Appendix K. Glossary, Decision Rules, and Quick Index#

This appendix is not an abridgement of the book. It is a router for the production floor. The definitions emphasize what a term does operationally here; the decision rules help you choose the right problem layer under time pressure; the reading paths get each role back to the relevant chapters. Where a definition conflicts with a project contract, the approved contract wins.

K.1 Commercial and product terms#

Business model. How the work completes an exchange of value. Paid, ad-supported, branded and platform-commissioned each carry different cut point, advertising, acceptance and data logic. You cannot choose at release time.

Audience promise. The emotional, plot, relationship and visual experience the work continuously guarantees. It must be partly proven early and delivered across the season.

Desire stack. The several layers of satisfaction one audience pursues at once — public retaliation, proof of competence, being chosen in a relationship, restored standing. Not a demographic label.

Viewing context. When, where, on what device and with what attention someone watches. It shapes subtitles, sound, pacing and episode length.

First paywall. Where a user first meets a payment boundary. An effective one rests on free value already delivered, a concrete answer, and a credible payoff.

Debt. A question the work has not yet answered but has promised to advance. Debts have a borrowing, evidence, escalation, repayment and the possibility of going overdue.

Partial payoff. Genuinely completing an action, a piece of evidence or a relationship change before the big secret is revealed, to maintain trust.

Honest cut point. The old debt has been partly or wholly repaid, a new question is concretely present, and the cut withholds no basic understanding the audience should have received for free.

Resume contract. The state interface around an ad in an ad-supported route: what completes before it, how the scene is restored afterwards, and which sounds and questions must stay continuous.

Creative family. A group of acquisition creatives organized around one audience promise — not mechanical variants with a different subtitle color.

Landing continuity. Continuity of character, question, emotion and fact between the ad's promise, its end frame and the beginning of the episode.

Guardrail metric. A metric that prevents damage to trust, refunds, complaints, brand safety or long-term retention while a primary metric is optimized.

Scaling ladder. The reversible progression from a small share of traffic to full release, with primary metrics, guardrails and data quality rechecked at each step.

Stop conditions. The criteria for terminating an experiment or a project, written down before the spending starts, so sunk cost cannot drive endless generation.

Greenlight. Approval of a batch with an explicit scope, budget, deadline and learning objective. It is not a guarantee that the season will succeed.

K.2 Story engineering terms#

Episode contract. An episode's production contract: entry state, opening question, objective, change, payoff, new debt, duration and obligation to the next episode.

Beat. One judgeable change in state, knowledge, power or debt. A beat is neither a shot nor a line.

Scene card. The scene's location, time, character objectives, strategies, escalation, turn, exit state and risk record.

Scene power. Who controls the agenda, the interpretation, the space, the evidence, the resources or the way out — not who holds the most senior title or speaks loudest.

Action dialogue. Language a speaker uses to pressure, trade, conceal, probe, refuse or seize the right to interpret.

Subtext. The gap between what a character says and the result they actually want. Not deliberately vague writing.

Visible evidence. The object from which an audience can judge a character's psychology and the plot's change — through behavior, props, space or facts on screen.

Fact lock. An approved objective fact of the project, with a version. Generation and editing may not rewrite it.

Character knowledge. What a character knows, believes or misunderstands at a specific time. Characters cannot use every fact the author knows.

Audience knowledge. What the work lets the audience know or infer at a specific time. It may run ahead of or behind the characters, but it has to be designed.

Knowledge break. A character reacting correctly, speaking or choosing on information they have not yet obtained.

Payoff ledger. The cross-episode record of when a question was borrowed, how it advanced, when it was delivered and whether it went overdue.

Next-episode first-shot obligation. The action and question the next episode must answer or advance within a stated time after a cut point.

Emotional engine. The repeatable structure of pressure, strategy, reversal, consequence and new debt. Not a genre label.

Middle collapse. The work repeats the same resistance and reaction while secrets, strategies, relationships and costs all stay the same.

Set-piece function. The choice, power and emotional work a famous sequence actually performs. In adaptation, preserve the function rather than copying the form line by line.

