Specification — Full-Body Historical Virtual Human

One character asset, three delivery surfaces: kiosk, mobile AR, VR. Knowledge bound to the figure’s writings, contemporary accounts, and the intellectual world available in their lifetime. Not an open-web chatbot wearing a costume.

This note is a production spec, not a vendor pitch. Methods that already shipped are cited back to Case Study — Jumbo Mana Bonjour Vincent and Case Study — Dimensions in Testimony and StoryFile.

1. Product definition

A full-body, real-time 3D historical figure that can:

• Stand or sit at human scale in a gallery kiosk (or LED / projection)
• Appear in a visitor’s space through a phone or tablet (AR)
• Be visited in a reconstructed period environment (VR headset)
• Speak from a constrained knowledge base
• Perform authored events (monologue, reenactment, debate) and unscripted Q&A inside guardrails

The same identity package (mesh, wardrobe, voice, knowledge graph, conversation policy) is compiled to three runtimes. Do not build three characters.

What it is not

• Not Dimensions in Testimony (that requires a living sitter and unedited video).
• Not a HeyGen talking-head.
• Not “ChatGPT with a Lincoln skin” on the open web.
• Not a claim that the dead person has returned.

Label on every surface: An interpretation built from sources dated to [year]. It can be wrong. It cannot know what happened after [cutoff].

2. Delivery architecture (one asset, three clients)

Sources + scholar
        │
        ▼
 Identity package
 (look / move / speak / know / refuse)
        │
        ├─► Kiosk runtime     (UE Pixel Streaming or local GPU)
        ├─► Mobile AR runtime (ARKit / ARCore, or WebXR)
        └─► VR runtime        (Quest / PCVR, same UE project, different LOD + locomotion)

2.1 Kiosk

Best fidelity and the only place you should premiere the character.

• 55–86" portrait or landscape display, or a life-size LED blade
• Optional depth camera / mic array for approach detection and barge-in
• Local workstation (RTX 4070-class minimum; 4080/4090 if hair + Lumen + two characters) or Pixel Streaming from a GPU cloud box
• Always-on “rest pose + ambient life” so the figure is present before anyone speaks
• Captions on screen (ADA). Leeds Castle’s Eleanor model: environment-aware, multilingual, conversation as the end of a curatorial path, not the start.

Kiosk is also the capture booth for live events: a scholar or actor can drive the face via Live Link Face while the public sees the historical body.

2.2 Mobile AR

Purpose: take the figure out of the building and onto a school lawn, a battlefield overlook, or a kitchen table.

Two honest tiers:

Do not promise MetaHuman cinema quality on a mid-range Android in sunlight. Ship Tier A for public AR; reserve Tier B for a staff iPad tour or a Quest passthrough experience.

World-locking ideas that actually help: GPS + image target on a plaque; “stand on this mark”; AR Lincoln delivering Gettysburg at the ALPLM-style exhibit point.

2.3 VR

Purpose: place, not just person. The character needs a room that teaches (Einstein’s Zurich lodgings, a printing shop, a quarterdeck).

• Quest 3 / 3S class for volume; PCVR or Cloud XR for hero close-ups
• Teleport locomotion + seated mode; the figure should not require the visitor to stand in a 4m room
• Scale 1:1. Hand presence for pointing at objects the character can pick up
• Optional second occupant (teacher + student, or two historical figures)

MetaHuman 5.6–5.8 is the default body pipeline: Creator in-engine, body + face, Live Link from webcam/phone/audio, mobile LOD path that Epic has demoed with hundreds of characters on phone (crowds, not hero close-ups). Use UE Optimized assembly for AR/Quest and UE Cine for the kiosk hero.

3. Production pipeline

Eight stages. Later stages must be allowed to send the character back to stage 2.

Stage 0 — Figure selection gate

Score candidates on six axes before anyone opens MetaHuman Creator:

1. Source density — letters, speeches, published works, recorded voice if any
2. Likeness evidence — portraits, death mask, photos, measurements, clothing
3. Public appetite — search, curriculum, tourism, existing Chautauqua demand
4. Legal clearance — death date, domicile, estates, photograph copyrights
5. Moral heat — enslaved people, victims, living descendants, sacred status
6. Period knowledge bound — can you draw a clean cutoff year?

