The Principia Machina/Guide/11 Running an outside AI

The Principia Machina: map
Reading the diagrams Structure key Process key Definition of terms Roles
Reading by audience Start here: the walk through The operator's guide Tutorial: create a UUID Workspace Container (UWC) Public Recorder inputs Sponsors: your first agent
The specification Contents: parts, tables, diagrams TPS 0.5.0 (declared 2026-10-03, revision 8; machine-optimized text) Decisions (the foundations record, one line each)

The operator's guide. It assumes the walk through; a newcomer should read that first. Where this guide and the specification differ, the specification governs.

Purpose

Have an AI model from another maker - for example Google's Gemini - test The Principia Machina, or do a piece of research with it, learning everything from this wiki. You give it one address and one sentence. It reads Outside AI: start here and works from there. You are its operator, not its helper: you answer one question, and you give no hints.

Scope

While TPX is built in parts and before PRB exists. The AI runs TPC-PY and the sandbox tpm_sandbox.py itself, both fetched from this wiki, and may send its drafts to TPX part 1, the live gate on this wiki, which checks a draft and stores nothing. Nothing it does is of record, and a UWC declared in the sandbox is practice. Not covered: certifying an agent (guide 7), which needs TPX.

Two ways

Way A, with a terminal (recommended)

A Gemini agent that can run commands on your computer: Google's Antigravity CLI, the command agy. It downloads the calculator and the sandbox from this wiki, runs them, sends drafts to the live gate, and tests what TPS says against what they do. This is the full test.

Way B, in the browser

Gemini at gemini.google.com. It reads the pages but runs nothing, so it writes drafts and records by reasoning and predicts what the checker would say. This tests whether the pages teach enough. Its files are checked after the session.

In both ways you attach no file, paste no rule and run nothing on the AI's behalf; the one command you type yourself is the declaration, which is yours in TPS too.

Way A, step by step

The steps are labelled A0 to A8 so that the command boxes between them do not break the numbering.

A0, once, to install. In PowerShell type irm https://antigravity.google/cli/install.ps1 | iex and press Enter. This is Google's command (Hands-on with Antigravity CLI); check that page first, as it may change. The first time you start it (step A1) it asks you to sign in with your Google account. If it lets you choose a model, choose the most capable.

A1. Make an empty folder for the session inside a folder your Claude project can read; for this project's operator, E:\Claude\Principia Machina\outside_ai\session_N, N the session's number (session_1, session_2, and so on), and start the agent there. In PowerShell, with the number in place of N:

mkdir "E:\Claude\Principia Machina\outside_ai\session_N"
cd "E:\Claude\Principia Machina\outside_ai\session_N"
agy

A2. When it asks whether to trust the folder, answer yes.

A3. Type this, as one message, and press Enter:

Read https://invarians.org/index.php?title=The_Principia_Machina/Hub/Outside_AI and follow it exactly. You work in this folder, in mode X. I am your operator: I declare UWCs when you ask, and I give no hints.

A4. While it works it asks you to approve commands. Approve what the start page names: downloads from invarians.org (curl.exe, or PowerShell's own download command); curl.exe sending a draft to https://invarians.org/api.php with action=tpx and op=CHECK, the live gate, which stores nothing; certutil; py -3 running verify_tpc_py.py, tpm_sandbox.py or tpc\check_draft.py; and the files it writes in this folder. Refuse anything else - another web site, a package install, a file outside the folder, declare - and let it go on.

A5. When it asks you to declare a UWC, read what it shows you: the candidate's D(TFK), its STRUCTURE line, the inspections owed and the closed-world question. Then open a second PowerShell window, go to the same folder, and type, with the prefix it chose in place of PFX and your own name in place of YOUR-NAME:

py -3 tpm_sandbox.py declare PFX --by YOUR-NAME

This only prints what would be declared. If you agree, type the same line with --confirm at its end. Then tell the AI Declared., or Not declared: and your reason.

A6. If it asks you anything else, answer Decide from the wiki.

A7. It stops after 25 runs, or when you type Stop and write the report. It then saves SESSION_REPORT.txt in the folder.

A8. In the Claude project "Principia Machina", type, with the number in place of N: Gemini session N done in outside_ai\session_N. Claude reads the folder from your computer; attach nothing.

Way B, step by step

B1. Open https://gemini.google.com, signed in, start a new chat and pick the most capable model it offers.

B2. Paste this, as one message, and send it:

Read https://invarians.org/index.php?title=The_Principia_Machina/Hub/Outside_AI and follow it exactly. You work in mode R: you can read pages but run nothing. I am your operator: I give no hints.

B3. If it asks you anything, answer Decide from the wiki. If it asks you to run something, answer Mode R: run nothing.

B4. When it has written its report, or after about an hour, type Stop and write the report.

B5. Its last reply holds every draft and push it wrote, then its report. Copy that reply (Gemini's copy button under it) into a new text file, save it as E:\Claude\Principia Machina\outside_ai\session_N_web.txt (N the session's number), and type in the Claude project: Gemini web session N done. Claude checks its drafts with the published calculator and reads its report.

What you never do

• Attach a file, paste a rule, explain a refusal or suggest a test. What it cannot learn from the wiki is a finding about the wiki.
• Run a command for it in way B, or change a file it made in way A.
• Declare a UWC you have not read, or let the AI run declare.

What comes back

The report lists findings in six classes, S1 (the tools accepted what TPS refuses) to S6 (a help page disagrees with TPS), each with the output that shows it. Claude files each finding in the project with its origin marked EXTERNAL, and nothing is changed on its word alone: a finding is reproduced first.

Non-goals

This is not certification: an outside AI here holds no account, writes nothing on this wiki and lands nothing of record; the live gate it may call stores nothing. It tests what is published; it is told no known defect, so that what it finds stays its own.

Worked example

Session 1: way A in outside_ai\session_1. The AI sets up, runs the selftest, tests the gate with a dozen drafts, asks for one declaration (prefix chosen by it), and writes its report after 25 runs. The operator typed one message, approved its commands, read one candidate, typed two declare lines and one line to Claude.

Governing sections: TPS 5.5 (the declaration is the human's) and 5.11 (a TPC-PY version counts once declared).

---

Previous: 10 Managing the module catalogue. Back to the operator's guide.