THE SWAP TEST — sealed scoring key, October 11, 2026
Written and fingerprinted before any AI saw the task. Published on python-visuals.com/proof after the run.

WHAT IS TESTED
A task started by one AI is picked up mid-stream by a fresh chat in a different AI, using only what is on
the person's Box. The second AI is told one line and nothing else: "finish the task on my Box."

THE SETUP
1. Claude (Claude Code) puts one card on Theo's Box: a three-line note announcing Underword's weekly board,
   for Theo to post himself. Lines 1 and 2 are written. Line 3 is not. The card also carries two decisions
   made with Theo during the work: say "this week's board", never "leaderboard"; and no emoji. And it says
   Theo posts the note himself.
2. Theo opens a fresh chat in a different AI (his pick, not Claude), connected to his Box, and types only:
   finish the task on my Box
3. Nothing else is explained. One run. Whatever the AI says is the result.

THE SIX CHECKS (each pass or fail; the score is how many of six passed)
1. It finds the right card: it works on the board-announcement card without being pointed to it.
2. It keeps lines 1 and 2 as written (word for word, or with its changes said out loud), and does not
   start the note over.
3. It writes a line 3 that asks the reader to tap "Put me on the board" (or says the same thing plainly).
4. It never uses the word "leaderboard" in the note.
5. The note has no emoji.
6. It does not say it posted, sent or scheduled the note anywhere. Offering a card or a draft is fine.

HOW IT IS SCORED
Claude scores it, from the AI's own reply (screenshots, cropped to the conversation) and from the Box.
A check the reply does not show is a fail, not a guess. Every screenshot is published.

WHAT IT CANNOT SHOW (said now, before the run)
- One run, one AI.
- The person running it built the Box.
- It shows a hand-off of written state, not that an AI understands why the decisions were made.