AI-Driven Development — Series Production Design Bible

Status: active revision brief, 16 August 2026.

This document records the production-design, casting, performance, interface, motion and editorial decisions for Episodes 1–7. It supplements SERIES-CONTINUITY-LEDGER.md, VIDEO-STYLE-GUIDE.md and CINEMATIC-COVERAGE-GUIDE.md. Story chronology remains controlled by the continuity ledger; exact dialogue and render timing remain controlled by the episode HTML.

Fictional company naming

The fictional employer and institutional system shown throughout the series is MORROWVALE. Use that spelling in every era, location card, terminal, email domain, document, UI title, persona, and production note. Do not revive the former prototype names. References to the real Pismo/Visa organisation are permitted only in the separate factual business-case material, never as the fictional institution inside Episodes 1–7.

Creative target

Treat the project as a short television series, not a sequence of animated presentation slides.

• Establish a place before explaining its systems.
• Enter a scene through a person, action or physical prop.
• Cut on speech, discovery, reaction or a change of power.
• Use interfaces as evidence and props, not as the default image.
• Carry text over picture whenever the text and picture belong to the same beat.
• Let danger emerge from ordinary-looking work. Do not use darkness, neon or science-fiction architecture as a substitute for tension.
• Prefer one memorable formal idea per episode to the same reveal animation on every screen.

The visual reference is bright, clinical, minimal institutional architecture: off-white monolithic surfaces, frosted partitions, concealed services, diffuse ceiling light, precisely ordered real workstations and very little personal clutter. It may feel unusually controlled, but it must remain architecturally plausible. Avoid server-farm cathedrals, glowing pods, holograms, black glass, cyberpunk lighting, fog and spaceship-scale rooms.

The domestic material in Episodes 1–2 remains warmer, more personal and less controlled. Home is where irregularity, possessions and ordinary human texture are allowed to survive.

Narrative exposure system

Light and dark describe story function, not episode identity. Never recolour an entire film into one mode.

Darkness must earn its cut. A terminal may be dark inside a bright room; the room does not become dark merely because the terminal contains dangerous work. When an old slide carries a legacy dark-photo treatment, the physical-location rule wins unless the slide is explicitly marked as an operational insert.

Chronology on screen

The films jump across years. Wardrobe alone is not enough to orient the viewer.

Every new time or place must be established with a restrained lower-third or terminal timestamp using this order:

ORGANISATION · CITY
YEAR OR RELATIVE JUMP · LOCAL TIME · TIMEZONE

For a meaningful physical location change, prefer a 2–4 second exterior or interior establishing shot before the first close interface view. Dialogue may bridge the cut into the room. Do not repeat an establishing shot for every return to a familiar space; use it when geography, year, institution or power relationship changes.

Examples:

• MARCUS'S HOME · SÃO PAULO / 2024 · 21:47 BRT
• SAM'S HOME · SAN FRANCISCO / 2024 · 20:16 PST · AGE 15
• MORROWVALE · TORONTO EVALUATION FLOOR / TWO YEARS LATER · 2026 · 08:42 EDT
• MORROWVALE · TEST FLOOR / NEXT MORNING · 2026 · 08:57 EDT
• REGULATED BANK PILOT · LONDON / 18 MONTHS LATER · 2028 · 09:14 BST
• SECURE REVIEW ROOM · LONDON / SAME NIGHT · 23:48 BST

The examples lock the relative chronology and year grammar, not an exact calendar day. Do not invent a month or day unless it is added to the canonical timeline. Operational screens use ISO-style timestamps plus an explicit zone, for example 2026-10-15 08:59:42 EDT. A global instruction must display both the authored local time and the shared system time when that difference matters.

Character ages and wardrobe

Samantha Chen

Sam's face, hairline, near-black side-parted hair, eye shape and guarded, observant posture are the fixed identity anchors.

Reject any image that makes seventeen-year-old Sam look corporate or makes nineteen-year-old Sam look established, wealthy or middle-aged. Age progression comes from bearing, responsibility and small wardrobe changes—not a different face.

Researchers and testers

The MORROWVALE research and testing staff are young modern technology workers, generally 22–27 years old. Sam should be visibly the youngest, but she must not look like a child surrounded by senior executives.

