public layer should allow identification and informed contact without exposing confidential prompts, personal information or security- sensitive logs.

23.3 Withdrawal, disputes and orphan assets

A rights holder may withdraw permission; a submitter may disappear; an asset may be disputed or technically obsolete. The system needs states for suspension, contested status, replacement, restricted access and orphaned responsibility. Historical records should not imply current permission. Machine clients must be able to recognise that an asset once existed but cannot presently be used. *TABLE 23 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 24 · Layers of Agent-Mediated Asset Transactions

24.1 The machine-readable rights envelope

An asset can express permitted uses, excluded uses, territory, duration, identity scope, price logic, attribution, technical constraints and whether human approval is required. These fields support discovery and preliminary negotiation. They do not remove exceptions, moral rights, collective agreements or jurisdictional rules. The envelope should state when machine interpretation must stop and a human decision must begin.

24.2 Separating probabilistic intent from deterministic settlement

A language model may infer that a user wants a licence, but inference is not authority. A transaction layer should present the exact asset, rights, price, counterparty and revocation conditions, obtain an approved authorisation and generate a durable record. Emerging agent-payment protocols are important infrastructure signals, while responsibility and consumer protection remain open design questions [R21–R22].

24.3 From a search market to a request market

Human users browse catalogues; agents can express structured requirements and ask the market to assemble a compliant solution. This can reduce transaction cost for localisation, asset replacement, music, voices and world components. It can also intensify concentration if one platform controls discovery, identity and payment. Open request formats and exportable records are therefore strategic infrastructure. *TABLE 24 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 25 · Human–AI Dual-Domain Value Flow

25.1 Why ‘AI-world GDP’ is not a direct additive total

An agent may call several models, purchase internal services and transfer a virtual asset before one final user pays for an experience. Counting every step as new value would count intermediate activity several times. Asset prices can also rise without corresponding current production. Measurement must distinguish gross transactions, revenue, income, capital formation, final use and value added, and must identify the economic principal behind automated activity.

25.2 Human–AI Dual-Domain Value Accounts

The human domain records creative labour, data preparation, capital, energy, rights and institutional services. The AI execution domain records inference and agent services attributed to a commissioning principal. The virtual-world domain records final use, maintained assets and internal circulation. The real- return domain records observable revenue, wages, licences and tax. Reconciliation removes intermediate transfers and shows how value and risk move between domains.

25.3 An AI-native audiovisual satellite account

A satellite account can estimate industry output, compensation, intermediate inputs, asset formation, imports, exports and energy use while preserving the main national accounts. Firms could begin with a voluntary sample: classify project expenditure, cloud operations, reusable world assets, human compensation and final revenue. The objective is a comparable evidence base, not a headline number detached from accounting boundaries. *TABLE 25 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 26 · Four Rights Layers in AI Cinema

26.1 Authorship: tool use and creative control

The U.S. Copyright Office and other authorities emphasise human authorship and control when assessing copyrightability [R15]. In AI-native work, evidence may include the creator's story constitution, selection, arrangement, revision and final approval. Merely requesting an unpredictable result is not equivalent to controlling expression. Different jurisdictions may reach different conclusions, so project records should support factual analysis without assuming a universal legal answer.

26.2 Separating training, project input and output

Training permission concerns the development of a model. Project-input permission concerns the assets supplied for a specific work. Output questions concern similarity, protected elements, contractual terms and human contribution. These layers should be reviewed independently [R16]. A production cannot cure an uncertain training or replica issue merely by placing a label on the final file.

26.3 Digital replicas and continuing consent

Consent for a face, voice or performance should specify media, context, territory, duration, autonomy, remuneration and withdrawal. A material change—such as allowing a replica to speak unscripted, enter a sensitive scene or operate after a performer is no longer available—should trigger renewed consent. The system must also maintain a practical means to stop new use and preserve the record of prior authorised releases. *TABLE 26 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 27 · Three Transparency Layers

27.1 A three-layer transparency architecture

The provenance layer records assets, tools, responsible parties and integrity references. The machine- readable layer communicates generation and modification status to platforms and agents. The perceptible layer informs users in a context-appropriate form. China’s labelling measures and EU transparency work both indicate movement toward coordinated explicit and implicit disclosure [R12–R14]. Implementation should remain accessible without overwhelming the audience.

27.2 A responsibility chain rather than a responsibility cloud

A model provider controls capability, safeguards and version disclosure. A creator controls inputs and creative choices. A producer controls workflow, rights and approval. A distributor controls labels, audience and platform policy. A runtime operator controls live policy, incident response and rollback. Several parties may share responsibility, but each should preserve the records that correspond to its control.

27.3 The institutional position of Index.Film™

Index.Film™ supports identification, recording, search and collaboration across work, workflow, character, world, behaviour and record assets. It should publish its metadata rules, correction process, access layers and dispute states. Inclusion is not endorsement; indexing does not confirm title; and an identifier does not replace legal registration or adjudication. These limitations are a source of credibility, not weakness. *TABLE 27 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 28 · Four Dimensions of Public Value

28.1 Cultural diversity and data concentration

Large models can reproduce dominant languages, aesthetics and commercial assumptions. Diversity requires more than translating an interface. Institutions should examine whose archives, performance traditions and narrative structures appear in data and evaluation; support local creators and languages; and enable communities to define sensitive uses. UNESCO's work on AI and culture underscores the need to connect innovation with cultural rights and plural participation [R18].

