Appendix C — Production Checklists

These are the book's gates, stripped to tick boxes. Each one is a stop condition rather than a suggestion: if a box is empty, the work does not move to the next stage, and the cheapest place to catch a defect is always the earliest one. Print the two you use weekly and keep them beside the machine; the rest are for the moments when a project starts, a persona ships, or a quarter ends.

The lists are deliberately flat. There is no scoring and no partial credit, because partial credit is what lets a defective frame into a published grid. If a checklist reads as too long for the work in front of you, that is information about the work rather than about the checklist.

Convention: Items are written as stop conditions, so an empty box means stop rather than note. Thresholds and dates come from the chapters cited in each heading. Where a threshold is a range, the chapter's range is reproduced rather than tightened here.

1. Character bible

From 2.3, 2.4, and 2.5. Complete this before the first generation. Stop condition: every box ticked, or the persona is not yet reproducible.

Identity

[ ] Legal name, spelling, and pronunciation.
[ ] Age, and whether it advances with the calendar.
[ ] Origin city and current city.
[ ] Languages spoken, and which is primary.
[ ] Education and current occupation.
[ ] Relationship status, family, and two close friends.

Physical

[ ] Height and build, with head-to-body ratio.
[ ] Distinguishing marks: scars, moles, tattoos.
[ ] Hair length, color, texture, and three staple styles.
[ ] Face: the twelve attributes the reference plate represents.
[ ] Hands: nails, jewelry, and which hand wears the ring.
[ ] Posture habits and two characteristic gestures.

Style

[ ] Palette: three dominant colors and two accents.
[ ] Brand posture: fictional labels or real ones.
[ ] Four defined occasions, each with one silhouette.
[ ] Seasonality: what the wardrobe does in winter.

World and props

[ ] Twenty recurring places named, three objects each.
[ ] Retired props and outfits recorded rather than deleted.
[ ] Every prop photographed as a clean reference plate.

Voice

[ ] Register, pace, and accent written down.
[ ] Three verbal tics or catchphrases.
[ ] Three things the persona never says, as rules rather than adjectives.

and disclosure

[ ] Ten canon facts, at least one falsifiable by a single post.
[ ] opened with birth year and current date.
[ ] In-character disclosure line written for the persona.
[ ] Brand kit: fonts, logo mark, and grid treatment.
[ ] Account-level bio written and published.
[ ] Pinned-post plan for the first month.
[ ] Bible versioned, and changes get a new version number.

2. Training dataset

From 4.4 and 4.5. Build this before you rent a GPU. Stop condition: every coverage box ticked or the gap named as a known limitation.

[ ] Twenty to fifty images gathered before any scoring.
[ ] Twenty to thirty survive into the training set.
[ ] Each image scored for identity first, coverage second.
[ ] Frontal, two fifteen-degree angles, two three-quarter views, two near-profiles.
[ ] Three pitch levels: level, slightly up, slightly down.
[ ] Four expressions: neutral, smile, talking, laughing.
[ ] Four lighting conditions: window, shade, softbox, one hard light.
[ ] Distance ranging from close-up to at least one full body.
[ ] Four or five distinct outfits.
[ ] Hair up, hair down, and one tied-back or wet frame.
[ ] At least three distinct backgrounds.
[ ] With and without glasses or accessories.
[ ] Five images held out and never trained on.
[ ] No watermarks, logos, or second people in any frame.
[ ] Every image resized and cropped to the training aspect.
[ ] Selection log records why each held-out image was excluded.

3. Likeness QA

From 4.6 and 5.6. Run the full battery after training, then the short form before every . Stop condition: nothing scoring 2 or below survives, regardless of the mean.

Nine framings

[ ] Frontal neutral, the reference condition.
[ ] Three-quarter left, where drift appears first.
[ ] Three-quarter right.
[ ] Profile left.
[ ] Profile right.
[ ] Low angle looking up.
[ ] High angle looking down, the selfie angle.
[ ] Close-up.
[ ] Full body.

Five lighting conditions

[ ] Window daylight, the baseline.
[ ] Ring light or on-camera flash.
[ ] Outdoor shade.
[ ] Softbox studio.
[ ] Hard direct flash.

Five emotional registers

[ ] Neutral.
[ ] Smile.
[ ] Laughing.
[ ] Talking with the mouth open.
[ ] Serious.

Scoring

[ ] Four dimensions scored separately, never averaged first: facial structure, skin and texture, body proportions, and "is this the same person?".
[ ] Batch work scores facial structure and same-person only.
[ ] Survivors scored individually after the contact sheet check.
[ ] Blinded comparison when two adapters compete.
[ ] Iteration register records base hash and chosen .
[ ] Short form accepted at eight generations in fifteen minutes for routine work.