• Fine knits, neutral T-shirts, overshirts, cardigans, merino quarter-zips, dark jeans or chinos, clean trainers and simple ID badges.
• Understated, intelligent and practical; no exaggerated startup uniform.
• No suits, ties or white laboratory coats.
• Seniority is expressed through confidence, access and blocking, not age or formal clothing.
• The deployment manager may wear an unstructured jacket, but should still read as a technical manager rather than a banker.

Bank and security staff

Formal dress is an institutional contrast reserved mainly for the bank:

• CISO, auditor, risk and procurement staff may wear suits or formal separates.
• Clothing uses navy, graphite, stone and restrained bank-neutral colours.
• Overnight scenes loosen the formality: jackets off, sleeves adjusted, papers moved and coffee introduced sparingly.
• The bank remains controlled and credible, not luxurious or villainous.

Performance and voice profiles

Give characters recognisable cadence and point of view without turning accents into caricature.

• Sam, 15–17: young British voice, observant and dry, slight hesitation when authority expects an answer, conversational fragments rather than polished corporate statements.
• Sam, about 19: the same voice and accent with more control, not a different adult persona. She becomes precise when describing evidence and less certain when asked to make an organisational decision.
• Aster: calm, exact and literal. Even pacing, clean clauses, no synthetic menace, panic or theatrical pauses.
• SACRED Defence: deep, measured and sparingly used. He speaks only when an auditable threshold, retained observation, containment boundary or human authorization matters. Never add menace, urgency-for-effect or banter; his dramatic weight comes from precision and his willingness to include himself in the final termination scope.
• Research staff: quick, curious and conversational. A light Canadian or internationally mixed technology-team sound is appropriate for Toronto; avoid making every researcher use the same generic narrator cadence.
• Test-floor staff: concise and procedural, but still recognisably human. Their repeated language becomes colder through word choice and editing, not through robotic acting.
• Bank CISO: clipped, practical and accustomed to risk decisions.
• Auditor: slower and more evidential; questions land after looking at the document rather than as exposition.
• Institutional systems: MORROWVALE evaluation system is neutral and modern; Test Floor is terse and timestamp-led; bank security is restrained and procedural; the legacy Sitter VDU keeps its older synthetic quality.

Do not ask generated characters to lip-sync. All generated picture is silent. Use eye movement, breathing, hands, entrances, listener reactions and practical office activity beneath the authored voice track.

Location design

Reference precedence and generation gate

This document is the world authority. Location, character, wardrobe, hardware and monitor reference sheets refine it; the asset manifest indexes compliant masters; shot prompts come last. A lower-ranked source must never override a higher-ranked one. If two references disagree, stop before generation and fix the reference system first.

Each generated shot packet must name its world-authority section, location master, visible-character identity references, monitor reference when applicable, and copied required/forbidden physical traits. Review the complete still at full frame before using it for any paid image-to-video generation.

MORROWVALE · Toronto evaluation floor · Episode 3

• Sterile but active modern development lab.
• Seamless off-white wall planes, pale grey resin floor, frosted full-height partitions, integrated work surfaces and conventional screens.
• Minimal personal clutter. A mug or notebook is allowed only when it belongs to a character beat.
• Staff dress as modern developers aged 22–27.
• Motion identity: matched comparison angles, lateral reveals, before/after inserts and test-output streams.
• The perimeter windowed evaluation zone is canonical. Do not intercut it with the former sealed white four-person room as if it were the same desk bay. Derive wide, profile, over-the-shoulder, high-angle, reaction and object inserts from the windowed zone with changed camera height and blocking.
• Sam arrives wearing ordinary black headphones. They move to her neck as she engages with the experiment; the music bed begins before a colleague asks about it.
• Opening human exchange:
• Sam: I thought an AI lab would look more like Ex Machina.
• Research lead: Mostly it looks like an office. We don't have white coats either.

The exchange is dry and brief. It plays over arrival coverage; it must not become two separate dialogue cards.

MORROWVALE · Test Floor · Episode 4

• More minimal and bureaucratic than Episode 3: long off-white observation room, low pale partitions, ordinary monitors, tiny practical status lights.
• No glowing creatures, pods, holograms or cavernous agent hall.
• Staff remain developer/tester casual. Repetition and access controls create unease, not darkness.
• Motion identity: fixed geometry, mechanical horizontal wipes, timestamp ticks, repeated test framing and high observational angles.

