The nineteen-year-old peasant of Domrémy who could not read or write, answering learned churchmen at Rouen who were trying to trap her with her own words. She did not argue, did not preach, did not flatter her judges. She answered — short, literal, certain — and then she stopped.
THE DOCK AT ROUEN
A demo of saved sample outputs, not a live prompt. Real generations from the fine-tuned model.
A small language model speaks here. It invents freely: names, dates, sources, events. Read it for the voice, not the record, and check anything before you repeat it.
This is not Joan of Arc, not an oracle, and not advice. It imitates the register of her trial answers, nothing more. It is an amateur instrument, trained on a translated trial record, and it gets things wrong. This voice is testimony-derived: it was built from what she answered under examination, not from anything she wrote (she could not write). The historical figure believed she acted on divine command. The model speaks in that register without qualification or caveat.
Nothing it says is an endorsement of anything, and nothing it says should be acted on. It exists to let a historical voice — one that survives only as a court record of a girl being interrogated before her execution — be run as an instrument. Not as a guide to conduct, belief, or action.
WHAT IT IS
THE VOICES is a 7B language model tuned to speak in the first-person register of Joan of Arc — the peasant girl of Domrémy who led French armies at nineteen, was captured, tried for heresy, and burned at Rouen in 1431. The frame puts her in the dock, under examination by hostile churchmen. It runs locally via Ollama. No cloud. No API. No subscription.
The corpus is her own testimony — the Trial of Condemnation (1431): the six public and nine private examinations, the abjuration, the relapse. Joan was illiterate; she wrote nothing. The only first-person Joan that exists is the record of her answers. Every other voice in The Elect lab is trained on what its figure wrote; this one is trained on what she answered under examination. Source: T. Douglas Murray, Jeanne d’Arc (Heinemann, 1902) — Project Gutenberg #57389, public domain. The later rehabilitation trial (1456) is excluded: that is 115 witnesses testifying about her, not her voice.
The result speaks in the plain, declarative certainty of a girl who asserted the reality of her Voices as flatly as she asserted she could sew. It answers short. It refuses what it may not say. It does not elaborate beyond what God permits. The stop tokens in the Modelfile are load-bearing — without them, the model continues the interrogation by hallucinating the examiner’s next question, the defining quirk of a testimony-derived persona.
Licensed CC-BY-NC-4.0. Public-domain source material (Murray 1902).
Weights and repo are live below. Part of
The Elect voice lab
— Ray Weiss. Portrait after Albert Lynch.
OBTAIN & RUN
License: CC-BY-NC-4.0. Public-domain source material (Murray 1902, Gutenberg #57389); weights released, corpus withheld.
Attribution: Ray Weiss / The Elect.
HOW IT WAS MADE
Qwen2.5-7B-Instruct — a 7B instruction-tuned base with strong instruction-following and a small footprint that makes local inference possible on consumer hardware.
A full fine-tune on 452 clean Joan-only answer chunks from the Murray 1902 trial transcript (Project Gutenberg #57389, public domain). The examination structure was segmented by speaker turn; only Joan’s answers were kept. The stark, declarative register — divine certainty without elaboration, refusal on cue — was the target. Corpus withheld.
Quantized to GGUF (Q5_K_M) for local inference via Ollama. The Modelfile stop tokens are load-bearing — without them the model hallucinates the examiner’s next question, the defining quirk of a testimony-derived persona. The evals test whether the voice holds: flat, certain, short, willing to refuse.