Appendix H. 30-Day Implementation Route and Maturity Audit#

This appendix is for a team that has finished the book and is ready to build a real production system. The 30-day goal is not automatic generation of a whole season. It is one pilot episode with stable IDs, an episode pack, assets and continuity, controlled generation, sound and editing, QC, rights, release and retrospective evidence.

H.1 Day 0: define what success means#

implementation_goal:
  pilot: one_episode_60_120s
  team_size: 3_8
  must_prove:
    - narrative_clarity
    - identity_and_state_continuity
    - traceable_generation
    - complete_audio_and_edit
    - five_layer_qc
    - rights_snapshot
    - safe_release_drill
  not_required_yet:
    - full_season_automation
    - autonomous_final_approval
    - multi_region_scale

Write down what you are not doing first, so the 30 days do not turn into an endless project of buying tools and building dashboards.

H.2 Days 1–3: business and story foundations#

Complete the intake card, audience evidence, business model, series promise, the pilot's episode contract, the budget cap and the stop conditions. Choose an episode with two characters, one key prop, one piece of precise text and one visible state change.

Acceptance: someone who was not involved can restate what changes in the episode and what its closing question is; the producer can say what this pilot is validating; the rights lead can list the unknowns.

H.3 Days 4–6: the episode pack#

Complete beats, scene cards, facts and knowledge, the line registry, shot requirements, the sound plan, risks and insurance routes. Read for time three times and run pack lint.

Acceptance: no factual conflicts, no character foreknowledge, duration in range, alternatives for high-risk actions, and an explicit obligation to the next episode. No final video is generated yet.

H.4 Days 7–9: the minimum asset set#

Build only what the pilot needs: lead identities, wardrobe, locations, props, graphic templates, voices and musical themes. Identity completes angle, expression, backlight and two-person tests; locations complete a floor plan and camera zones.

Acceptance: every locked asset has an ID, version, provenance, rights and approval; retired candidates are quarantined; key text does not depend on generated pixels.

H.5 Days 10–11: state and continuity#

Write character, prop, location, time, weather, light and knowledge state for every scene. Write start, peak and end for complex actions. Establish the ten most important lint rules, even if they are enforced by hand at first.

Acceptance: any two adjacent shots can be compared on start and end state; prop transfers have events; axis and eyelines are derivable from the floor plan; a time jump does not reset all state.

H.6 Days 12–14: storyboard and animatic#

Turn beats into shot functions, and make static thumbnails and key state frames. Assemble an animatic with sound, adding temporary dialogue, music and graphics. Cut redundant shots, and fix the risks and the budget.

Acceptance: the main power change reads with the sound off; the conflict follows on audio alone; the duration is in range; and every expensive shot has a function nothing else can serve.

H.7 Days 15–18: controlled generation#

Low risk first, high risk after. Every call has a prompt manifest, a budget, an attempt log, acceptance criteria and stop conditions. Key actions get insurance coverage first. Candidates are quarantined and approved independently.

Acceptance: no stale assets, no facts outside authority, no candidates without provenance; failures are classified explicably; three failures on the same invariant stops; cost per approved second can be computed.

H.8 Days 19–21: sound and editing#

Lock the voice bible, the pronunciation lexicon and dialogue selection. Complete V0, a muted read, an audio-only listen, the pacing version, music cues, ambience and Foley, subtitles and graphics. Remix after picture lock.

Acceptance: voices are consistent across shots; music gives way; it is intelligible on a phone speaker; subtitles and graphics bind to facts; the timeline references approved takes.

H.9 Days 22–24: five-layer QC#

Check story, continuity, audiovisual, technical and rights as separate layers. Issues have evidence, severity, root cause, routing and regression. Fixes produce a new release candidate.

Acceptance: zero blockers; zero majors or a legitimate exception; key facts exact on OCR; identity, props and knowledge continuous; the rights snapshot complete.

H.10 Day 25: the release drill#

Generate files from the immutable release candidate manifest, have a second person verify, and publish locally or in a platform sandbox. Verify playback, subtitles, sound, cover art, events and rollback. Preserve the receipts.

Deliberately upload a version with a wrong checksum; the gate must catch it. If only the operator's memory of the right file protects you, the release system is not fit for purpose.

H.11 Days 26–27: failure and recovery#

Inject at least two of: duplicate events, a vendor timeout, a stale asset, budget exhaustion and a rights revocation. Observe detection, containment, evidence and recovery, with someone who did not build the system working from the runbook.

Acceptance: no duplicate spend, no unauthorized publishing, no lost state, no overwritten history; after recovery, the budget, task, asset and approval invariants all agree.

H.12 Day 28: creatives and feedback#

Produce three creative families, each declaring its real promise, its proof and its landing point. Build the event dictionary and the experiment registry. Simulated data is fine for rehearsing funnel diagnosis and converting comments into issues.

Acceptance: creatives neither leak nor invent a payoff; metrics locate the problem layer; data-driven suggestions are written back through change requests rather than editing locked story directly.