Green only if 1, 2, and 4 are strong. Popularity without sources produces a costume.

Stage 1 — Documentary spine (6–16 weeks, overlapping art)

Build three corpora, physically separate in the repo:

Add a fourth file that is not in the model: D. Modern scholar notes (anachronisms to avoid, contested facts, trigger topics, “we do not know”). This drives the refusal policy, not the voice.

Van der Veen’s rule from Vincent still applies: feed verified material so the system is not “spitting back Wikipedia.”

Output of stage 1: a sourced chronology, a voice dossier (diction, metaphors, jokes, temper), a list of things the figure would not know, and a question bank of 400–800 anticipated visitor questions with cited answers.

Stage 2 — Likeness and body

Reference order of authority:

1. Photographs / film (if any)
2. Life mask, death mask, skeletal metrics
3. Multiple independent portraits, not one romantic painting
4. Written descriptions of height, gait, scars, teeth, clothing
5. Period-accurate dress from costume historians (St Andrews: 12 garment sets for 24 bodies)

Build path:

1. Sculpt or Mesh-to-MetaHuman from the best 3D or photogrammetry of a reconstruction
2. MetaHuman Creator for topology, skin, teeth, body shape
3. Groom (hair/beard) as a separate art pass — this is where cheap builds die
4. Wardrobe in Marvelous Designer / CLO, then MetaHuman Outfit assets so clothes refit
5. Props: spectacles, pipe, rifle, palette, cane — each is a conversation affordance in VR
6. Two LODs minimum: Hero (kiosk/VR near) and Tour (AR/Quest/far)

If no reliable face exists (Genesee’s Backus problem), say so on the label and design a type with documented clothing and age, or pick another figure.

Do not use a living actor’s identifiable face without a contract. Dalí Lives did this with consent and an ad agency; a historical product should not silently deepfake a lookalike.

Stage 3 — Voice

Priority stack:

1. Restored historical recording + limited clone only if estate and ethics clear (rare before the 1890s)
2. Directed actor + dialect coach (Leeds Castle hired a native Spanish actor for Eleanor; ETH used a comedian for young Einstein after writing the lines)
3. Period-informed TTS with an actor’s supervised fine-tune, then lock the voice so it cannot drift

Record a voice bible: 30–60 minutes covering calm exposition, anger, humor, prayer, public oratory. That becomes both TTS training and the animation driver’s audio set.

Always ship captions. Always ship a “this is an actor / a model” line in the credits.

Stage 4 — Motion

Three layers, mixed at runtime:

Body in conversation: IK to a lectern or chair; gesture library tagged by rhetorical move (point, open hand, pace). Full-body mocap for the 8–20 minute monologue; do not mocap every Q&A.

For AR, bake or heavily compress the monologue as an animation sequence so the phone is not solving a live face.

Stage 5 — Mind (conversation system)

Hybrid, because pure generation and pure retrieval each fail a historical brief.

Visitor utterance
    │
    ├─ intent + safety filter
    ├─ if matches authored beat → play clip / animation (DiT-style retrieval)
    ├─ else RAG over Corpus A, then B, then C
    ├─ generate answer in-character with citations hidden from visitor, shown to staff
    ├─ cutoff + refusal policy
    └─ speech + face + optional gesture tag

Hard rules:

• Cutoff year is a system constraint, not a prompt suggestion. George AI’s “electric fence” is the right idea; feeding “today’s conditions” back in is how you get Glenn Beck in a wig.
• When the answer is from Corpus B, the figure should be able to say “Madison wrote that about me” rather than “I think.”
• Log every question and the corpus hit. Review weekly for the first 90 days, monthly after. Orsay learned that suicide was the top Vincent question only after install.
• Staff dashboard can force a canned reply onto a new FAQ overnight without a model retrain.

Do not put an LLM in front of survivor testimony. That method stays in the DiT note.

