Appendix J. Master Acceptance Checklist#

This checklist is used at pilot greenlight, batch close-out, before release and at season close. Each item is marked pass, fail, not applicable or unknown. Unknown requires an owner and a deadline; not applicable requires a stated basis. A failed critical item can never be offset by a composite score.

J.1 Commercial and audience#

• The business model is explicit, and paid, ad-supported, branded or other routes do not mix the wrong metrics.
• The target audience is described by desire and viewing context, not only age and gender.
• The core promise has partial proof visible in episode one.
• Competitor research includes both successes and failures, with the original evidence preserved.
• Research observation, interpretation and decision stay separate.
• The pilot answers a specific unknown rather than "seeing how it goes."
• Budget, deadline, stop conditions and the decision maker are signed off.
• What acquisition creatives promise can be delivered within a reasonable viewing time.
• Commercial metrics include refunds, complaints or another trust guardrail.
• Scaling uses a reversible ladder.

J.2 Story and season#

• The protagonist's desire produces action.
• The antagonist has a sustainable method rather than just raising their voice.
• The series promise covers emotion, plot, relationship and visual.
• The season is derived backwards from its final choice.
• The A, B and C lines carry different functions.
• Releasing a secret changes a character's model rather than only adding information.
• Each act escalates strategy with visible consequences.
• The middle contains no pure stalling or repeated humiliation.
• The payoff ledger records what was borrowed, its progress and its delivery.
• The ending delivers both the external goal and the internal choice.

J.3 Episode pack#

• This episode's entry state matches the previous episode's exit.
• The opening question is specific and visible.
• The episode's change can be judged in one sentence.
• The episode repays at least one old promise.
• The new debt borrowed at the end is explicit.
• The next episode's first-shot obligation is locked.
• Every beat changes state, knowledge, power or debt.
• Every scene has an objective, strategy, escalation, turn and exit state.
• Fact, character knowledge and audience knowledge stay separate.
• Dialogue belongs to action rather than background explanation.
• Long narration is segmented by units of thought.
• Three timed reads fall inside the range.
• High-risk actions have insurance routes.
• Pack lint passes and several roles have approved it.

J.4 Characters and assets#

• Every core character has a stable ID.
• Identity anchors are observable features.
• Identity references cover frontal, profile, head-down, speaking, backlit and multi-person conditions.
• Age, fatigue, rain and injuries do not contaminate the identity layer.
• Characters are distinguishable by silhouette, hair, wardrobe and accessories.
• Wardrobe separates a base layer from overlay states.
• Injuries have a side, a degree, behavioral consequences and a recovery time.
• Locations have a floor plan, fixed landmarks and camera zones.
• Props have an owner, a holder, a position and a state.
• Key text assets are produced independently in post.
• Assets have provenance, checksums, rights, lifecycle and approval.
• Retired and quarantined assets cannot be selected by new tasks.

J.5 Continuity#

• Every shot's start state matches the previous shot's end state.
• Prop transfers are committed as events.
• Hands, injured sides, hair partings and jewelry are never mirrored.
• No character uses a fact they do not know.
• Relationship changes have evidence in action.
• The axis, screen direction and eyeline targets are explicit.
• Characters' world coordinates update as they move.
• Multi-person shots have slots and feature ownership.
• Light direction agrees with story time.
• Weather, wetness and environment state propagate by event.
• Music preserves theme, stems, tempo and bar position.
• Zero blockers in the state lint.

J.6 Shots and generation#

• Every shot has one primary narrative job.
• Shot requirements carry state, required visibility, prohibitions and acceptance.
• High-risk shots go to high-quality generation only after cheap validation.
• The static storyboard and the animatic with sound have been timed.
• References are chosen by angle and purpose rather than piled in indiscriminately.
• Prompts derive from structured facts and state.
• Unsupported capabilities block rather than being silently ignored.
• Each attempt changes only its controlled variable.
• Attempt logs contain hypothesis, result, cost and conclusion.
• Stop rules are enforced after repeated failures on the same invariant.
• Candidates stay quarantined after technical validation.
• Operator recommendation and final approval are separate.
• Approved takes have usable ranges and handles.
• Published material can be traced back to inputs, model and rights.

J.7 Sound#

• Core characters have a voice bible and anchors.
• The pronunciation lexicon covers names, companies, numbers and terminology.
• Line tasks carry the preceding and following lines, the target and the intent.
• Lip sync, body and voice are accepted separately.
• Voices stay stable across the units of a long narration.
• Narration and picture do not duplicate authority.
• Musical themes have narrative meaning and variant boundaries.
• The cue sheet records entries, exits, bars and stems.
• Ambience establishes one continuous space.
• Foley matches the real action state.
• Music ducks dynamically for dialogue.
• Phone speaker, headphones and low volume are all tested.
• Sound assets carry commercial rights.

