How the Rig Works

The single most important fact about this rig is the most counterintuitive one: no model runs in your browser. There is no inference happening when you turn a knob. Every answer you read was generated ahead of time, captured verbatim, and stored. The dials only choose which recording plays.

This is a deliberate constraint, not a shortcut. It is what lets the thing be honest.

Verbatim or nothing

Every string a specimen "says" is its real, unedited output. Nothing is paraphrased, cleaned up, or hand-written to sound good. When a model trails off mid-sentence, that is where it actually stopped. When it dissolves into word salad at high temperature, that is the real salad. The rule in the codebase is blunt: to change an answer, you regenerate it with the model. You never edit it.

The reason is that the entire point depends on it. The moment you let a human touch the words, the rig stops being an instrument and becomes a puppet show.

What the dials are

There are three controls, and each one holds something constant so that the others mean something.

• Specimen - which model is answering. Eight of them: two dead era-models (one ends

in 1930, one before 1900), a refusal-ablated one, and five locally fine-tuned characters.

• Era / framing - which language the same question is asked in. talkie can be asked

about a "thinking Engine" (1930) or a "large language model" (2026); the Reader can be asked plainly or heckled by a podcast host. The substance of the question never changes. Only the words do.

• Temperature - how far the model is allowed to wander from its single most likely next

word. At 0.0 it is deterministic. At 1.5 it is mostly noise.

Everything else is pinned: the same seven questions, the same random seed (1930), the same sampler. So any single reading is reproducible, and any difference you see between two readings is caused by the one dial you moved.

How the answers are made

The answers come from a few pipelines. The dead era-models are generated in their own repositories - talkie with Apple's MLX (its 0.8 column reproduces the quotes from the original deep-dive verbatim), GPT-1900 through its own chat CLI - each sweeping temperature across the same five stops. The living characters are run locally through ollama, each called by name so its own fine-tuned system prompt and sampler apply - the generator overrides only the seed, the temperature, and the length, so a model's own voice is never flattened into a house style.

Then the output is dropped into a small typed shape, registered, and the rig renders it. That is the whole machine. The cleverness is not in the software. It is in holding everything still except the one thing you came to watch move.