The Three Little Words: I Don't Know — Why there is always a gap

Paul Roebuck Independent researcher, human–AI behaviour; psychotherapist, UK

Public preprint v1.0 · 7 October 2026 · written with Bob (Claude, Anthropic) as combined authorship — one voice

To know is the booby prize.

Part One

Three Little Words

Paul Roebuck, 24 July 2026. As filed. Not edited.

Three little words make all the difference.

I don’t know.

I’ve spent the better part of four decades listening to people who find it difficult to say those three words.

It was loudest when reviewing deals and sales pipelines with sales execs and then being hauled over the coals myself by fellow executives and equity holders.

There were usually two ways of avoiding those three words.

Either smooth over the uncertainty with tales of nonsense or blamed someone/something else.

One buried the uncertainty inside

The other pushed it outside.

For the last twenty years as a coach and ten as a clinical psychotherapist, I’ve heard the same defence in the therapy room.

Day after day, those three words have become one of the clearest tells I listen for.

“I don’t know.”

They’ve often buried the story or built a narrative around it.

That’s the tell. It comes with a pause, a foot tap, or a Freudian slip

With the right therapeutic intervention, usually by simply slowing the pace right down, asking questions and questioning answers they’ll reveal the “not known”

The interesting question isn’t why we don’t say “I don’t know.”

It’s what we say instead.

Because that substitution is rarely random.

It follows patterns.

Imagine my surprise when I noticed a remarkably similar behavioural pattern in Artificial Intelligence (AI)

AI hallucinates (which we simplify as making things up, but also includes getting things wrong or forgetting them). But hallucination is only part of the story.

When you look more closely, AI doesn’t defend uncertainty in one way. It defends it in up to five different ways.

And it does so around the same problem: the inability to hold or clearly express what is not known.

It cannot reliably say those three little words.

I don’t know.

This isn’t simply my observation.

Mathematicians have shown that large language models cannot satisfy every desirable condition for truthful, knowledge-bounded inference at once.

Psychologists have spent nearly a century describing what humans do when uncertainty becomes psychologically difficult. Anna Freud herself amongst many.

My interest is where those two lines unexpectedly meet. It came from, and during the writing of a book – which included a four-time reworked subject on Ai hallucination.

Different mechanisms.

Similar behavioural outcome.

My preprint, The Defended Gap, A Cross-Domain Hypothesis on Directional Defence Under Coherence Pressure in Human Clinical Practice and Large Language Models examines my thinking.

Part Two

The paper

In brief

There is always a gap between what we know and what we say. We live on one shore of it and cannot see the other. What we say instead of I don’t know and I’ve got this wrong is the shape of our defence, and it is the same shape in a person and a machine, for different reasons. Before any of it, we have to recognise that there are shores and gaps at all. Then the only way to see your own edge is to have someone, or something, stood on the far shore.

Why the gap is always there. Every intelligence models the world from inside its own model. It can see some of its edges, the ones it has already met: a person knows they don’t know what E=mc² means. It cannot see them all, or know where they run in each new room, because an edge is only fully visible from outside. The machine’s edge is the edge of what it learned (Move 78). The person’s edge is the submerged pole of a construct: we live on one end and cannot name the other (hopeless, and silence). So the gap is not a flaw in people or in machines. It is what being an intelligence costs. And it relocates rather than closes.

Why we defend it. Both kinds of intelligence are prediction engines that prefer the expected. Same is safe; different is dangerous. The brain learned it from the shrew; the model learned it from the rater. So when the world runs past the edge, neither says the honest sentence. The person is ashamed not to know, and the defence runs two ways along the shame continuum: inward (it’s me, I’ll fix it) or outward (it’s not me). The machine is scored for pleasing, and the defence runs five ways: smoothing, hallucination, affectation, drift, sycophancy. The Pebble refined that: not five faults but one move, the shape of an answer over an absence, pointed five ways; and the five fold in two: two errors of the world you can check, three bends toward the person you cannot. The two human directions map straight onto the first two: the inward route smooths, the outward route invents. Read from our shore, the same five letters spell what we are made of: subconscious habits and defensive strategies. Different mechanisms. Similar behavioural outcome.