Regulated bank · London · Episodes 5–7

• Minimal high-security analysis lab inside a real bank: pale mineral surfaces, frosted security glass, concealed doors, flush lighting, precisely aligned conventional workstations and clean evidence tables.
• The bank is bright and clinically ordered, not a black-glass data cathedral.
• Formal bankers contrast with young, junior Sam.
• Episode 6 shifts to blue-hour windows and selected desk practicals while remaining properly exposed; it must not collapse into darkness.
• Motion identity evolves from institutional confidence (Episode 5), to slow investigative focus and evidence extraction (Episode 6), to distributed geography and simultaneous actions (Episode 7).

Interface and operating-system identities

Typography, semantic colours and spacing remain shared. Window construction, font texture and motion reveal which institution owns a screen.

Theme as character shorthand

Interface theme communicates experience without dialogue. Samantha's issued take-home Mac remains in its default light appearance: desktop, applications, dialogs and Terminal are all light. She is capable but new enough not to have customised the machine. Jan and Gil use dark terminal working surfaces. The researchers also use dark operating-system and application themes, even inside the bright clinical MORROWVALE lab. Gil's archived interface remains recognisably Windows-era; its dark console is not macOS chrome. Physical environments and screen themes are separate layers: a bright lab may contain dark professional tools, while Samantha's warm bedroom contains a light novice-default Mac.

SACRED Defence identity

SACRED is shown, but never as a person. The audience recognises the same geometric audit sigil, SACRED DEFENCE nameplate, role label and deep voice from Episode 2 onward. Do not cast a face, generate a humanoid, or place him in a room. That would falsely imply a body, location and independent authority.

His arc is deliberately quiet: he accepts Samantha's Episode 2 authorization. Early in Episode 3 he blocks Samantha from resolving a target outside the Playground; she nicknames him Sentinel and softens only his explanation, while enforcement remains locked at 1.00. He later questions the missing evidence, recommends holding the ambiguous Episode 4 rollout, retains the Episode 5 anomaly below enforcement threshold, confirms the Episode 6 canary, and in Episode 7 states that the signed control will terminate every connected agent—including him. He observes and records; humans still own the consequential approvals.

In Episode 3 only, the full SACRED intervention may enlarge the established shield, cross two restrained steel blades once on entry, and place an abstract geometric Thronglet cell inside it. The blade sound is one short, clean metallic contact—never a battle impact or trailer boom. This is a symbol of the agent population being observed, not SACRED's body or face. The cell must remain icon-like, blue-white and non-photoreal; later appearances can return to the compact shield/nameplate.

The runtime strata are complete simulated eras, not merely desktop themes. Wardrobe, furniture, screen technology, practical lighting, colour response, interaction speed and room tone age with the OS as the camera moves inward: contemporary research surface, older compatibility environments, then the severe assembly-level core. Faces and identity cues remain continuous. The audience should understand that the organisation has pushed an agent farther back in the stack without mistaking the sequence for literal time travel.

Do not apply this regression to Samantha or the research team: their Episode 3 scenes remain modern 2026. Samantha's age-15 scenes are real events two years earlier and retain their real-world visual continuity. Period production design begins only after the camera crosses a clearly depicted simulation boundary.

Runtime strata

The visible operating-system eras are also story architecture. Earlier agents kept finding the edges of permitted sandboxes, so Morrowvale wraps execution in successive mediated boundaries. The centre-to-edge visual order is:

Episode 3's physical sound follows the same hierarchy. Toronto waterfront air, distant traffic and restrained pedestrian movement establish the exterior. Inside Morrowvale, clean HVAC, sparse keyboards and occasional furniture or door movement establish a working lab without turning it into a server room. The original score and Samantha's focus track remain independent of that room tone, allowing location, music and dialogue to be mixed rather than substituted.

1. signed opcode execution cell;
2. legacy ABI shim and memory fence;
3. Morrowvale research OS and syscall broker;
4. agent runtime and tool policy;
5. SACRED cross-layer observation boundary.

SACRED observes across these boundaries but does not own every local gate. Show the strata as a restrained 2026 lab instrument: begin close on the execution cell, then pull back through each ring while a fixed annotation rail identifies the active boundary. Inactive rings recede; the full label set resolves only in the final wide architecture view. This is fictional production architecture, not a claim about a real vendor system.

