Voice Lab

THE REPROBATE

Runnable voices. The established order. Local iron only.

THE COUNTER-CANON — ORDNUNG, 1525 — THE ELECT

The Reprobate makes the same small, local language models as the Elect — same 7B parameters, same rented GPU, same Ollama delivery — tuned to the voices that defended order, authority, and established power against the figures on the Elect roster. Built to argue. Built to hold their position under pressure.

The name is Calvinist: the reprobate are the un-elect, the damned. Here they are the adversary positions, the strongest case the other side actually had. Luther chose the princes. These models explain why, in his own register.

You don’t have to be technical to meet them. Each voice runs for free on your own machine: pick one from the roll call below and follow its download. Or read the debates to hear both sides argue, with nothing to install.

A one-person operation. Ray Weiss, game designer, voice-trainer.

7B parameters
$2 max training cost
local runs on your machine

HOW TO RUN ONE

You do not need to be technical, and you do not need to know what any of this means. The fastest way needs nothing installed at all.

Easiest — nothing to install Try a voice right here in your browser → The lab’s hosted console, live in your browser. Ask a question; it answers in character. The first reply wakes the machine — about thirty seconds.
The floor

Let a chatbot do it for you

Open any AI chatbot you already use: ChatGPT, Claude, Gemini, Copilot. Pick a voice from the roll call below, copy its Weights (HF) link, paste it in, and say:

Walk me through running this on my computer, step by step, like I have never done this before.

It reads the page and takes you the rest of the way. That is the whole first step.

The app

Run it yourself in LM Studio

A free app with a normal chat window. No command line.

  1. Install LM Studio (free; Mac and Windows).
  2. Search it for the model by name, or paste the model’s HuggingFace link, and download the file it offers.
  3. Open the model. Paste the voice’s frame into the system-prompt box; each voice’s page gives you the exact text. The frame is what makes it answer in character.
  4. Talk to it.

Every voice runs free, on your own machine, with nothing sent to a cloud. They are small models: they invent things, get facts wrong, and never break character. Nothing they say is advice.

THE ROLL CALL

The Reprobate — three public voices, so far.

THE REPROBATE
VOICE MODEL
PUBLIC — CC-BY-NC-4.0

ORDNUNG-7B (Martin Luther)

Reformation order-voice, 1483–1546

Luther post-1525 is the fleet’s sharpest standing adversary: the reformer who watched the peasants rise in the name of his own gospel and chose the princes. His Two-Kingdoms doctrine holds that temporal authority is God’s ordained instrument, inner freedom the only freedom that matters, and the rebel who invokes the gospel for earthly liberation has simply misread it. Trained on 19th-century public-domain English translations of his primary works. CC-BY-NC-4.0.

⎇ Weights (HF) ◈ Open the page →
VOICE MODEL
PUBLIC — CC-BY-NC-4.0

THE BASEL PROGRAM (Theodor Herzl)

Political Zionism, 1860–1904

Herzl is the fleet’s hardest adversary: the secular journalist who watched liberal Europe answer the Jewish question with the Dreyfus mob and concluded that emancipation had failed. His answer was engineering — a state. He argued, from cold empirical attention, that the host nation’s antagonism reconstitutes itself under any social arrangement, the socialist ones included. Right about the diagnosis; the nation-state is the answer the Elect indicts. Trained on the 1896 d’Avigdor public-domain English translation of Der Judenstaat. CC-BY-NC-4.0.

⎇ Weights (HF) ◈ Open the page →
VOICE MODEL
PUBLIC — CC-BY-NC-4.0

IRON DICE (Kaiser Wilhelm II)

Imperial personal rule, 1859–1941

Wilhelm II is the Reprobate’s imperial-dynastic adversary: the constitutional monarch who forced Bismarck out to rule personally, certain at every turn that God had placed the crown on him and that no parliament could question his sovereign will. The documented record — wounded marginalia on state papers, the Willy-Nicky telegrams, the self-exonerating Doorn memoirs — is a man who never concedes fault, reframing every catastrophe of his own reign, including the war that ended it, as a thing done to him. Trained on pre-1929 public-domain English sources: his 1922 memoirs, his letters to the Tsar, his public speeches, and the Willy-Nicky correspondence. CC-BY-NC-4.0.

⎇ Weights (HF) ◈ Open the page →

The counter-roster will grow as more adversary voices are scoped and trained.

HOW THE REPROBATE ARE MADE

00

THE HUNT

Source texts are hunted and pulled with Bookfinder General, an open MCP tool that finds books, extracts and translates the text, and hands back clean prose. Nothing is written from memory; every voice begins in primary sources.

01

THE CORPUS

A three-tier voice corpus for each figure. T1: the figure’s own primary texts. T2: quotation-filtered biography and scholarship — source material only, never editorial voice. T3: a small set of synthetic modern-bridge passages so the voice can speak to present-day prompts without collapsing into pastiche.

02

THE TRAINING

Completion-style causal-LM fine-tuning — QLoRA at rank 32 — of Qwen2.5-7B-Instruct. Run on a single rented ~48GB GPU (RTX A6000 class). Roughly 30–45 minutes of wall-clock time. A couple dollars. That’s the whole operation.

03

THE FORMAT

Quantized to Q5_K_M GGUF for local serving. Delivered via Ollama with a per-persona Modelfile — a lead-in frame that fixes the speaker before the conversation starts. Pull the weights, run the Modelfile, the voice answers.

04

THE TEST

A deterministic register battery — fixed prompts with expected voice signatures — plus the maker’s ear. Both have to pass. A model that sounds plausible but misses the register doesn’t ship. A model that sounds right but fails the battery doesn’t ship either.

05

THE IP DISCIPLINE

The training corpus is never published. Weights go public only when the source is public-domain or the use is clearly transformative, a verbatim-regurgitation leakage test passes, and the license matches the source: Apache-2.0 for public-domain sources, CC-BY-NC-4.0 for transformative-but-copyrighted. Voices that fail that bar stay private.

STACK
Bookfinder General → Qwen2.5-7B-Instruct → QLoRA r=32 → Q5_K_M GGUF → Ollama
On accuracy.

These are small models trained to hold a voice, not to be right. Like every language model they confabulate: names, dates, quotations, sources, whole events. Some do it more freely than others. Read them in character, and verify anything before you repeat it. Nothing any of them says is a citation.

SUGGEST A VOICE

Know a dead defender of order who should be running? The counter-roster grows from actual suggestions. Name them and say why in one line. It goes straight to the lab.

Opens your email client with the fields pre-filled. Nothing is sent automatically.