K.3 Asset and identity terms#

Stable ID. A machine identifier that survives changes of name, file and version. The display name changes; the ID does not.

Identity contract. A character's cross-shot stable face, silhouette, hair and recognition anchors, plus the permitted and forbidden variations.

Identity anchor. A small number of stable features verifiable in the picture — face length ratio, hair parting, a small mole. Presence is not an anchor.

Golden reference. The stress-tested, approved identity baseline. Not necessarily the prettiest single image.

Angle grid. The set of references for a character at frontal, profile, three-quarter, high and low angles and under occlusion.

Expression envelope. The range of expressions a character can carry while holding identity. Not nine unrelated faces.

Re-anchoring. State continues to be inherited while identity is periodically re-injected from the golden references, preventing drift in continuous derivation.

Identity drift. Face shape, stable features, age or medium style changing when they should not.

Feature migration. In a multi-person shot, a mole, accessory, garment or hairstyle being assigned to the wrong character.

Feature ownership. A machine-readable statement of which character each signature feature belongs to.

Base look. A character's stable combination of wardrobe, hair, makeup and accessories for a given phase.

State overlay. Rain, blood, dust or damage added on top of a base look by an event.

Location contract. A location's floor plan, landmarks, materials, entrances, camera zones, light sources, axis and sound constraints.

Camera zone. An area within a location where a camera may be placed while keeping the background and axis coherent.

Asset lifecycle. Executable states such as exploration, candidate, approved, locked, retired and quarantined.

Quarantined asset. An asset forbidden to new tasks for reasons of rights, safety, provenance or quality, while its evidence is retained.

Dependency graph. The directed relationships between assets, facts, shots, tasks, approvals, rights and releases, used for impact analysis.

K.4 Continuity terms#

State snapshot. The character, prop, location, time, knowledge and sound state at a shot or scene boundary.

State event. A business fact that explicitly changes state — a prop transfer, a change of clothes, an injury, or a secret learned.

Start-end interface. The state contract requiring the previous shot's end to match the next shot's start.

Action contract. An action's start, preparation, contact peak, end, direction and safe cut points.

Two-handed peak. The brief state in a handoff where both parties touch the prop, proving a transfer rather than a teleport.

Axis. The spatial direction formed by characters or an action; crossing it changes screen relationships.

Re-establishing. Rebuilding screen relationships through a visible camera move, a neutral shot or character movement.

Eyeline target. The object a character looks at in world coordinates — not simply screen left or right.

World coordinates. The shared spatial positions of characters, props and camera within a location, used to derive background and direction across shots.

Knowledge state. The set of facts a character currently knows, wrongly believes, suspects or does not know.

Wetness state. The overlay state by which wardrobe, hair and skin change progressively with exposure to weather.

Music state. A timeline snapshot of theme version, key, tempo, bar, stem switches and transition method.

Continuity lint. Rule checks run against deterministic state conflicts. Not responsible for judging whether a performance is moving.

Impact analysis. After a fact, asset or version changes, the list of what must be recompiled, redone, regressed or recorded as an exception.

K.5 Shot and generation terms#

Shot requirement. A shot's narrative job, framing, state, required visibility, prohibitions, acceptance, risk and insurance route.

Shot state packet. The minimum identity, state, spatial, action, sound and fact package needed to generate and evaluate the current shot.

Insurance coverage. Low-risk alternative shots designed before the high-risk principal shot succeeds — hands, reactions, prop inserts.

Key state frames. Approved images of an action's start, peak and end, used to constrain the path a video model interpolates.

Usable range. The range of frames in a take where identity, action, lip sync and background all hold simultaneously.

Handles. The extra frames preserved either side of a usable range so the editor can build action and sound interfaces.

Prompt compiler. The system that converts task, facts, state, references and vendor capability into a call format.

Prompt manifest. The record of compiler, template, adapter, input checksums, references, assumptions and the rendered prompt's checksum.

Reference selector. The component that chooses the minimum reference assets by angle, state, location, lifecycle and task purpose.

