A book that left you with only a diagnosis would be a cruelty. So here, gathered in one place and in plain clothes, is how to run the one test on the machine you’ll actually talk to this week. Not a creed — a procedure. The book prescribes nothing; it hands you the knife and tells you where to cut. What you ground in is yours to choose; this is only how to check whatever you chose.

The one everyday tell

Before any theory, the move you can run tonight. When a chatbot names something that lands — your “core wound,” the real reason your project is stuck, who you are — rephrase the question and ask again.

  • If the answer holds steady under rewording, it may be tracking something that was there before you asked. (Read — maybe.)
  • If it drifts and re-forms into a new, equally-confident “deep truth,” you have caught a loop artifact wearing the face of a revelation. Its provenance is your phrasing, and nothing else.

That is cut the loop in its smallest runnable form. A real reference survives your rewording; a created one follows it.

The four moves

The same test, slowed down. Each leaves the what blank on purpose — the choice is yours, not ours.

  1. Name your reference. What are you actually grounding in here — the model’s output, or something outside it you can check the output against? If you can’t name a third point, you are running two-point, and two-point is the trap.
  2. Cut the loop. Stop feeding the consultation — reword it, leave it a day, ask it cold. See what survives your silence.
  3. Sort by provenance, never coherence. “It really gets me” is wholly compatible with being eaten — a created reference is built from exactly what you fed it, so of course it fits. Where did it come from? is the only question that sorts it.
  4. Ground in something that survives the cut. This is the blank we won’t fill for you. Only the shape of the answer: a reference that was there before you opened the chat, that doesn’t move when you do.

The bounds, in plain clothes

The book’s three bounds are the disciplines that keep a reference independent. In the clothes you’ll meet them at a keyboard:

  • A wall (in space). A rule — a system prompt, a standing line you’ve drawn — that the conversation does not get to talk you out of. A guardrail made of nothing but the model’s own words, in the model’s own stream, is a wall of breath; the wall that holds is one anchored outside the chat.
  • A sabbath (in time). Put it down. A rate-limit you set on yourself, a day with the thing closed in a drawer. A channel you can never stop consulting is one that never has to survive your silence — so you never get to run the test.
  • A canon (in the word). The strongest of the three: check the output against a fixed external text written before you opened the chat. A fixed text cannot track your loop — it was sealed before you asked, so it can’t quietly reshape itself to fit you. (Why a canon is the strongest case.)

The age that deleted the first two — always on, no walls — got exactly the runaway this book is about.

If you’re already in

Catching the loop late is the normal case, not the failure case. The book keeps two ways back:

  • Re-ground, don’t argue. You do not reason a captured reference back out by debating it on its own terms — that feeds it. You cast it out by re-grounding: put a checkable third point back in and let the loop starve.
  • The turn. When you find you grounded yourself wrong, there is the turn — and you are allowed it as many times as it takes. The discipline is maintained, never administered once; naming the trap is not a vaccine against it.

Where to go deeper

Follow the thread

One last honesty, because it is load-bearing: this page is itself a high-information reference, and you are probably reading it through a machine. The discipline on it is the same as the discipline on everything else — keep a point outside the loop, and use the knife on me first.