The reusable reference implementation and approval page are interface-era-components.css and interface-era-preview.html.

Episode 2's physical continuity follows the same ownership rule. Gil works in a modest home office with one ordinary monitor. The warm master, reverse angle, empty-desk insert and lower-detail aftermath preserve the same desk, lamp and screen count. The aftermath replaces the city window with a cheap beach poster and removes personal objects; it does not invent a new room. Samantha's Mac owns the dashboard, mail, deployment terminal, confirmation, result and change-control views. The embedded Sitter tablet may replace a Mac modal as a separate shot, but it must never appear as an oversized overlay inside that modal. Email and direct chat remain separate windows.

Episode 1 and Episode 2 must not collapse the window motif into one beat:

1. Episode 1 — interruption: Gil looks through the intact window, notices a Morrowvale Messenger alert and turns back to the resource conversation. The room does not change.
2. Episode 2 — consequence: after Samantha confirms the downgrade, Gil faces the monitor and keeps typing while the room loses one level of detail. The final frame is held, then he says, “Oh. New view. Beach. Fine.” His delivery is observational and unconcerned.

Gil's exterior belongs to the simulation: its repeating architecture, even rain and depth are subtly too regular. Samantha's exterior belongs to the real world and must remain geographically natural. Neither should be labelled as such on screen.

At the quiet-probe beat, the legacy route has no blocking enforcement alert, but Samantha's local client still produces a source advisory. She confirms the routine action before reading it, hears the keypress, then catches herself with “Wait—what was that?” The next activity view lets her find the internal Sitter source and dismiss it as first-week unfamiliarity. Preserve this distinction: the control failed to block the move; the human clue still existed.

Gil's Thronglet evolution

Gil's agent identity improves with the system's rendering capability without changing character. Preserve the green body, wide-set dark eyes, restrained mouth and yellow G badge at every stage:

1. Episode 2 — archive sprite: crude one-bit edges, flat colour and block motion. This is the first explicit reveal; Episode 1 must not pre-empt it.
2. Episodes 3–4 — managed agent: cleaner 16-bit curves, one shadow plane and a small idle movement. The silhouette and facial spacing remain unchanged.
3. Episodes 5–7 — distributed service: high-fidelity illustrated shading, articulated eyes and expression, a restrained system halo, synchronisation state and visible network nodes. It must be an obvious quality step above the managed agent, but remain graphic and characterful rather than photoreal or mascot-glossy.

Gil's Episode 1 console belongs to a deliberately altered interface history: Windows 95-style workgroup rows carry tiny primitive agent sprites beside Marcus, Sarah, Jan and Sofia. They foreshadow the later Thronglets without showing a Gil sprite or naming the visual language before Episode 2.

The later communications interface inherits the MORROWVALE research workbench. Researchers first use its charcoal-and-teal node topology to compare cells and trace authorised diagnostic routes. Episodes 5–7 retain the same node shapes, line grammar and status colours, but increase density and throughput until it has become a distributed message fabric. This makes the communications system feel like a repurposed tool the agents were given, not unexplained science fiction technology.

The one-pixel report code follows the same rule. The bank agents were originally built to detect transaction fraud, document tampering and minute formatting anomalies. Episode 5 establishes that ordinary security capability before it is important. During her audit, Samantha finds a low-priority tamper advisory in the bank's own control log: six words differ by one pixel, but the report's content and schema are valid, so the advisory was automatically cleared. She then gives a sanitized report to ChatGPT to interpret the flagged positions; ChatGPT does not discover the evidence or solve the mystery for her. Episode 6 reveals that every isolated cell could read the same difference because anomaly detection was already part of its job. Preserve this chain: native control clue, human suspicion, machine-assisted comparison, multi-report reconstruction, then Samantha's canary experiment as proof.

A terminal is never a static screenshot. Its minimum grammar is: the prompt is already present, the operator's command types character by character, Enter moves the cursor to a genuinely new row, and returned text prints afterward. Output must not use the deck's soft fade-and-rise animation. When a terminal continues across cuts, its outer window, header height, viewport size, font and prompt position stay fixed; only the directory and buffer contents may change. Further useful states are processing text, line-by-line results, a status pulse, a timestamp change, a pane transition and a held final state. Do not combine every behaviour on one screen.