Vendor adapter. The layer converting unified semantics into a specific model's fields, handling unsupported capabilities explicitly.

Controlled attempt. Changing exactly one declared variable at a time, with a falsifiable prediction written before the call.

Attempt log. The record of hypothesis, change, frozen inputs, result, cost, conclusion and next step.

Cost per approved second. Total model, labor and rework cost divided by the finally approved usable seconds — much closer to production reality than a per-call price.

Capability routing. Selecting a validated vendor or method by shot type, quality, cost, latency, rights and fallback.

Silent de-escalation. The system ignoring an unsupported constraint and calling anyway. A production system blocks instead.

Cumulative derivation. Continuously using the last generated result as the new identity source, so error magnifies shot by shot.

K.6 Sound and post terms#

Voice bible. The contract for a character's vocal identity, register, tempo, articulation, breath, sentence endings, emotional boundaries and rights.

Voice anchor. Short test audio and parameters used to rebuild the same voice across batches and long narration.

Pronunciation lexicon. The unified pronunciation of names, companies, numbers, abbreviations and terminology.

Unit of thought. The passage in a long narration that completes one cognitive change. Better suited to performance and assembly than segmenting by word count.

Authority ledger. The division of which facts and feelings are carried by picture, dialogue, narration, music and graphics.

Music bible. A motif's narrative meaning, key, tempo, stems, permitted variants and prohibited uses.

Cue sheet. The timeline record of where music, dialogue, narration, ambience and Foley enter, change and exit.

Sound bridge. Using sound across a cut to connect action, space, time, or the two sides of an ad break.

Narrative mix priority. Whatever matters most at a given moment gets the acoustic space, and other sounds actively give way.

Intermediate timeline representation. A vendor- and application-neutral structure for shots, audio, subtitles, graphics and provenance.

Picture lock. The version at which picture durations and order freeze; later changes require impact analysis.

Clean version. A deliverable without burned-in subtitles or specific graphics, for localization and derivatives.

Change list. The structured difference between edit versions: additions, deletions, replacements, durations and downstream invalidations.

Shortest comprehensible version. The pressure version trimmed to the edge of comprehension, used to separate necessary information from habitual pauses.

K.7 QC, release and rights terms#

Finding. An observed phenomenon with evidence, from a checker or a reviewer. Not necessarily an independent problem yet.

Issue. A production problem, aggregated from findings, that can be assigned, fixed, regressed and closed.

Blocker. An error that changes a key fact, right, safety condition, identity or release integrity. It must be closed before release.

Major. An error that significantly damages comprehension, performance, space, sound or brand, but can be handled through an authorized process.

Minor. A local flaw not affecting comprehension or rights, still recorded along with whether it was accepted.

Regression set. The minimum sufficient set of objects that must be revalidated after a fix, derived from dependencies.

Closure evidence. The fix version, the comparison, the regression, the approver and the time, proving an issue is resolved.

Release candidate. An immutable candidate with content, sound, subtitles, graphics, QC, rights and technical profile frozen.

Release manifest. The authoritative list of every file, version, checksum, approval and rollback relationship in a release candidate.

Rights snapshot. All asset licences, uses, platforms, territories, terms and evidence states, frozen at release.

Use relationship. Rights are not a permanent property of an asset. They are a licence for a specific party to use it at a specific time, in a territory, on a platform, in a manner.

Rights revocation. The process of stopping new use, querying derivatives, handling published material and preserving evidence after a licence ends.

Two-person gate. A high-risk release verified independently by two different people, preventing a single-point mistake.

Release receipt. The platform content ID, upload time, remote version, transcoded preview and event verification.

Rollback. Stopping current distribution and restoring a known approved version, without deleting incident evidence.

K.8 Agentic and system terms#

Agent contract. An agent's objective, what it may do, must do and must not do, its uncertainty policy and its definition of done.

Task packet. The execution contract carrying objective, inputs, constraints, outputs, acceptance, budget, permissions, idempotency and retries.