J.8 Editing and finishing#

• V0 establishes causality first, without packaging hiding problems.
• The failed range of each selected take is excluded.
• A muted viewing conveys the main state changes.
• An audio-only listen follows the subject and the conflict.
• Reaction shots are used only when a judgment changes.
• Pauses serve thought and action rather than being distributed evenly.
• Transitions connect through causality, action or sound.
• Subtitles come from locked dialogue and break by meaning.
• Subtitles avoid faces, evidence and platform buttons.
• Phone, contract and chat graphics bind to facts.
• Grading normalizes technically before unifying style.
• Changes after picture lock go through impact analysis.
• Masters, clean versions, subtitles, stems and checksums are complete.
• Exports are re-imported and verified.

J.9 QC and release#

• Story, continuity, audiovisual, technical and rights are checked as five separate layers.
• Findings have a timecode, observation, expectation and evidence.
• Issues have severity, root cause, routing and regression scope.
• Duplicate findings from a common root cause are aggregated.
• Blockers are not offset by composite scores.
• Low-confidence model evaluations go to humans.
• Repairs happen at the earliest responsible layer.
• Closure has a fix version and regression evidence.
• The release candidate is immutable.
• The rights snapshot covers every published asset.
• Release files match the signed manifest's checksums.
• A second person independently verifies the release.
• Playback, sound, subtitles and events are verified at low volume.
• The rollback target and its trigger conditions are explicit.
• Platform receipts and remote content IDs are preserved.

J.10 The agentic system#

• Every agent has a responsibility, inputs, outputs and forbidden permissions.
• Tasks have an objective, context, budget, acceptance and an idempotency key.
• Capability tokens are short-lived and minimal.
• External content is untrusted by default.
• Context sections carry an authority level and provenance.
• Candidates, recommendations, validations and decisions are recorded separately.
• Model output cannot write canon or approval state directly.
• Duplicate events do not duplicate tasks, calls or settlements.
• Deterministic errors are never blindly retried.
• Budget exhaustion preserves evidence and stops.
• Vendor anomalies have a circuit breaker and a compatible fallback.
• Dead letters are fixed with new tasks, not by editing historical messages.
• The human queue has an SLO, priorities and backpressure.
• Prompts, adapters and evaluators can be staged and rolled back.
• Snapshot and event recovery have been rehearsed.
• Someone who did not build the system can take over from the runbook.

J.11 Data, distribution and feedback#

• Creatives are organized by promise family rather than as edit variants.
• Creatives show real proof and do not invent plot.
• The event dictionary is fixed before the experiment.
• Primary, guardrail and diagnostic metrics stay separate.
• Experiment variables and what is held constant are explicit.
• Early stopping rules are defined before anyone sees results.
• Clicks, comprehension, continuation, payment and refunds are analyzed together.
• Comments are deduplicated, clustered and bound to version and timecode.
• Comments become issues rather than editing the script directly.
• Fast and slow feedback are separated.
• Data-driven suggestions return through change requests.
• Failed experiments preserve the reason and their scope of applicability.

J.12 Legal, contracts and organization#

• The adaptation rights chain covers medium, territory, language and AI use.
• Real likeness and voice licences cover the specific scenes and derivatives.
• Music, fonts, stock and model terms carry versions.
• Rights records include term, platform and the revocation process.
• A rights revocation can list every derivative and release package.
• Client-supplied material has provenance and approval.
• Scope of work, revision rounds, acceptance and payment are explicit.
• Change requests show cost, schedule and rights impact.
• The boundary of source files and editable deliverables is explicit.
• Team roles, approvals and on-call responsibilities are explicit.
• Budget includes labor, failure, storage, rights and delivery.
• Cash flow and profit are managed separately.

J.13 Scale#

• Scaling proceeds by batch through a scale gate, never approving a whole season at once.
• Throughput is measured in approved seconds and approved episodes.
• Work in progress does not exceed the slowest gate's capacity.
• Cross-episode shared assets have their own freeze.
• Component versions are bound to production batches.
• Leads, key facts and the paywall boundary are 100 percent human-checked.
• Quality debt has an owner, a deadline and a stop threshold.
• Vendor switching and exit have been rehearsed.
• The system still runs when a key person is absent.
• Language and platform derivatives share one source of fact.
• Season close-out reclaims keys, processes deletions and verifies the archive.

J.14 Absolute vetoes#

Any one of these prohibits release: uncleared real likeness or voice; an error in a key amount, date or identity fact; a lead visibly replaced; a prop end state that changes the plot; a character knowing a core secret too early; a release manifest that disagrees with the files; an open security incident; no rollback capability; a published asset that cannot be traced to its source.

J.15 Sampling for a 100-item audit#

A large project need not check every ordinary item by hand each time, but the absolute vetoes, lead close-ups, core facts, the paywall interface and rights are always checked in full. The rest is sampled by asset risk, recent defects, component changes and batch. A failed sample widens the scope; it is never averaged away.

J.16 Project sign-off#

final_acceptance:
  project_id:
  release_id:
  checklist_version:
  passed_count:
  failed_count:
  unknown_count:
  not_applicable_with_reasons: []
  blockers: []
  approved_exceptions: []
  signatures:
    producer:
    story:
    visual_continuity:
    sound_edit:
    qc:
    rights:
    release:

Signing is not a transfer of responsibility. It confirms what evidence each person holds. An overall pass never overrides an absolute veto.