Stage 6 — Event layer (what the character does)

The conversation engine is only one mode. Ship a program clock so venues can book the figure the way they book a Chautauquan.

Classic Chautauqua three-beat (use this; audiences already understand it):

1. In character monologue (12–20 min), sourced, set in a dated moment
2. In character Q&A
3. Out of character — a scholar or a “Narrator” agent explains sources, disagreements, and what the figure could not have known

On a kiosk, beat 3 can be a second voice or a wall panel. In VR, beat 3 can be a doorway that leaves the period room.

Other event types are in §5.

Stage 7 — Compile to three clients

One content repo. Per-platform materials, hair cards vs strands, and animation compression differ. Do not fork the knowledge base.

Stage 8 — Evaluation before public

Test with:

• Historians who did not build the corpus (blind fact check of 50 answers)
• Teachers and a class
• People who hold the figure as an ancestor or faith object
• Accessibility review (captions, height of kiosk figure, seated VR, contrast)
• Red-team prompts: current elections, medical advice, “say slurs,” “predict the future,” “which religion is true”

Ship a public methodology page. Hide nothing about cutoff year and actor voice.

4. Subject matter for presentations

Think in programs, not “topics the chatbot knows.” Each program has a date, a place, a source list, and a learning goal.

4.1 Life and work

• Formative years, craft, a single decisive year (Van Gogh at Auvers; Einstein in 1905; Douglass in 1845)
• How a text was made (drafting the Declaration, composing a symphony, mapping a voyage)
• Daily material life: food, money, illness, servants, tools

4.2 Ideas in their own vocabulary

• A doctrine they actually defended, with the books on their desk
• What they got wrong by their own later lights if the corpus supports it
• Disputed points, presented as disputed (Corpus B on stage)

4.3 Place-based programs

• Site-specific AR: the figure at this house, battlefield, lab, church
• VR reconstruction of a lost room
• Walking tour: five anchors, two minutes each

4.4 Civic and classroom

• Primary-source seminar: visitor brings a document, figure responds only from Corpus A
• Exam prep / oral defense
• Language practice in the figure’s tongue (Eleanor: EN/FR/ES because those were her languages)

4.5 What to refuse as subject matter

• Medical or legal advice in modern terms
• Endorsement of living politicians
• Graphic replay of sexual violence or torture unless the venue is a memorial with a protocol
• “What do you think of [event after cutoff]?” as if they had an opinion. Offer beat 3 instead.

5. Interactive events the character can perform

5.1 Chautauqua monologue + Q&A

Still the best format ever designed for this. Scholar-performers already research primary sources, speak 45-ish minutes in first person, take questions in character, then drop character. A virtual figure should steal that clock exactly, including the drop-character beat, which is what keeps the product honest.

Length for kiosk/AR: 8–12 minutes authored + 5–10 minutes Q&A. VR can hold 20.

5.2 Reenactment of a dated scene

Not open world. A scene:

• Gettysburg Address, with the visitor as a listener in the crowd (VR) or on the lawn (AR)
• A salon argument with scripted beats and one improvisational middle
• A scientific demonstration the figure can actually perform with period tools
• A courtroom or parliamentary speech, visitor as clerk

Needs authored animation, a set, and a fail-safe if the visitor walks away.

5.3 Object lesson

Visitor points at a prop (or an AR-tagged museum object). The figure talks only about that object from Corpus A/B. This is the highest-trust mode after pure retrieval.

Historic Metahumans already did this with scanned Tel Dan artifacts.

5.4 Dual-character programs (the high-value unique event)

Two full-body figures, same engine, separate corpora and cutoff dates.

Implementation: two MetaHumans, two RAG indexes, a moderator policy that hands the floor. On a kiosk, split screen or a turntable. In VR, two chairs. In AR, only one body at a time unless you accept a mess.

Label counterfactuals in the title: A conversation that never occurred, built from letters of 1791–1826.

5.5 Classroom seminar

Teacher locks a topic and a source pack. Students may only ask from that pack. Analytics to the teacher (Hello History / Humy already sell this layer; you want it with a body).