H.13 Day 29: cost and the scale gate#

Compute fixed costs, cost per approved second, human minutes, failure rates, review capacity and three budget scenarios for the first ten episodes. Find the bottleneck and decide whether to approve the next batch.

Do not approve a season because one episode finished. The scale gate looks at narrative comprehension, identity pass rates, approved-second throughput, traceability, rights and cash flow at minimum.

H.14 Day 30: project review and retrospective#

Play the work first, then pick a shot at random and trace back to task, prompt, state, cost, QC, approval and rights. Demonstrate one failure, one repair, one human overriding an automated recommendation, and one recovery.

Sort the retrospective into immediate improvements, needs more evidence, project-specific and disproven. Every improvement has an owner, a deadline and verification.

H.15 Minimum team#

A team of three: one producer-writer, one visual and generation, one sound, editing and QC — with a second person required to double-check rights and release. A team of six can separate producer, writer, visual continuity, generation, sound and editing, and QC and engineering. Roles can be combined; approval boundaries cannot disappear.

H.16 Minimum stack#

You can start with Git or a versioned directory, SQLite or structured tables, object storage, a task queue script, media probing, OCR, subtitle tooling and an editing application. What matters is not the number of products but stable IDs, immutable versions, manifests, events, budget and approvals.

Do not begin by wiring up a ten-agent group chat. Make one shot traceable from pack to release, then expand.

H.17 Maturity 0: accidental generation#

Characteristics: chat threads and personal folders; unversioned prompts; good results picked by hand; unclear rights; nothing reproducible. The goal is not to optimize the model, but to establish project IDs, asset records and approval state.

Promotion evidence: any published asset can be traced to its source, owner and rights.

H.18 Maturity 1: a reproducible pilot#

Characteristics: an episode pack, locked assets, attempt logs, basic QC and a delivery list — while still depending on a few key people.

Promotion evidence: another group can re-run one episode from the documentation, and state and cost are explicable.

H.19 Maturity 2: controllable batches#

Characteristics: state events, a dependency graph, dynamic auditing, budget reservations, workflow gates, human queues and release manifests. Several episodes run in parallel with shared assets under control.

Promotion evidence: an asset change produces a complete impact list; recovery after a failure works; the first ten episodes hold a stable approved-second throughput.

H.20 Maturity 3: production at scale#

Characteristics: a component release train, calibrated evaluators, staged rollout, backpressure, automatic rights gates, and derivation across languages and platforms. The team does not rely on a single hero.

Promotion evidence: vendor switching, a key person's absence, and a batch rollback drill all pass.

H.21 Maturity 4: a learning production system#

Characteristics: two-speed feedback, an experiment registry, declared scopes of applicability, cross-project recipe promotion and quality debt governance. Automation expands validated capability, not permissions.

Evidence: human overrides, failures and negative learning all improve the system while creative autonomy is preserved.

H.22 The maturity audit table#

Score each 0 to 3: absent, an individual habit, a team process, technically enforced.

Any rights, release or key-fact item scoring 0 cannot be covered by a high total.

H.23 Anti-maturity signals#

• Candidates keep increasing and approved seconds do not.
• Automated scoring gets finer and humans still cannot explain a decision.
• Tools keep multiplying and facts are still found in a group chat.
• Incidents are solved by one engineer firefighting remotely.
• Client changes go straight into generation with no impact analysis.
• Campaign data changes the story daily with no experiment registry.
• Images, music or voices cannot be shown to be licensed for commercial use.
• "Final" files keep accumulating and release candidates have no immutability.

H.24 The 90-day extension#

Days 31–60 scale to three episodes in parallel, validating shared identity, locations, music and state propagation, and introduce dependency queries, automated lint and budget reservations. Days 61–90 complete vendor switching, a rights revocation, disaster recovery, a small multi-language sample and a release train drill.

The condition for scaling is that the slowest gate has capacity — not that the generation API still has quota.

H.25 Annual governance#

Quarterly, audit asset rights, vendor terms, evaluator drift, recovery drills and keys. Twice a year, review series templates and quality debt. Annually, review which automation genuinely raised approved output and audience trust. Retire ineffective rules and expired recipes; a knowledge base has a lifecycle too.

H.26 Final completion checklist#

• One episode is fully traceable from commercial premise to release.
• One case — the book's or your own — runs through every workstation.
• Character, state, space, action, sound and music stay continuous.
• Prompts and model calls have versions, budgets and stop rules.
• Subtitles, graphics and precise facts come from authoritative sources.
• Five-layer QC and the rights gate pass.
• The agentic system holds no final creative or publishing authority.
• Releases can be rolled back and incidents can be recovered.
• Campaign feedback returns through experiments and permissions.
• The team knows which capabilities are reusable and which still need validation.

H.27 Closing implementation principles#

Lock the cheap decisions before the expensive generation. Establish state before chasing continuity. Preserve evidence before automating optimization. Make the system fail safely before expanding its autonomy. Deliver on the promise to the audience before optimizing commercial metrics.

A real production system does not promise to succeed on the first attempt every time. It promises that every attempt has a boundary, every failure leaves information, every decision can be traced, and every work is reliably finished within budget, within rights, and within human judgment.