Case Study — Dimensions in Testimony and StoryFile

What visitors actually experience

A visitor stands in front of a life-size (or large) display of a seated witness. They speak a question. Speech recognition + natural language processing selects a pre-recorded video clip of that person answering the closest matching question from a multi-day interview. The clip plays. The person on screen is not animated, not lip-synced to a TTS voice, not a MetaHuman. S172 S175

Shoah Foundation’s rule, stated in the public production notes: footage of the interviewee’s response is not animated, edited, or manipulated, so the file remains usable as a primary source. S175

That single constraint is the product.

Who owns what

Stephen Smith’s presence on both sides of the ledger (Shoah Foundation leadership historically, StoryFile co-founder) is why press sloppily says “StoryFile’s Dimensions in Testimony.” Accurate sentence:

StoryFile grew out of the DiT method. DiT remains a Shoah Foundation collection. StoryFile sells the generalized engine and new captures (including local Holocaust centers that were never filmed by Shoah).

Orlando’s 2026 recording of Suzanne Schneider is a clean example of the split: the center already used DiT on field trips; local stories were not in the Shoah set, so they hired StoryFile and budgeted ~$250k per subject. S171

Pipeline — Dimensions in Testimony

1. Choose a witness and get consent

Subjects are living witnesses to genocide or liberation, not reconstructed famous dead. Languages in the collection have included English, Spanish, Hebrew, German, Mandarin, Russian, and Swedish. The set includes Holocaust survivors, at least two WWII liberators, a Nanjing Massacre survivor, and a war-crimes prosecutor. S175 S177

Consent here is not a likeness checkbox. It is an elderly person agreeing to sit in a dome for days knowing the result will outlive them and answer strangers.

2. Multi-day structured interview

Typical production: five days. Same clothing every day. After each answer the sitter returns to a common rest pose so clips can be concatenated without a jump. S175

Question load:

• Eva Mozes Kor at CANDLES: 1,500+ questions. S179
• Common public figure: up to ~2,000 questions covering life before, during, and after, plus off-topic small talk so “hello” and “how are you” have somewhere to land. S175 S179
• Eva write-up in Tablet: ~30 hours of interview, then humans mapped each reply to dozens of variant questions, producing on the order of 30,000 Q–A pairings. S173

Pinchas Gutter, then 80, was the 2012 prototype: flown from Toronto to ICT’s light stage in Playa Vista. S174

Early capture (2013–2017): 116-camera dome at ICT. Traum et al. describe a hybrid of RED Epic 6K cinema cameras for the hero stereo view plus cheaper HD cameras for other angles. S178

From 2018: a mobile rig, so crews could go to aging survivors instead of flying them to Los Angeles. S175

A later offshoot films survivors on location (e.g. at Auschwitz) with a multi-camera pod for VR/geolocated playback — related, not the museum Q&A theater. S174

3. Indexing, not generation

This is retrieval.

1. Transcribe and time-code every answer.
2. Humans (and later ML assist) attach many phrasings to each clip.
3. At runtime, ASR hears the visitor, NLP picks the best clip, the player rolls video.
4. Every asked question can be logged and reviewed. Tablet describes a back-room dashboard of recognized questions vs. returned clips. S174

If the visitor asks something never recorded, the system should fail closed (a bridge line, a clarification), not invent a memory. That failure mode is the ethical load-bearing wall. Generative histobots fail the other way.

4. Display

Museum “Dimensions in Testimony Theater”: large or life-size figure, spoken Q&A, often after a short introduction to the person. UK Forever Project is a useful contrast — linear testimony first, then moderated questions — where DiT is question-led from the start. S176

Also shipped inside IWitness so classrooms can “meet” Pinchas Gutter without standing in a gallery. S231

Permanent and rotating installs include Illinois Holocaust Museum, Dallas Holocaust and Human Rights Museum, CANDLES, Nancy & David Wolf Holocaust & Humanity Center (Cincinnati), Nova Southeastern’s Weiner Center, and others. S170 S231

Pipeline — StoryFile (the productized cousin)

Same retrieval philosophy, three grades of capture.

