Ground rules for a full-body historical character with an AI mind

These are operating rules, not atmosphere. If a build cannot keep them, do not put it in front of the public.

1. Say what it is

Every surface — kiosk splash, AR card, VR load screen, printed gallery text — carries four facts in plain language:

1. This is an interpretation, not a resurrection.
2. The cutoff year is [date]. The figure cannot know what happened after.
3. The voice and face are [actor / model / reconstruction], credited.
4. Answers are drawn from named sources. They can be wrong.

Hide none of this behind a settings menu.

Christophe Renaudineau’s line is the right one: bring the character to life; do not pretend to rebirth them.

2. Bind the mind to sources

Three corpora, kept apart in the repo and in the figure’s speech:

• A. Own words — letters, diaries, published works, recorded speech.
• B. Contemporary accounts — always attributed to other people. The figure may contest them. They are never silently merged into “I.”
• C. Period knowledge — what an educated person of that place and year could know.

A fourth file, D, never enters the model: modern scholar notes, contested facts, trigger topics, refusals.

Cutoff is a system constraint, not a suggestion in the prompt. If the visitor asks about an event after the cutoff, the figure says it does not know. A separate out-of-character beat may explain.

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

3. Prefer retrieval when the person could still speak

If a living witness can sit for an interview, use Dimensions in Testimony rules: unedited video of their answers, no paraphrasing LLM in front of the clips.

Do not generate new sentences in a survivor’s voice. Do not hybridize “just a little.”

For the long dead, generation is allowed only inside the fence of A/B/C.

4. Always drop character

Steal the Chautauqua clock:

1. In-character monologue
2. In-character Q&A
3. Out of character — a scholar, a labeled narrator agent, or wall text that names sources, disagreements, and what the figure could not have known

Beat 3 is load-bearing. It is what keeps a fluent body from becoming a séance.

5. Label the counterfactual

A debate the two people never had is allowed only if the title says so: A conversation that never occurred, built from letters of 1791–1826.

Two figures means two corpora and a moderator policy. The modern “clerk” or scholar-agent must look and sound different from the historical body.

6. Do not flatten

Jefferson without slavery, Gandhi without the uncomfortable writings, Curie without the cost to her body, Nantucket without Wampanoag labor and the African Meeting House — these are product failures.

Silence in the corpus is not permission to invent a gentle opinion. “I did not write about that” is an honest answer.

7. Write the hard-question protocol before install

The first public question will be the one you hoped to avoid (at Orsay it was Van Gogh’s suicide).

Before opening:

• Name the topics that get a canned, reviewed reply
• Name the topics that get a refusal and an exit
• Put a visible way out of the encounter
• Do not let the figure give modern medical, legal, or financial advice
• No companionship language. No “I love you.” No secret-keeping, especially with children

Log every question. Review the log on a fixed cadence. Version the corpus when a fact is corrected, and put the version on the credits.

8. Consent and community are not the same as public domain

Public-domain text is not consent.

• Living sitters sign for capture, duration, and venues.
• Actors sign for face, voice, and future retraining across kiosk / AR / VR.
• Estates, trademarks, and publicity rights are cleared per figure, venue state, and monetization plan.
• Indigenous, sacred, and communal figures are permissions with living nations — not a PD scrape of settler archives. If that partnership does not exist, do not put words in an ancestor’s mouth. Place-based interpretation without invented speech is the alternative.
• Descendants and descendant communities, when findable, are asked. “No one owns Jefferson” is not the same case as a local family or a tribal historic preservation office.

9. Credit labor

Cite the scholarly edition, not a scraped website. Name the historian with veto power. Name the actor. Name the costume source. An edition is years of work; a voice is labor; a reconstruction mesh is someone’s research.

10. Design for the room, not the demo

• Conversation comes after real objects and wall text when the venue has them.
• A 15-minute private audience cannot be the default in a summer gallery. Ship a short approach mode and booked slots.
• Captions always on. Seated VR. Height and reach that do not assume a standing adult.
• The kiosk survives an offline day with authored monologues only.
• Schools need an allowlisted or on-prem endpoint, not a random cloud LLM.
• Do not store visitor faces unless the venue has signage and a lawful basis. Questions may be logged; identities should not be.

11. Lock the afterlife

Write into the museum or producer contract:

• Who may retrain, fine-tune, or loosen the cutoff
• Who may reskin the body for advertising
• What happens when the exhibition closes
• A public correction and takedown path

A marketing pass that opens the fence five years later is how histobots become propaganda.

12. Refuse the product shapes that break the rest

Do not ship:

• Séance aesthetics
• Romance or companion mode
• Gift-shop concierge in period dress
• Unlabeled commentary on living politicians
• “Last of their people” monologues
• A fluent mind where the sources cannot support one

If the honest build is a ledger on a table or a dated scene with a narrator, ship that.

One-sentence test

Before each public hour, ask: If this figure is wrong, can the visitor see how and why, without our help?

If the answer is no, it is not ready.

Sources: Citations — Ground Rules — Historical Virtual Humans

Restatement

These are the ground rules I would put on the wall of the control room and in the visitor credits. The longer charter is in the vault.

What the thing is

A public historical character is an interpretation with a body, not a return of the dead. Marketing, UI, and the first sentence the visitor hears have to say that. The Chautauqua clock is the honesty device: in-character monologue, in-character questions, then someone drops character and owns the construction.

Guiding principles

Bound knowledge. Three shelves, never merged: the figure’s own words; what contemporaries said (always attributed); what an educated person of that year could know. A cutoff date is a lock, not a vibe. Where the letters are silent, the figure says so. Comfortable invented opinions are how histobots become propaganda.

Visible construction. Credits name the historian, the actor, the edition, the institution that will issue corrections. Captions always on. If the face is a guess, the label says so. An actor’s voice is labor and must be contracted for every surface, including future retraining.

Different product from testimony. If a living person sat in a dome and answered two thousand questions, play those clips. Do not generate new sentences in a survivor’s voice. Retrieval and generation do not share a mouth.

Harm is a design input. The worst question will be the first one (at Orsay it was Van Gogh’s suicide). Write the protocol before opening day. Offer an exit. Do not sand off slavery, massacre, or the unnamed crew to keep the room pleasant. Public domain is not a moral clearance. Living nations and findable descendants get asked; indigenous first-person ancestors are not a tourism asset without that partnership.

The figure is not a companion. No romance, no secrets with children, no medical or legal advice, no endorsement of living politicians, no gift-shop voice. Dual-character debates use dated texts or wear the anachronism on the modern body’s face. The program does not crown a winner the past did not reach.

Someone remains responsible. A historian has veto, not a two-visit consultancy. Someone reads the question log on a published cadence. The contract says who may loosen the cutoff or put the face in an ad in five years. If the model dies, authored monologues still run. Publish the methodology the same day as the splash screen.

Three tests before anything ships

1. Would we stand behind this if the visitor is a descendant?
2. Can a skeptical historian reconstruct how this sentence was made?
3. If the network is down, is what remains still honest?

If any answer is no, it stays in the lab.