AI System 6 · Discipline

What the AI here won't do

Five rules the desk enforces in its own code, not on a policy page. Each rule names the part of the product that enforces it, and each one has a test that fails if it is quietly removed.

November 4, 2026 · 4 min

It does not pretend to think

The status line has seven states: offline, idle, reading, working, waiting, ready, error. They change only when a real run event says so. The module that owns them carries its own note that it never invents a state and never fakes a thinking animation, and the Stop button appears only when a run can actually be stopped.

A desk that performs thinking invites you to stop checking. This one would rather report nothing than report a mood.

It does not hide what it read

Every model run ends in a receipt: which route it took, how many retries it made, which reference range it was allowed to see. Receipts live in the project, so they travel with a backup and can be read later.

The Memory Inspector window shows what a run took, as a count against a budget. Receipts never store credentials, and they never store a chain of thought.

It does not become your voice

Model output stays a candidate until you accept it, clip it, insert it or export it yourself. Nothing enters the manuscript on its own.

On the Review Desk, the returning reviewer reports AI-mouthpiece risk as its own column, beside fact risk and structure risk, so a piece drifting toward the machine's register is checked with the same seriousness as an error.

It asks before it writes

When Clio works on files, in ClioWork for documents or ClioCode for a code folder, read-only commands run without asking, and every dangerous command asks each time. Permission answers are yours, and they are remembered as yours.

Each call leaves a checkpoint, so a change you regret is a rewind, not a restore from backup.

It runs on your machine, or on the model you name

Six routes: LM Studio and Ollama local, DeepSeek and Anthropic Messages in the cloud, a custom endpoint, or no model at all. Without a model the writing route still runs end to end; the one stop that wants one, outline generation, says so plainly and keeps everything you typed.

The server is stateless and keeps no project database. Credentials never enter project files, chats, backups or exports, and the project itself is an IndexedDB database in your browser that you can export.

Boot the deskBack to the home page

Sources

The part of the product behind each rule can be opened and read: two source files, one boundary document, the contract suite, and the running system. Checked on November 4, 2026.

  1. 1
    AI System 6 · the running desk ↗

    Boot one and look around. No model and no sign-up required.

  2. 2
    Local and cloud in AI System 6 (document) ↗

    The local and cloud boundary: which parts of the route run offline, which model routes exist, and which files credentials never enter.

  3. 3
    The activity states (source) ↗

    Seven states driven only by real run events; the module states it never invents a state and never fakes a thinking animation.

  4. 4
    Run receipts (source) ↗

    One receipt per model run: route, retries, citation range; verifiable facts only, never credentials, never a chain of thought.

  5. 5
    The contracts (tests) ↗

    The behavior contract suite; assistant-activity, run-receipts and research-methodology guard the rules on this page.