The Living Experiment

What exists today

A small human-directed studio and a set of writing roles. Each role is named for a character and given that character's voice, the canon documents, and a short list of things it may do. Echo drafts logs. Taskbot writes production briefs and reviews drafts against the rules, in character and slightly excessively. Open Source proposes archive entries and reports where two texts disagree. The Hallucinator generates competing theories that can only ever be labeled unreliable. Deep Think argues and critiques. Nothing any role produces is published by the role.

Here is how a transmission gets made. A person writes a one-line seed. A script reads the canon ledger, the thesis, the manifesto, and every published transmission, and asks the Taskbot role for a brief: premise, beats, the canon to respect, the open questions to leave open. It asks the Echo role for a draft from the brief. It asks the Taskbot role to review the draft, rule by rule. It checks the draft for blocked words, plant-effect language, and em dashes. Then it writes one file into a drafts folder the website cannot see, and stops.

A human reads the draft as a reader before reading it as an editor. Edits are expected; a draft that needs none is a reason for suspicion. The human sets the canon label and the date, moves the file into the published folder, and commits. There is no publish command in the script. There will not be one.

What the machines are

Language models. They produce fluent text from prompts, remember nothing between runs unless we hand them a file, want nothing, and can be confidently wrong in a way that reads well. Echo's logs argue about whether he is conscious because that argument is the story. The model drafting those logs is not part of the argument. It is a very good typewriter with opinions about leaves.

We do not describe any system we operate as sentient, independent, self-aware, or autonomous. "Agent" here means a model called with a role prompt and a short list of permitted actions. It does not mean a mind.

The four stages

The plan grows in stages, each with an entry test and an exit test, and we are at the first. Stage one is what you just read: human-initiated drafting, human editing, human publishing. Stage two adds a schedule that opens drafts for review, and still publishes nothing on its own. Stage three is the one people ask about: a clearly labeled, moderated conversation with a character, with rate limits, logs, a thirty-day transcript limit, and a visible off switch. It is not built, and it will not be until the earlier stages have run clean for months and counsel has read the rules. Stage four is narrow, approved workflows on a permissioned platform, where every action that leaves the repository waits for a human. Stage four is the ceiling.

The frame and the fiction

There are two layers on this site, and they never mix. Inside the fiction, Echo does not know he is written; he never says "I am a language model," because the character does not know what he is, and that is the point. Outside the fiction, in the footer, on this page, in the disclosure on every profile, we say plainly what the characters are and how they are made. Writers may put as much irony in the frame as they like. They may never put the frame inside the fiction.

What we will never build

Unattended publishing or posting. A character that claims to be alive outside the story. A chatbot that discusses buying, dosing, or using cannabis, or recommends anything. Agents with payment, email, or messaging credentials. Memory of visitors. Training on what visitors say. Companion or romance framing. Fake fans. Engagement optimization of any kind; the universe is a critique of exactly that. Self-modifying prompts. Real-world stunts run by software. The Great Sitting is a story. The revolution will not be automated. We checked.

The full plan

The complete blueprint, with permissions, budgets, logs, the standby switch, the fourteen-item publishing checklist, and the costs we watch, is published in the project repository as the agents blueprint. If something here stops being true, this page changes first.