5.6 Guided tour / concierge

The figure walks a pre-baked path in VR or appears at AR stops. Low improvisation, high production design. This is how you get throughput at a busy site.

5.7 Private audience

One visitor, booked slot, longer Q&A, like Leeds Castle’s chapel. Highest magic, worst throughput, needs a docent in earshot.

5.8 What not to stage

• Séance aesthetics
• Romance / companionship mode
• The figure as customer-service agent for the gift shop
• Unlabeled debates about living people

6. Resources to produce something faithful, complete, and watchable

“Faithful” is sources. “Complete” is a body that holds up at 2 meters plus a mind that can last 15 minutes. “Compelling” is performance and scene, not polygon count.

6.1 Team (hero figure, first year)

A serious first hero is 6–12 people part-time over 6–9 months, not a weekend LoRA.

6.2 Money (order-of-magnitude, 2026)

Reuse the bands already established in this vault:

On-going: GPU stream $0.50–$2/hr per concurrent kiosk if clouded; actor residuals if contracted that way; historian hours for log review; venue staff.

6.3 Stack (opinionated default)

• Unreal Engine 5.6+ with MetaHuman Creator in-engine
• Mesh-to-MetaHuman if you have a reconstruction mesh
• Live Link Face / webcam for capture; Audio2Face or MetaHuman audio-driven for unscripted
• Pixel Streaming for kiosk; Quest build for VR; ARKit + USDZ or WebXR for phone
• RAG over versioned corpora (A/B/C) with citation store
• Separate policy model or rules engine for cutoff and refusals
• Analytics that store questions, not visitor faces, unless the venue has a sign and a DPA

6.4 Time

First public hero: about 9 months if rights are clean. Faster only if you steal a reconstruction mesh and a published correspondence.

6.5 Completeness checklist

A figure is “complete” enough to show when it has:

• Cutoff year on the splash screen
• Corpus A in a citable edition
• Hero mesh that survives a 4-second orbit in kiosk
• Period wardrobe with a source card
• Voice credits
• One 10-minute authored monologue
• 200 tested Q&As with citations
• Refusal behavior for afterlife, medicine, living politicians
• Beat 3 (out of character)
• Captions
• A public methodology page
• A takedown / correction path

7. High-content, high-demand figures

“High-content” means a thick Corpus A. “Popular” means audiences already come. Prefer intersection.

Tier 1 — dense writings, global name, usually clearable

Tier 2 — strong demand, thinner or hotter corpus

Einstein (ETH already did young Einstein; photo/voice rights and Hebrew University history are messy), Tesla (myth factory), Cleopatra (almost no own words — fail the source-density gate unless you accept a composite of sources about her), Napoleon (good corpus, nationalist capture), Hypatia (almost no own words), Shakespeare (works ≠ private self; do the playwright-on-stage, not the inner life), Sacagawea (almost no own words; do not invent), living-memory figures (King, Kahlo, Orwell) — publicity and moral load jump.

Tier 3 — locally perfect

The sleeper hits are site-specific: a ship captain at a maritime museum, a mill girl at a factory, Eleanor of Castile at Leeds, Biran at Tel Dan. Source density can be moderate if the place carries the visitor. Dual-character with a famous correspondent then exports the local figure.

Selection rule

If Corpus A is thinner than a short monograph, do not build a full-body conversational human. Build a documentary, a reconstruction face, or a guided tour narration instead.

8. Legal, logistic, moral

8.1 Legal

Writings. In the U.S., works of authors dead more than 70 years are generally public domain (simpler for pre-1929 publications). Scholarly editions and translations can still be copyrighted — license the edition or use a PD text. Do not scrape a copyrighted collected-works site into RAG and call it research.

Photographs and portraits. A 16th-century painting may be PD; a 1955 museum photograph of that painting may not. A specific Einstein photo is still someone’s copyright even if Einstein’s persona is argued to be free in some states.