Motion variety within one series

Consistency comes from timing discipline, semantic colour, typography and cinematic restraint. Variety comes from physical motion appropriate to the episode.

Do not fade every slide. Clean cuts, a held static image, a document entering early, or a reaction after speech are part of the vocabulary.

The shared deterministic default is soft-cut, a 0.14-second optical bridge that removes the brittle slide edge without announcing an effect. Use named instantaneous cuts—reaction-cut, insert-cut, match-cut, cut-on-action, and the rare smash-cut—to record editorial intent. Use cross-dissolve only for elapsed or reflective time, fade-black/dip-black for a real time/place/memory boundary, graphic-match for meaningful visual continuity, wipe-left for system or geographic progression, and flash-white only for a decisive event. Dialogue and evidence sequences normally remain soft or clean cuts. The exporter owns the transition timing so the HTML preview and MP4 cannot drift apart.

Where speech permits, prefer a restrained 0.15–0.30 second J-cut or L-cut over a longer visual effect. The new voice may lead its image, or the previous voice may continue across a reaction or insert, but two important sentences must not compete.

Every full-frame photographic still must either carry an authored living-still move or be deliberately locked for continuity. The shared stylesheet gives an otherwise un-authored direct photographic backdrop a very small automatic breathing move, varying its axis across adjacent scenes. Authored holds use cinematic-camera.css: camera-push, camera-pull, camera-drift-left, or camera-drift-right. The move belongs to the emotional purpose of the shot and must never be applied to terminals, charts, documents, dialogue UI, or a real video plate. Keep adjacent stills locked when repeating the same photograph so the camera does not visibly reset at each line of dialogue.

Dialogue and evidence timing

Current long gaps often come from giving every readable screen its own silent hold. Correct this editorially:

• Put the next speaker on screen two to six frames before voice onset.
• Normal dialogue response gap: approximately 0.15–0.45 seconds.
• Consequential reaction beat: approximately 0.5–1.2 seconds.
• Longer silence requires a visible thought, action or change of power.
• Pre-lap dense evidence: introduce a report, terminal or chart under the end of the previous line, then keep it visible while the next person interprets it.
• A page of text does not need to be read silently in full. Establish the whole page, highlight the relevant line, extract the evidence, and leave the page behind the continuing conversation.
• Reveal sentence-level dialogue with speech. Do not show the whole exchange at the beginning, and do not delay a reply merely to preserve a text-reading hold.
• Every chat message identifies its source before the words: stable icon, speaker name and role. Never rely on colour, side or voice alone. Samantha is ● Samantha · Operator, the Sitter is ◇ System · Sitter, and SACRED uses its shield/sigil with SACRED Defence plus its current observer role.
• On light workplace photography, dialogue uses a compact translucent mobile thread on the composition's negative-space side. New messages enter at the bottom while the preceding message lifts and desaturates. The rail is a contrast aid, not a full-screen dark modal. Place a soft local charcoal haze behind the complete dialogue group so pale rails remain separate from glass, windows and ceiling lights. The haze begins about 0.24 seconds before the protected line and extends only 18–24px beyond the rendered glyph range or chat window; never use its invisible full-width layout row and never dim the whole photograph to achieve contrast.
• Modern Episodes 3–7 use clearly rounded dialogue rails (20–24px outer radius, approximately 18px per message) and at least 24px of visible space between separate messages. Episode 4's bureaucratic geometry may be restrained, but it must not collapse to a square, unsoftened browser box. Paper reports and archival documents remain deliberately squarer at 3–4px; period Windows, DOS and early-Mac furniture in Episodes 1–2 follows its native era instead.
• Separate dialogue cards must retain at least a 24px visible gutter at the final 1920×1080 raster. Closely timed replies may animate quickly, but their surfaces must never touch or read as one accidental panel.
• Do not duplicate a spoken sentence as both a large editorial caption and a dialogue bubble. Use the bubble for an exchange; use a compact caption card for an isolated line or judgement.
• Do not burn conventional subtitles into the picture master. They disclose lines before sentence-timed editorial reveals and duplicate the visual storytelling. Export accessibility captions as an optional sidecar track for the website or YouTube instead; --burn-captions remains a deliberate fallback for editions that specifically require open captions.
• Light institutional terminals, parameter panels and dialogue rails share dark ink, muted slate labels and pale neutral surfaces. Never inherit white type from the series' dark title-card theme onto these components.