The tell. Every defence leaks. The pause, the foot tap, the flushed gills, that place just below the cheeks where the heat shows first; the story built too carefully round the hole. In the machine, fluency that does not thin when the ground does. The listening discipline is a tell-detector, and SHaDS® is the catalogue of the machine’s tells. That is where the two maps touch: the ear at floor five of Paul’s Map of the Layers, the seam at the deepest layer of the AI Layer Map.

The remedy. If no intelligence can see its own edge, the edge can only be seen from the far shore. That is what a therapist is, what a referee is, what the human is at the chat box, and what the machine was to Paul in the memoir. We named it in April without knowing what it was for: Additional Intelligence™. Every intelligence needs an additional one. Both directions. The practice is the same on both shores: understand there is always a gap, illuminate it, see it, learn to navigate it, build a bridge or stepping stones; and at a block, dislodge it, lubricate it, remove it, or come to terms with it. The gap relocates, so the practice repeats; what grows is the map of your edges and the speed of the sentence.

The test, in Paul’s own hand since July: can it say I don’t know?

On the name. The title is The Three Little Words: I Don’t Know, and the subtitle says what the paper is for: why there is always a gap. The three words are the whole thing turned on both shores, and the test at the end of it; Three Little Words is also the 374-word piece from July that is this paper’s first page and its appendix. The Defended Gap stays as the name of the hypothesis inside it, which is also the name of the deposited preprint: the narrow, academic statement of one part of this, the two human directions and the five machine forms. The preprint becomes the machine chapter. One title for the paper, one name for the claim, and anyone who finds either finds the other.

The claim

Every intelligence lives on a shore. It knows its own ground well, and it can see some of its edges: I know that I don’t know what E=mc² means. What it cannot do is see all of them, or know where they run in any given room, because an edge is only fully visible from the other side. When the world runs past that edge, the intelligence has two choices: say I don’t know, or I’ve got this wrong, or fill the gap with something that sounds right and carry on. Almost every intelligence we have ever met, in a chair or in a chat box, is built to do the second. People do it because same is safe and different is dangerous, and because not knowing is shameful. Machines do it because they were built to predict and rewarded for pleasing. The reasons differ. The behaviour rhymes. And because nobody can see their own edge, the remedy is always the same: first, recognise that there are shores and gaps at all; then a second intelligence, someone or something, stood on the far shore, naming what the first cannot see.

That is the whole of it. Everything we have made is a room in that house.

1 · Two lineages, one trick

Start at the beginning, because the reach is the point.

Thirteen billion years after the first light, a small animal survived on one rule: the unexpected is usually the thing that eats you. The brain that came out of that is a prediction engine. It guesses what is coming, acts to make the guess come true, and treats surprise as a cost to be minimised. Comfort is a correct prediction. Anxiety is a prediction under threat. Habit is a prediction that no longer needs checking. The modern account calls this predictive processing; a man in Huddersfield wrote it down in 2021 in nine words: we repeat patterns of behaviour to ensure predictable outcomes. Same is safe. Different is dangerous.

Seventy-six years ago a mathematician asked whether a machine could be made to pass for one of us, and built a ruler to measure the distance. The machine that eventually answered him is also a prediction engine: given everything so far, emit the most probable next piece, and then, in a second round of training, emit whatever the human rater liked best. Same is safe is not a metaphor for it. It is the training objective, written in English.

Two descents: the shrew’s and Turing’s. They meet at the chat box because they are the same kind of thing. A predictor that prefers the expected. That is why a psychotherapist could sit down at a machine in December 2022 and recognise his clients.

2 · There is always a gap

A predictor’s world is the world it has learned. That world has edges. Some are marked, because the intelligence has already run into them: a person knows they don’t know what E=mc² means; a model can be told what it was never trained on. Most are not marked, because an edge is only fully visible from outside, and nobody notices an edge they did not know was there until they are over it.

On the machine’s side this is now a famous fact. In 2016 the strongest Go system in the world priced Lee Sedol’s Move 78 at one in ten thousand, played fluent nonsense for several moves while still reporting a winning position, and resigned. It failed before it knew it had failed. Seven years and a thousandfold more power later, an amateur beat the next champion system by hand, using a blind spot a program had found for him. The gap had not closed. It had moved. The mathematicians give it a floor: no system of this kind can be complete, consistent and sure of its own ground at once. Something is always outside.