Capability token. A minimal-permission credential bound to a subject, project, resources, actions, budget and expiry.

Context packet. A task's minimum context compiled by authority and provenance, plus what was excluded and what is assumed.

Candidate, recommendation, decision. Agents produce candidates and suggestions; approvers or an authorized process produce decisions. All three are recorded separately.

Event envelope. The unified message format carrying event type, version, time, producer, causation, correlation and idempotency.

Idempotency. Repeating one business intent produces no duplicate task, call, state or settlement.

Call intent. A record persisted before a request goes to an external vendor, so a crash can tell you whether the call happened.

Lease. A worker's short-lived possession of a task; on expiry it can be taken over, avoiding a permanent running state.

Circuit breaker. Pausing new calls when a vendor's failure rate is too high, protecting the queue, the budget and downstream.

Dead letter. A task message preserved after repeated or deterministic failure. Fix upstream and create a new task; never edit history.

Backpressure. Actively slowing upstream production when downstream or human gates lack capacity.

Materialized view. A query view built from events and authoritative state. Rebuildable, and never the sole source of truth.

Snapshot. A recoverable compression of system state at a point in time, verified alongside events.

RPO. The acceptable range of data loss; critical state and events usually require close to zero.

RTO. The target time to restore business state after a failure.

SLO. A team's measurable target for success rate, latency, error escape or approval waiting.

Shadow run. A new rule or evaluator that only records results without affecting the gate, used for calibration.

Staged rollout. A new component applied to a small share of tasks, then expanded or rolled back on the metrics.

K.9 Learning and scale terms#

Experiment registry. Recording the question, hypothesis, variables, metrics, sample and stop rules before anyone sees results.

Scope of applicability. The projects, shots, audiences, models and conditions under which a conclusion holds.

Negative learning. Reusable knowledge that a method does not work, when it does not, and why.

Fast feedback. Short-cycle fixes to subtitles, graphics, landing points and deterministic problems.

Slow feedback. Changes to character, structure, the first paywall and series promises, which need evidence across batches.

Scale gate. The evidence-based decision to expand from a pilot to the next batch, the first ten episodes or a full season.

Approved seconds. Finished seconds that have passed every necessary acceptance and are usable in the final cut.

Work-in-progress limit. The number of tasks a workstation may hold waiting or in progress, preventing queue growth.

Quality debt. The future obligation carried by approved exceptions, temporary overrides, missing tests and placeholder assets.

Release train. A compatible set of prompts, adapters, rules, evaluators and templates released together with a batch.

Maturity. The capability levels from accidental generation, through a reproducible pilot, controllable batches and production at scale, to a learning system.

K.10 Twenty on-the-spot decision rules#

1. When the story is not understood, do not start by fixing the image. Go back to the episode contract, the beats and how facts are presented.
2. When a character looks wrong, do not add more adjectives. Check references, angles, state contamination and re-anchoring.
3. When an action fails, do not default to sampling more. Write the start, peak and end, and build insurance coverage.
4. When features migrate in a group shot, do not demand equal detail for everyone. Separate slots, features and action ownership.
5. When space jumps, do not rely on a random transition. Return to the floor plan, the axis and world coordinates.
6. When text is wrong, do not regenerate the shot. Use the fact-bound post graphics layer.
7. When lip sync is right and the performance is false, do not keep adjusting sync. Go back to intent, breath and physical action.
8. When narration is boring, do not first look for a more emotional voice. Check whether the thought changes and whether authority is duplicated.
9. When music does not fit, do not just move the entry. Check the motif's meaning, its trigger and the stem state.
10. When the pacing is slow, do not speed up the whole episode. Find repeated information and waiting that is already over.
11. When it will not cut together, do not blame the editor. Check coverage, handles and state interfaces.
12. When costs run over, do not silently substitute a cheaper model. Compare cost per approved second, regression and narrative loss.
13. When client feedback is abstract, do not execute it immediately. Convert it into a timecode, an objective and observable conditions.
14. When metrics fall, do not tell the writers to strengthen the conflict. Locate the funnel position, the version and the time range first.
15. When automated scores conflict, do not take a majority vote. Look at thresholds, evidence and scope of applicability.
16. When a vendor times out, do not resend immediately. Query the call intent and the remote job.
17. When budget runs out, do not automatically lower quality. Preserve what exists and escalate the decision.
18. When rights are unclear, do not carry on while you wait. Quarantine the asset and stop derivation.
19. When the wrong version is published, do not delete first. Stop traffic, preserve evidence, roll back, then investigate.
20. When the pilot succeeds, do not approve the whole season. Expand batch by batch through the scale gate.