28.2 Human agency and the creator economy

Creators need meaningful control over tools, credit, income, portability and refusal. Audiences need the ability to understand, correct and exit adaptive systems. A platform can increase nominal participation while reducing bargaining power or making creators dependent on opaque ranking and generation services. Human agency is therefore an economic and institutional metric, not only an ethical aspiration.

28.3 Energy and sustainable generation

Evaluation should report energy per accepted output, per user session and per active world, while total energy and regional grid effects remain visible. Model efficiency, caching, edge placement, workload scheduling, renewable supply and selective generation can reduce demand. The industry should avoid the assumption that generated abundance has no physical infrastructure cost [R19–R20]. *TABLE 28 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 29 · Three Observation Windows, 2026–2035

29.1 2026–2028: the infrastructure period

The priority is controllable generation, long-form state systems, native audio-video workflows, provenance, labelling and professional evaluation. AI-native films become more common, while real-time products remain bounded. Organisations that build portable assets, rights records, model-routing and acceptable-output metrics create advantages that survive rapid model turnover.

29.2 2029–2032: the form-expansion period

If latency, cost and device gates improve, interactive cinema, generated live events, shared film–game worlds and spatial experiences can move from demonstrations to repeatable products. New subscription and event revenue may emerge. The critical governance shift is from file-level records to event-level responsibility and from content permission to agent and world permission.

29.3 2033–2035: the persistent-world period

If identity, multi-user state, portability, economics and governance mature together, persistent cultural worlds may become a significant medium. They will not replace films or games. They will provide a layer in which works, characters and communities continue across devices and forms. Cross-border disputes, statistical treatment and long-term stewardship will become central institutions. *TABLE 29 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 30 · Four Future Scenarios

30.1 Governed growth

Capability and revenue improve while standards, labelling and responsibility mature. Professional AI-native production expands, and audiences accept adaptive experiences because control is legible. The winning strategy is investment in trusted assets, evaluation and interoperable governance rather than maximal autonomy.

30.2 Platform enclosure

A small number of integrated platforms control models, identity, distribution, payment and world state. Creative tools are powerful but export is costly and terms can change. Organisations need multi-platform architecture, contractual exit rights, independent records and assets that can be migrated without losing identity or permission history.

30.3 Trust crisis

High-profile harm, rights disputes, synthetic deception or safety incidents reduce audience and institutional confidence. Regulation tightens and insurance costs rise. Provenance, explicit human responsibility, controlled product scope and rehearsed incident response become conditions of market access rather than optional compliance.

30.4 Open worlds

Open standards, agent transactions and portable identity allow creators and communities to operate across multiple worlds. Innovation and plural participation increase, but so do cross-system abuse, fraud and dispute complexity. Opportunity depends on strong identity permissions, contestable governance, interoperable evidence and fair settlement. *TABLE 30 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 31 · Action Network for Six Stakeholder Groups

31.1 Creators and production organisations

Map story, character, world, workflow and evidence assets. Define AI-native creative intent before selecting tools. Establish state and version systems, acceptable-output metrics, rights review and final human approval. Run one bounded project that can be evaluated honestly rather than distributing experimentation across every production without learning.

31.2 Model, technology and platform providers

Disclose capability and material limitations; support export, version history and provenance; separate probabilistic proposals from transactions; provide safety degradation and incident tools; and publish changes that affect rights or output. Platforms should make responsible operation easier than opaque workarounds.

31.3 Distribution, investment and insurance

Use AI delivery schedules that identify assets, models, permissions, labels, evidence and runtime responsibility. Evaluate acceptable-output economics and persistent-world liabilities. Price risk according to control and records rather than treating all AI involvement as identical. Require a practical exit and archive plan.

31.4 Policy, research and education

Develop shared terminology, representative test sets, public-interest research and statistical pilots. Education should combine cinema, interaction, systems, data, rights and operations. Certification should recognise demonstrated capability and professional responsibility rather than short-term familiarity with one interface. *TABLE 31 · CHAPTER ANALYTICAL FRAMEWORK*

FIGURE 32 · Joint Public-Interest Programme

32.1 Definitions and evaluation as public infrastructure

The P/R model, G0–G5 stages and E1–E4 evidence language should be revised through cases and peer review. Benchmarks should test long-form continuity, control, real-time recovery, provenance and human agency while leaving artistic quality open to criticism, juries and audiences.

32.2 Index.Film™ and market collaboration

The initial programme should publish minimum fields for six asset types, access levels, correction and dispute processes, and machine-readable permission boundaries. Market pilots can test how identifiers and records reduce transaction cost without claiming to determine ownership or artistic value.

32.3 A global city and talent network

Cities can host joint production, testing, education, festivals, residencies and market access. Participation should be based on concrete projects and reciprocal benefit rather than nominal membership. The network should make local creative cultures visible while improving cross-border interoperability.

32.4 Research and annual publication

Each annual edition should update technical evidence, regulatory developments, market data and scenario signals. It should report errors and changed judgments. A standing research process can invite institutions, creators, performers, engineers and public-interest groups to test the framework and propose revisions. *TABLE 32 · CHAPTER ANALYTICAL FRAMEWORK*

Appendices The appendices convert the white paper's concepts into practical terms, evidence requirements, metadata, revision protocols and action tools.