4. Pre-publication

From 5.6, 6.7, and 11.3. Run this on every asset. Stop condition: any single failure returns the asset to the pipeline or discards it.

[ ] Technical scan at 100 percent: hands, limbs, shadows, merged objects.
[ ] No defect findable by a viewer in under two seconds.
[ ] Text inside the frame spelled correctly.
[ ] Identity scored against the , side by side rather than from memory.
[ ] Canon check against the ten facts and the timeline register.
[ ] Edit attempts capped at three; a fourth failure discards the frame.
[ ] Video generated voice first, then motion, then .
[ ] Delivery checklist passes for every platform in use.
[ ] Captions burned in, inside the safe zone, with a separate transcript kept.
[ ] Hook occupies the first two seconds, and it is generated rather than a fade.
[ ] Disclosure label present before the asset is scheduled.
[ ] Provenance record written: prompt, , references, timestamp.
[ ] Metadata recorded and the asset filed under the correct folder .

5. Pre-campaign compliance

From 11.1 through 11.8. [AS OF 2026-10] Run before any paid campaign, new platform, or new product. Stop condition: likeness documentation and disclosure placement are not negotiable.

[ ] Account-level declaration current on every platform in use.
[ ] Asset-level disclosure placed where it is unavoidable, not available.
[ ] Interaction-level disclosure set in the chat opener.
[ ] Both disclosures used when the post is paid, photorealistic, or advice-giving.
[ ] Reference face tier confirmed: invented, composite, or derived.
[ ] or likeness license on file for every human reference.
[ ] No real person named in any prompt.
[ ] Face search run on the generated identity before committing to it.
[ ] Retention periods set for logs, consent records, and invoices.
[ ] Data subject request process written and walked through once.
[ ] VAT path chosen, with consent recorded for non-refundable digital sales.
[ ] Platform policy reading dated within the current quarter.

6. Weekly operator

From 7.6. Four slots and one daily block. Stop condition: the drops below twenty approved frames.

Slots

[ ] Production 1, Monday, ninety minutes: generate, QA, schedule the first half.
[ ] Production 2, Wednesday, ninety minutes: second half plus video.
[ ] Production 3, Friday, sixty minutes: bank building, generating ahead of need.
[ ] Review, Sunday, forty-five minutes: analytics, refill, asset audit.

Daily

[ ] Chat block kept to its planned window.
[ ] Comments harvested and repeated questions logged.
[ ] Published assets counted against the plan.

Hygiene

[ ] run before every batch.
[ ] Twenty approved frames in reserve.
[ ] Backup restore tested on one folder this week.
[ ] Failed runs cleaned so no orphaned assets accumulate.

7. Quarterly audit short form

From 12.5, condensed to one page. Answer in writing with a date, score pass or fail, and stop at the first failure in each section. Stop condition: a failed section is fixed or the thing it covers is wound down before the next quarter opens.

Technical

[ ] Adapter, , and custom-node versions pinned and recorded.
[ ] Base model hash matches what the adapter was trained on.
[ ] Smoke test passes on three frames at production settings.
[ ] Reference library complete, with no missing entries.
[ ] Backup verified by an actual restore.
[ ] inside its normal band.

Distribution

[ ] Disclosure posture current for every post in the quarter.
[ ] Secondary platform active rather than existing.
[ ] captures growing quarter over quarter.
[ ] Cadence sustained at the planned rate.
[ ] Format mix intentional, with one test run this quarter.
[ ] Repeated questions from comments fed into the content plan.

Financial

[ ] known and compared to last quarter.
[ ] Pipeline spend within the planned tier, with variance explained.
[ ] Prepaid credit balance under one quarter of consumption.
[ ] Revenue composition known, with shares per stream.
[ ] and computed for the quarter.
[ ] Subscription utilization measured, with a tier decision if under a third.

Compliance

[ ] AI Act transparency duties reviewed against the current transition state.
[ ] Platform policies read, with the reading dated.
[ ] Likeness documentation current for every reference in use.
[ ] Retention periods actually implemented, not merely written.
[ ] Withdrawal rights and VAT handling reviewed for digital sales.
[ ] Every revenue stream has a decision attached, or has been stopped.

How to run these checklists

Three rules make the difference between a checklist and a ritual. First, tick boxes against the asset rather than against your memory of the asset — the scoring rules in section 3 exist because a face judged from memory drifts with it. Second, when a box fails, stop rather than annotate: a note that says "hands are rough but usable" is how a rejected frame reaches a published grid, and the retry costs less than the correction.

Third, keep the two lists you actually use on paper: section 4 for every post and section 6 for the week. The rest are event-driven, and a quarterly audit run from the same page in the same sitting is worth more than four audits run from memory. If a section fails, write the failure down with its date. The register is what turns a checklist into a trend line, and the trend line is what tells you whether the pipeline improved or you merely got faster at the same defects.