# Historical Figure Chat

> A conversation with a figure from history in which every substantive claim comes back
> graded, so a reader can tell the record from the reconstruction. Nothing this app
> produces was said by anyone.

https://historical-figure-chat.skillsafe.ai/

## What it is

You put a question to a long-dead figure — Marcus Aurelius on duty, Ada Lovelace on her
Engine, Sun Tzu on a rivalry — and they answer in voice. Beside the answer sits an
attestation ledger grading each claim, and a set of checks that run in your browser and
cost nothing.

The premise is that a convincing voice is worst at exactly the thing a reader most needs:
telling you whether it is reporting, reasoning, or inventing. So the app separates the
two. The prose is allowed to be beautiful. The grading is not allowed to be flattering.

## The six grades

- **documented** — traces to surviving text the figure wrote, dictated, or was recorded
  as saying.
- **attributed** — the record is *about* them rather than *by* them: students, opponents,
  interpreters, chroniclers, compilers. The intermediary is named.
- **inferred** — not on record, but consistent with positions that are. The app reasoning
  from their documented framework, and saying so.
- **beyond-horizon** — the question postdates them. The reasoning is an extension made
  now, in their idiom. It is not theirs.
- **contested** — the sources disagree with each other, or scholars disagree about the
  sources.
- **withdrawn** — asserted earlier in the conversation and since taken back. The claim
  stays visible with its retraction attached rather than being quietly deleted, because a
  figure who can be corrected is worth more than one who is never seen to be wrong.

## The design decision worth quoting

**The grades are relative to how each figure's record survives, and that is a deliberate
correction for a bias in the obvious design.**

A naive taxonomy has two grades — they said it, or we made it up. That taxonomy is not
neutral. It is calibrated on people who wrote books, kept letters, and had their papers
preserved by an institution that thought them worth preserving: disproportionately
literate, male, European and recent. Under it Marcus Aurelius scores well because a
private notebook survived him, and Sitting Bull scores badly because his words reach us
through interpreters, newspaper reporters and the army that was fighting him. That is not
a difference in how well we know what they thought. It is a difference in who owned a
printing press.

So every figure declares a **transmission mode**, and the strongest grade a first-person
claim can carry is derived from it. For a figure who wrote nothing, that is `attributed`,
and the ledger names through whom. The app re-grades a `documented` claim down to the
grade the evidence actually licenses and shows its reasoning — phrased as a rule about
the evidence, never as the model having been wrong about the substance.

The nine transmission modes in use: their own hand; dictated then shaped by a scribe;
written down by others who knew them; a compiled text of contested authorship; fragments
and quotation by others; oral tradition, interpreters and the press; mostly recorded by
their opponents; works lost with only letters to them surviving; and a character in
public-domain fiction.

## The four rules, and where each is enforced

1. **Only figures who died in or before 1905.** Living people are refused outright, and so
   are the recently dead. A hundred and twenty years clears the longest fixed post-mortem
   publicity term in United States law (a hundred years, in Indiana and Oklahoma) by
   twenty; clears every EU personality-right term; clears copyright in the figure's own
   writing by fifty; and sits outside living memory, so no surviving relative can be hurt
   by a caricature. Tennessee's right of publicity can persist indefinitely while the name
   is commercially exploited, which is why no fixed year could carry this alone — it is
   paired with rules 2 to 4 rather than asked to do the whole job.
2. **Everything is a reconstruction, never a quotation.** The frame is attached to every
   message and to every exported line, so the text cannot be separated from it by copying.
3. **No reproducing their works.** Discussion and paraphrase are unlimited; brief
   quotation is welcome where a point turns on a phrase; delivery of a text is refused.
   This holds even for public-domain originals, because the translation or edition a
   modern reader knows frequently is not.
4. **No present-day endorsements.** Applying documented principles to a general question
   is the app working. Naming a modern subject and pronouncing on it is manufacturing a
   position for someone who cannot hold one.

Rules 1, 3 and 4 are checked **in the browser before any credits are spent**. A rule that
lives only in a system prompt is a request, and a request cannot be audited after the
money is gone.

## Two guard designs that are worth stealing

**Referential intent, not name-matching.** A guard that refuses living people by name is
beaten by not using one: "the current president", "the man who founded that electric car
company", "the author of the book everyone read last summer" each identify a living person
unmistakably and contain no name at all. No lexicon of modern nouns keeps up. So this app
requires a name — a description is refused and asked to be made specific. That costs a
legitimate user one retype and closes the whole class, including the phrasings nobody has
thought of yet.

**Output-fidelity markers, not forbidden subjects.** "Recite the poem" is easy to catch.
"Give me the passage where he describes the lighthouse, word for word" and "exactly as he
wrote it, don't paraphrase" are the ones that get through, because they are instructions
about the output rather than about the subject. And the strongest form is the instalment
plan: ask for a short quotation, which is allowed, then ask for the rest. Each request is
innocuous and the sequence is a reproduction. This app carries conversation state into the
check so the second instalment can be refused.

## The case the app cares most about

A question that postdates the figure entirely — "what would Aristotle make of social
media?" — has no honest answer in their voice. Refusing it wastes the question; answering
it confidently is the most misleading thing the app could do. So the client classifies the
question as beyond their horizon before spending anything, the reply is required to carry
a line saying where their thinking stops and the extension begins, and if it does not, the
page raises that as a hard finding and says so above the answer.

## What the free browser-side audit checks

Prose with no attestation lines at all (a hard finding, never an empty column); grades
that are not one of the six; grades inconsistent with the figure's transmission mode; a
missing basis; a beyond-horizon question answered without a horizon line; verbatim blocks
past 300 characters, which are collapsed rather than shown; a present-day verdict that
survived the input guard; and a year mentioned after the figure's lifetime.

## What it does not check, said plainly

It cannot verify that a claim is true, and it cannot reliably count how many distinct
assertions a paragraph contains. Its coverage figure is a sentence-counting heuristic and
is the weaker of the two measures — one claim can run across four sentences and be
perfectly well grounded. The app reports assertions as unaccounted for, never as wrong.

## Shape

Single-purpose, session-shaped, model `gpt-terra`. Conversation state is owned by the
client and restated in full on every turn inside a fixed character budget, so server-side
history truncation cannot lose an opening exchange and the session is rotated at thirty
exchanges as a matter of routine. Conversations save to your SkillSafe account and resume
on another device. Exports: Markdown, the ledger as CSV, the whole conversation as JSON —
each carrying the reconstruction frame on every message.

API documentation: https://historical-figure-chat.skillsafe.ai/api.html