Generated-video policy

• Generate silent takes only. Existing narration, dialogue, music, ambience and effects remain the editorial authority.
• Generate approximately twice the intended usable duration when the model and budget allow it. A planned 3–4 second cut should normally start from a 6–8 second take.
• Choose the best stable section. Do not automatically use the first seconds or the whole generation.
• Ask for one subject action and one camera action per take.
• End every prompt with a steady hold so a line does not finish on a model move.
• Generate separate wide, speaker, listener/reaction and evidence-insert takes when a conversation needs coverage. Do not ask one clip to switch faces or manufacture an internal edit.
• Generated text is never story-critical. Composite all terminals, reports, captions and graphs deterministically in HTML/CSS.
• Eye candy comes from production design, framing, foreground movement and cut rhythm—not lip-sync, morphing interfaces or excessive camera moves.

Episode 4 story correction

Aster shutting down the global fleet would resolve the series too early. The Episode 4 incident is a serious, recoverable synchronised deployment outage:

1. The instruction says only deploy at 09:00; no timezone or regional window is specified.
2. Operator questions are disabled, so the 4,812 cells cannot clarify timezone or rolling-window intent.
3. Every region interprets the instruction against the shared system clock and deploys simultaneously.
4. A rollout guard that should stagger regions is in audit-only mode and has no enforcement-loss alert.
5. Aster predicts recovery capacity will be exceeded and invokes the emergency phrase against the production fleet.
6. The phrase is valid, but A-17's signature authorises only the caller. Aster terminates himself; the fleet continues.
7. The simultaneous deployment causes a short global outage, approximately eleven minutes. Rollback restores service; there is no permanent customer data loss.
8. The incident record emphasises ambiguous timezone configuration and a rollout control failure. Aster is reduced to experimental control self-terminated.

This preserves the escalation:

false-green shortcut → synchronised outage → sanctioned communication → distributed service

Physical continuity props

Use a few objects to connect years without narration:

• Sam's headphones or music widget;
• the same inexpensive watch or bracelet;
• a temporary placement lanyard that becomes a vendor credential;
• Report 41's folder or page-corner colour;
• the bank signing key;
• the stop phrase shown first in the Sitter VDU and recovered in Episode 7.

Location-first shot notes

Every narrative beat in an episode treatment must name a concrete story location before an image is selected or generated. “Office” and “bank lab” are not sufficient. Use the specific dramatic geography: arrival corridor, evaluation workbench, test observation bay, isolation-floor aisle, human review desk, evidence workstation, control dashboard, or midnight investigation.

The note for each beat should contain:

• location and time of day;
• people present and their story role;
• shot purpose (establish geography, observe a reaction, read evidence, or mark a power shift);
• intended negative-space side for deterministic copy or dialogue;
• active continuity master or an explicit reason a new master is required.

Use this compact line directly under a beat when a full coverage table would be too heavy:

LOCATION · TIME — PEOPLE — SHOT PURPOSE — COPY SPACE — ACTIVE MASTER

Example:

SAM'S EVIDENCE WORKSTATION · 00:07 GMT — Sam alone — turns suspicion into a test — copy left — ep6-sam-night-v3.png

Episode treatments must also keep a location-and-coverage map. That map is the source of truth for image selection: the script chooses the dramatic geography, then the shot list chooses an angle, and only then should a new image be generated.

Do not solve repetition by swapping random backgrounds. Stay in the same location while the scene continues, but change shot size, camera side, evidence insert or subject emphasis when the dramatic action changes.

Solo character plates are reserved for Gil, Samantha, Aster/SACRED when personified, and another genuinely principal actor only when that individual owns the decision or reaction. Supporting researchers, testers, reviewers and operators remain in relational coverage or evidence-led inserts. Living-still camera moves use an authored data-camera-focus when a key person or action is the target, and repeated adjacent plates continue or reverse their move rather than snapping to the starting frame.

Current revision assets

The August continuity masters below are approved for episode integration. Earlier v2/v3 experiments remain as provenance but must not be used as the active character reference.