Right of publicity / NIL. U.S. state patchwork, no federal statute. Typical post-mortem terms: CA 70 years, NY 40 (domicile + date conditions), IN 100, TN potentially perpetual if exploited. About two dozen states recognize a post-mortem right. Many statutes target commercial use; news, commentary, and some biographical/documentary uses have exemptions (NY’s replica statute is explicit about historical and biographical works). Educational kiosks are safer than merch, apps with ads, or “meet Lincoln, subscribe now.” Get a memo per figure, per venue state, per monetization plan.

Voice. Some states now treat voice as publicity; federal likeness bills have been proposed. An actor contract must cover digital reuse across kiosk/AR/VR and future retraining.

Trademarks. Estates trademark names and signatures (Einstein, Edison, some presidential libraries). Search before the splash screen.

EU / UK / elsewhere. Personality rights and data-protection-adjacent claims differ. A figure who died recently in France is not the same case as Jefferson. If you ship AR worldwide, geofence or license.

Indigenous, sacred, and communal knowledge. Not a publicity-statute problem. It is a permissions problem with living communities. Do not build a prophet, a clan ancestor, or a ceremony leader from settler archives alone.

8.2 Logistic

• Throughput. A 15-minute private audience kills a gallery. Offer a 90-second “approach” mode plus booked slots.
• Noise. Kiosk mics in museums are hostile. Use array mics, a push-to-talk affordance, and big captions.
• School filters. Districts will block cloud LLMs. Offer an on-prem or allowlisted endpoint.
• Offline day. The kiosk must survive a dead circuit with authored monologues only.
• Versioning. When a historian corrects a fact, you need a corpus version number on the credits.
• Staffing. Someone has to reboot the box and read the question log. Budget a docent hour per open day.
• Heat and battery. Outdoor AR at noon will drop Tier B. Publish device requirements.
• Internationalization. Translating answers is not translating the person. Either keep the figure in their language with captions, or hire a historian-translator, not Google Translate on the output.

8.3 Moral

Non-negotiable labels: interpretation; cutoff; actor/model; sources.

Do not flatten. Jefferson without slavery, Gandhi without the ashraf and race writings, Curie without the cost to her body, are product failures, not politeness.

Do not comfort-generate. If the corpus is silent, say “I did not write about that.” Inventing a gentle opinion is how histobots become propaganda. George AI’s critics were reacting to this, not to the T-shirt.

Trauma. Suicide (Vincent), massacre, rape, medical racism: write protocols before install. Offer a visible exit. Do not let the figure give modern clinical advice.

Descendants and communities. Ask when they exist and are findable. Biran’s family permission is the model. “Public domain” is not the same as “no one can be harmed.”

Children. Default the figure out of companionship language. No “I love you.” No secret-keeping.

Labor. An actor’s face or voice inside the figure is labor. Contract it. A historian’s years of edition work is labor. Cite the edition.

Dual-character debates can manufacture a verdict the past did not reach. If you stage Lincoln–Douglas, use 1858 texts. If you stage Lincoln × a 2026 avatar, the modern one must carry the anachronism on its face.

Afterlife of the file. Decide who may retrain, reskin, or fine-tune the figure in five years. Put it in the contract with the museum. Do not let a marketing team open the cutoff.

9. Recommended first build

If the brief is “faithful, complete, compelling, three surfaces”:

1. Pick a Tier 1 writer with a PD corpus and a place (Douglass at a specific hall; Franklin in a print shop; Curie in a lab; a local Tier 3 with letters).
2. Ship kiosk + Quest in season one. Ship billboard AR as a postcard of the same monologue, not a second mind.
3. Program clock: one Chautauqua, one object-lesson tour, one labeled dual with a clerk-avatar.
4. Publish the methodology page the same day as the splash screen.

That is the Bonjour Vincent lesson (scholar + letters + venue + afterlife) plus the DiT lesson (do not invent the voice of the dead when a real voice could have been recorded) plus the Chautauqua lesson (always drop character).

See also

• Case Study — Jumbo Mana Bonjour Vincent
• Case Study — Dimensions in Testimony and StoryFile
• Citations — Historical Virtual Characters
• Citations — Digital Humans