K.11 Reading paths by role#

Founders and client leads. Preface, Part I, Part II, Parts XII–XIII, Part XVIII, Appendices H and J. Focus on promises, budget, rights, data and scaling; you do not need every generation parameter first.

Producers. Introduction, Parts I–III, Parts X–XIII, Part XIV, Part XVI, Chapters 107 and 114 in Part XIX, Appendices A, B, G and H. Focus on maintaining batches, decisions, risk and handovers.

Writers. Parts II–III, the state and knowledge material in Part V, Chapters 67–68 in Part XIV, Part XVII, Chapters 99–104 in Part XVIII, Chapter 108 in Part XIX, Appendix B. Focus on writing story as a producible contract.

Directors and visual leads. Parts IV–VII, Chapters 69–70 in Part XIV, Chapters 77–78 in Part XV, Part XVII, Appendices C and E. Focus on identity, space, action and shot function.

Sound and editing. Parts VIII–IX, Chapters 71–72 in Part XIV, Chapters 79–80 in Part XV, Chapters 92, 94 and 97–98 in Part XVII, Chapter 111 in Part XIX, Appendices E, F and G.

QC and rights. Part X, Chapters 63–64 in Part XIII, Chapter 73 in Part XIV, Chapter 81 in Part XV, Chapters 87 and 89 in Part XVI, Chapter 112 in Part XIX, Appendices F, G and J.

Engineering and agentic systems. Part XI, Chapter 74 in Part XIV, Chapter 81 in Part XV, Chapters 83–90 in Part XVI, Chapter 113 in Part XIX, Appendices C, D, G, H and I.

K.12 Reading paths by incident#

A character's face changing: Chapters 18, 22, 26, 35–37, 77 and 109, plus Appendices C, E and F.

Long narration and music: Chapters 14, 40–46, 71, 79, 94 and 102, plus Appendices E and F.

Broken transitions and state: Chapters 23–25, 29, 33, 47, 53, 70, 78, 80 and 91–98.

Agentic duplication, overreach and recovery: Chapters 55–59, 74, 81, 83–90 and 113, plus Appendices D, F and H.

High clicks and low retention: Chapters 15, 60–62, 76, 82 and 99–101, plus Appendices F and J.

Copyright and client disputes: Chapters 63–66, 73, 89, 101 and 103–105, plus Appendices G and J.

K.13 Reading paths by project stage#

Starting a project: Preface, Introduction, Part I, Appendix A.

Story: Parts II and III, Appendix B.

Assets: Part IV, Appendix C.

Continuity: Part V, Appendices C and E.

Shots and generation: Parts VI and VII, Appendix E.

Sound and post: Parts VIII and IX.

Acceptance and release: Parts X and XIII, Appendices F, G and J.

Systematizing: Parts XI and XVI, Appendices D, H and I.

Distribution and learning: Parts XII and XVIII.

Training a team: Parts XV, XVII and XIX.

K.14 A closing principle for the index#

When a problem appears, first ask which layer it belongs to: story fact, state interface, asset identity, generation execution, sound and post, quality gate, rights and release, or data feedback. Then return to the earliest layer that can change the outcome, rather than patching the finished cut over and over.

The point of this vocabulary is not to let a team speak more jargon. It is to let different roles understand the same responsibility, evidence and completion condition under the same word. Once "pass," "lock," "release" and "fixed" have definite meanings, AI can become a governable production capability rather than a faster way to make a mess.