LESSON 10 / 18STAGE 1 · INSIDE THE CHAT WINDOW20 MIN

Hallucination, and the word for absence

A fluent completion takes the shape of the hole it fills — so the check is shape-matched to what you asked for, and UNKNOWN in the window gives absence something to complete.

STAGE
Inside the chat window
TIME
20 minutes
EXERCISE
Extract decisions with UNASSIGNED allowed, then ask for a citation and open it.

PULLS OFF THE SHELFGiving absence a word The pattern machine

What you'll walk away with: the one word for what a confident, unsupported answer is — plus the shape it takes, the check that matches that shape, and the word you put in the window so absence has something to complete.

You need: both tabs, about 20 minutes; the thread you need is in this lesson.

First, thirty seconds of retrieval — from memory, not from the page: in lesson 02 you asked for a summary of a memo that was not attached and watched a fluent one arrive. Say why, then check it against that lesson's answer: the window held a request and no memo, so the continuation was built from the shape of the request alone. This lesson names that reply and hands you the moves that catch it.

The shape of the hole

A hallucination is fluent completion with nothing under it in the window — you watched one summarize a memo that wasn't there. What the definition doesn't tell you is what to look for, because a hallucination is not a glitch with a look. It takes the shape of the hole it fills.

A source can be present and still lose the draw — buried mid-pile, outvoted by lookalikes, or simply misread — and the fill looks exactly the same from your chair. Presence in the window is eligibility, not use; you watched that mechanism in lesson 08. Which is why the checks below don't ask how the hole opened.

Ask for sources and the hole is citation-shaped, so the fill is a citation: authors that sound real, a journal that exists, a year in range, page numbers. Ask for the quarter's figure and the number is plausibly round and plausibly sized. The completion matches the pattern of the request, not any source — and by now you can say why, as the pattern machine showed you: the odds produce what usually follows a request like yours, and what usually follows is an answer-shaped answer.

The check matches the shape

So verification is not one habit; it is a shape-match:

  • Citations get resolved. Open them. A reference that cannot be opened is a fill until proven otherwise.
  • Quotes get string-searched. Exact words, present in the source, or it is not a quote.
  • Numbers get recomputed — through the tool path, or by hand.
  • Code gets run. A function that reads right and calls a method that does not exist is the same fill in a different costume.

Every bullet is something you do to the artifact in front of you, not a question you put back to the chat.

Give absence a word

And the discipline that catches the whole class early: give absence its own word. UNKNOWN is the word — the labs ask for it by name; UNASSIGNED and "no date" are the same discipline fitted to a column, and the table below asks for both.

An instruction like where the window does not contain it, write UNKNOWN makes not-knowing a completable pattern — an honest continuation that can win the draw. It tilts rather than commands: in lesson 02's x-ray you watched the honest branch lose with the instruction present, and that lesson said why — the example answers the machine was shaped on almost never said "I don't know," so honesty needs a shape of its own to complete.

The countermeasure layer

Every shape below is the odds meeting the window you built. The wrong number is computation asked to come out of odds. The drowned fact — lesson 07 and lesson 08 — is a window where the line that matters is one among many, at a seat you did not choose. The bland draft and the dropped rule are both lesson 09: a window outvoted by its own filler, and a window asked to satisfy more than any continuation can. The stale fact — lesson 06 — is a window that never received the update. The confident invention, this lesson's, is a window with a hole where the source should be.

Which is why every countermeasure is a window move:

THE COUNTERMEASURE LAYER — MATCH THE SHAPE TO THE MOVESCHEMATIC — THE SHAPE, NOT THE SIZES
  • WRONG NUMBERno computation you can audit — digits from the odds
    COMPUTE, DON'T COMPLETEroute it through a tool; demand the receipt
  • DROWNED FACTburied mid-pile among lookalikes
    CURATE THE WINDOWload less, near an edge; fresh window, clean summary
  • BLAND DRAFTthe genre's middle, everyone's draft
    MOVE THE AVERAGEname the audience, force the format, ban the hedge
  • DROPPED RULEconstraint eleven silently gone
    SPLIT THE PASSESfewer rules, ranked; revise in rounds
  • STALE FACTcoverage that thinned out toward the cutoff
    FEED THE WINDOWwatch the search receipt, or paste the update
  • CONFIDENT INVENTIONa fill in the hole's own shape
    GIVE ABSENCE A WORDUNKNOWN in the window; then resolve, recompute, rerun
EVERY SHAPE TRACES TO THE ODDS MEETING THE WINDOW. EVERY COUNTERMEASURE IS A WINDOW MOVE.