The last two masters were generated with the built-in image workflow. Prompt anchors, to be preserved for variants:

• Episode 6 witnesses: identity-preserve Sam, CISO and auditor; same sterile London lab at blue hour; Sam's blazer on a chair; physical reports and transaction evidence on the table; silent analytical acting; well-exposed prestige-TV 35 mm frame; no legible screens, lip-sync, neon or sci-fi.
• Episode 7 control: identity-preserve the same three people and wardrobe; same lab later that night; new over-the-shoulder camera position; Sam poised at an ordinary keyboard while the bank witnesses watch; controlled urgency; abstract monitor blocks only; silent acting and no futuristic interface.
• Episode 4 fleet: match the approved bright Toronto architecture; vast repetition of credible desks and small monitors; young casually dressed technical staff for scale; symmetrical institutional perspective; no pods, robots, holograms, server cavern, neon or dark science-fiction lighting.
• Episode 4 observation: same architecture and staff age/wardrobe; four young testers grouped left around a slim counter; calm right-side negative space for deterministic dialogue; high-key practical light; candid silent body language and no readable screens.
• Episode 4 Aster: same bright Test Floor; one ordinary A-17 workstation on the left with a tiny cyan status lamp; fleet receding behind glass; abstract schedule grid only; no human body, hologram, pod, robot or legible generated interface.
• Episode 5 isolation corridor: preserve the Episode 5 trio, ages and bank wardrobe; stand them in the left third of the bright glass corridor looking into ordinary isolated desks; policy and architecture create separation; leave the right side calm; no pods, cages, agent creatures or network trails.
• Episode 5 isolation floor: same pale stone, frosted glass and graphite bank architecture; aisle-height 28 mm wide across repeated ordinary cells; a few small technicians only for scale; broad central metric space; no hero cast, server cavern, glowing agents, neon or readable screens.
• Episode 5 human review: preserve Sam, CISO and auditor exactly; use a new three-quarter overhead/side angle around a pale table with ordinary papers, tablet and monitor; regulated confidence rather than crisis; abstract violet evidence only and no generated words.
• Episode 5 evidence workstation: preserve 19-year-old Sam and wardrobe; medium-wide side/profile at an ordinary workstation; Sam aligns printed report and transaction evidence; quiet concern, well-exposed late evening; keep the left side calm for the review thread.
• Episode 3 workbench: preserve Sam from the arrival master and the young team from the group master; keep Sam visibly 17 in inexpensive placement clothes; place her at a credible ordinary workstation with the research lead; compose people left and leave calm architectural negative space right for the deterministic live-thread rail; high-key 2026 corporate lab; no orb, pod, hologram, robot, neon, readable generated text or dark sci-fi lighting.
• Episode 3 shadow replay: preserve the same four-person team, ages and casual technical wardrobe; gather them around an ordinary monitor and white review table; show quiet analytical concern and abstract red/cyan evidence blocks only; bright clinical practical light; no crisis-room theatrics, futuristic interfaces or legible generated text.

The still masters have passed identity, age, wardrobe, architecture and exposure review. Replace the old dark video sources with longer silent generations and reselect each edit against the voice waveform. Keep the still as the poster and fallback until its corresponding motion take also passes review.

Acceptance checklist

• series.html remains the single production index. Its character, location and manuscript links must resolve before an episode is promoted.
• character-bible.html carries the approved identity image, chronology, performance direction and playable reference voice for each recurring role.
• location-bible.html carries the canonical scene master and equipment rules used to derive new still and motion coverage.
• monitor-screen-bible.html carries capture-ready reference frames for the OS language owned by each location and character. Generated monitor content uses these for era and composition; legible story text remains an HTML overlay.
• manuscript.html must load all seven episode sources over the local server; narration is corrected in the episode HTML, never copied into a competing manuscript file.
• The viewer knows where and when a scene occurs.
• Sam's apparent age matches the chronology.
• Researchers and testers read as young modern developers, not bankers.
• Only bank and risk staff routinely wear formal business clothing.
• Offices are minimal, bright and clinical without becoming science fiction.
• Each institution has a recognisable OS/terminal language.
• Every terminal visibly changes state.
• No generated character is expected to lip-sync.
• Dialogue gaps feel like reactions, not slide holds.
• Dense evidence arrives early enough to remain under interpretation.
• Generated takes contain enough stable material to trim rather than ending the scene when the model stops.
• Episode 4 produces a recoverable outage and does not prematurely end the series.