StoryFile Life (consumer)
Webcam or laptop. Choose StoryLines (curated question packs). Record answers. Cloud stores clips. Relatives talk to the file later. Heather Maio-Smith has said a typical Life user records 250–325 answers, not 2,000. S235 Free trial advertised; public pricing has moved around (waitlist noted in 2025 commentary). S236

Creator / Conversa SaaS
No-code tools: write questions, manage streaming clips, train the matcher, publish. Aimed at businesses and developers. S237

Enterprise / museum studio
Green-screen studio, hundreds of questions, life-size playback, education wrap. This is what Orlando paid for in 2026: StoryFile crew at Green Slate Studios, ~450 questions for Suzanne Schneider, walk-by conversation on a life-size screen planned for a 2027 museum. Two local liberators also planned. S171

StoryFile’s public technology name for the matcher is Conversa. Company language: retrieval-based, not generative. National Medal of Honor Museum staff quoted to that effect. S236

Funding reported around the consumer launch: $2M (2020) then $4M (2021). S234

Cost

Published numbers, not guesses:

Why Pinchas cost ~7× Orlando: you are paying for inventing the dome, the NLP, and the display language the first time. Later captures reuse Conversa / the mobile rig.

Personnel on a DiT-grade shoot:

• Witness + family/consent counsel
• Interviewer trained in trauma-aware testimony
• Multi-cam / lighting crew (or mobile volumetric pod)
• Wrangler to keep wardrobe and rest pose consistent
• Archivists / indexers (the unglamorous majority of the budget after day five)
• NLP engineer
• Museum educator who writes the pre-roll and the “this is a recording” signage
• Ongoing reviewers of the question log

That last role never goes away. New visitors invent new questions for decades.

Market success

DiT is the rare historical-avatar project with a reason to exist that does not depend on novelty.

The last survivors are dying. Dallas’s own page notes Sam’s death in 2026. Museums that already run DiT on every field trip (Orlando’s line) will keep paying because the alternative is a silent gallery. S170 S171

Success metrics that actually matter here:

• Number of biographies in the collection (dozens, not thousands)
• Museums and campuses with a theater or IWitness access
• Languages
• Whether students treat the figure as a person they may question, not a clip they must sit through
• A 2016 Shoah evaluation called the tech a “valuable education tool” — modest language, on purpose S175

It is not a growth-stage consumer app. Hello History’s 200k downloads are a different business. DiT will never have 1,210 characters. Each new person costs a quarter million and a week of an elderly witness’s life.

StoryFile’s bet is that the method generalizes: grandparents, Medal of Honor recipients, corporate founders, local liberators. That is a real market (legacy video, HR, museums) and a different ethical load than generating Van Gogh.

What “success” is not: photorealism scores, LLM leaderboard position, or GMV in seven hours. Those metrics belong to the other half of this research.

Ethics that the pipeline encodes

These are not after-the-fact essays. They are production rules.

1. No invented testimony. If it was not said on the record, it cannot play. Contrast George AI and Bonjour Vincent, which must improvise the moment a visitor leaves the corpus.
2. Primary-source integrity. No face swap, no TTS, no emotion engine on top of the face. S175
3. Logged questions. The institution can see what the public asks and whether the matcher is drifting. S174
4. Consent of the living. Family permission shows up again in Historic Metahumans (Biran); DiT requires it of the speaker themselves.
5. Failure is visible. A missed match is honest. A fluent lie is not.

The scholarly literature around DiT (Tablet, Sussex recommendations, Princeton “ethics of the algorithm”) argues the remaining risk is selection: which clip the model thinks you asked for, and which questions were never recorded. That is a real problem. It is a smaller problem than a generative Washington discussing 2026.

When to use this method vs a histobot

Use DiT / StoryFile retrieval when:

• The person can still sit for an interview, or already did
• Legal and moral requirement is “their words”
• The venue is a museum, courtroom-adjacent education, or family archive

Use a Bonjour Vincent-style generative histobot when:

• The person is long dead
• The corpus is letters, not a voice
• A scholar will live inside the prompt and the logs
• You can label the thing as interpretation

Do not hybridize casually. Putting an LLM “in front of” survivor clips so it can paraphrase them destroys the property that made DiT fundable.

Primary sources in the bibliography

S171 S172 S173 S175 S178 S179 S232 S234 S236 S237

Full list: Citations — Historical Virtual Characters

Opposite method: Case Study — Jumbo Mana Bonjour Vincent