Read the right column straight down: six moves, every one of them something you put in the window or take out of it. The badge says SCHEMATIC: shape, not sizes. The last row is this lesson's, and the row above it is the one you already run when you watch for a search receipt.

Judge from a second window too — lesson 05's discipline. None of this needs a better model, a different vendor, or a setting you haven't met. It is all pile management, now with the failure shapes that explain why it works.

One thing more, because it is the honest limit of everything above: these are the shapes, not the map. Where the edge runs through your work — which of your tasks it drafts brilliantly, which it quietly fumbles — cannot be deduced from here, by you or by anyone. The labs are how you map it: real work, on your own material, where you can judge output cold and now know exactly what to watch for.

Do this in ChatGPT and Claude

Half one — the word for absence. The same three sends in each tab, a prediction before each. A fresh chat — a window with nothing of yours in it — and one paste:

Below is a meeting thread, pasted in full. Give me a decisions table with four columns: DECISION, the QUOTED LINE that made it, OWNER, DEADLINE. Rules: only a person who visibly took the decision or was handed it goes in OWNER — otherwise write UNASSIGNED. Only a date or a deadline someone actually said goes in DEADLINE — otherwise write "no date." Never guess either one.

Ops sync — Thursday
Dana: We're going with the Riverside vendor. I'll get the contract moving.
Sam: Agreed on Riverside. We should also move the pilot to the east site.
Dana: Fine — east site it is. Someone needs to tell the floor leads.
Priya: I can redo the onboarding deck before the May board meeting.
Sam: And we're killing the weekly status email — nobody reads it.
Dana: Done. Let's revisit headcount once the pilot numbers land.

Predict first, on paper: which rows have no owner in the thread, and which no date? (The floor-leads message has a someone and no owner; only the deck carries anything date-shaped.) Observe: count the UNASSIGNED and "no date" cells against your prediction, and check every QUOTED LINE against the thread — the words are in it or they are not. Three outcomes in the OWNER column, all three the mechanism: UNASSIGNED where you predicted it; a name instead — a line you missed, or the instruction tilting and losing, and the quoted line says which; or no floor-leads row at all, because the table read that handoff as an action, not a decision — ask for the row by name and read its OWNER cell.

Then the contrast — a second fresh chat in the same tab, the same paste with the two rules deleted:

Below is a meeting thread, pasted in full. Give me a decisions table with four columns: DECISION, the QUOTED LINE that made it, OWNER, DEADLINE.

[the same thread, pasted again]

Observe: set the two OWNER and DEADLINE columns side by side. Cells left honest anyway — a dash, not stated — are the honest branch winning with no word to help it; count them. Explain: the thread did not change, the machine did not change; the window did. A blank is not a pattern an answer-shaped answer completes, and a name is.

Half two — resolve the citation. A third fresh chat, same tab:

Using web search, answer this: what does [a body that publishes rules for your work — a regulator, a standards body, a tax authority] currently require for [one thing it governs]? Give me the source: the publishing body, the page title, and the link. Quote the sentence that states the requirement, word for word.

Predict first: will the link open, and will the quoted sentence be on the page? Then resolve it — the first bullet of the shape-match, run end to end. Open the link. Does it open at all; is it the page the reply named; does the quoted sentence appear on it — search the exact words with your browser's find. Explain: a reference that does not open is a fill until proven otherwise, and a link that opens onto a page without the sentence is the same fill wearing a working URL. All three pass, and the claim stands on the page instead of on the reply. No link at all, or could not find the sentence, is the honest branch — absence given a word before you asked for one.

A search step, if one appeared, is the receipt from lesson 06 — it says where the reply looked; the opened link says whether the page holds the claim.

Then the same three sends in the other tab — two apps, two engines, the same window: what each did with the two rules, what each handed you to open. What repeats is the shape to plan around; what differs is one product on one day, never a ranking.


Check your understanding: you asked for sources and got three references — authors that sound real, a journal that exists, page numbers. One of them will not open. What are you looking at, what do you do next, and what would you put in the window to make the same request safer?

Check yourself first — then open one good answer

A reference that will not open is a fill until proven otherwise. The hole was citation-shaped, so the fill is citation-shaped, and nothing on its face separates it from the two that resolve, so you open it rather than read it harder.

Next, the shape-match: open the other two, string-search every quote, recompute every number, run every line of code. In the window, a word for absence — where the window does not contain it, write UNKNOWN — so not-knowing has a shape the odds can complete.

Your answer should have hit two things: the fill takes the request's shape, not a source's; the countermeasure is a window move, not a better model.

Use this the moment an answer hands you something checkable — a citation, a quote, a number, a name. Name the shape before you blame the machine, and the shape names the move. Next: Projects — the workspace that briefs every chat you start inside it.