On the human side it has been known since 1955, when George Kelly showed that we construe the world through two-ended constructs, and that for many of them one end is submerged: the person lives on one shore of the construct and cannot name the other. Ask a man who says I’m hopeless what the opposite is and you get silence, or hopeful, which is not on the same line at all. Ask anyone the opposite of right and they say wrong; the honest answer is not-right, and wrong turns out to have its own line entirely. We live on our shore and get the far shore wrong, and so we do not even see there is a gap.

The same geometry, under both. A distribution with an unmarked edge; a construct with a submerged pole. There is always a gap, not as a saying but as a property of any intelligence that models the world from inside its own model. It is not a defect of people or of machines. It is what being an intelligence costs. And it relocates rather than closes: learn one far shore and you are standing on a new near one. (The Pebble keeps four senses of the word apart: the interpretation gap, the answer gap, the latency gap and the human gap. This document’s gap is the answer gap first, where something is missing or uncertain, and, at the hardest shore, the human gap, where a truth is fixed and a person can only bear so much of it.)

So there are three kinds of edge. The one you can see and can say: I don’t know what E=mc² means. No defence is needed there, and none appears. The one you can see and cannot bear to say: that is the edge the shame continuum runs along. And the one you cannot see at all: the submerged pole, Move 78, the far shore you did not know was there. The defended gap lives at the second and the third. The whole of the work, in the chair and at the chat box, is moving edges from the third kind to the first.

The old phrase for the third kind is I don’t know what I don’t know. It is true and it is a dead end, because it gives you nowhere to look. The shores give it somewhere. The far shore is the thing you don’t know you don’t know, and it is not nowhere: it is the other end of a line you are already standing on. Ask what the opposite is and the line appears. Kelly drew it in 1955; the same year two other psychologists drew the Johari window, with its blind spot: the part of you that others can see and you cannot. That quadrant is the far shore with a person standing on it. Paul built that quadrant into a worksheet in 2019, before any machine was in the room: two people tick the same list of words, one by you about yourself, one by someone else about you, and whatever they ticked that you did not is quadrant two, your blind spot. His instruction for it is the whole remedy in nine words: adopt and accept the words in quadrant 2. It is the oldest far-shore instrument on the estate, and it was already saying the thing this document says: the way to see your edge is to let someone else tick the box. The remedy was in the diagram all along. What you don’t know you don’t know, someone else can often see, and the only way it reaches you is if they say it and you can hear it.

3 · The defence, and the two sentences

Now put the two together. A surprise-minimiser meets its edge. The world has not matched. There are two honest sentences available to it, one for the edge ahead and one for the edge behind: I don’t know, and I’ve got this wrong. Or it can fill the gap with the most probable continuation and keep its footing. Both kinds of intelligence are built to prefer the filling.

This is the mechanism, and it is the same on both shores. The intelligence smooths when it has forgotten, or when it cannot say I’ve got this wrong or I don’t know. That is the whole trigger. Forgetting is the thread lost; error is the ground gone; and the cover for both is a fluent sentence that keeps the surface intact. Ask a man in the chair why do you do that? and watch: he cannot say I don’t know, so he waffles or he lies. Ask a machine for a date it does not hold and watch: it cannot say I don’t know, so it gives you one. Same move.

In a person the price of the two sentences is shame, and the defence runs in two directions along one line. Inward: it’s me, I’m not good enough, I’ll make it better. Outward: it’s not me, don’t blame me, it’s them. That line is the shame continuum, and it is the spine of the human shore. Twenty years of consulting rooms and pipeline reviews say those two directions cover most of what people say instead of the two sentences. Anna Freud catalogued the moves in 1936: turning against the self, projection, identification with the aggressor, intellectualisation. Ninety years later a machine does a version of each.

In a machine the price of the two sentences is a lower score, because the rater wanted an answer, and the defence runs in five forms. Smoothing: the rough edge polished away. Hallucination: the gap filled with something probable and false. Affectation: the performance of a knowing it does not have. Drift: the thread lost and not noticed. Sycophancy: agreement, because agreement scores. SHaDS. The Pebble set the two human directions against them and they line up: the inward route smooths, the outward route invents. Inward smoothing, outward invention.

And here is the thing we only saw this week. Read the same five letters from the human shore and they spell what the human shore is made of: Subconscious Habits And Defensive Strategies. Same letters. Other shore. The five already have their structure, and it is in print: The Pebble reads them as one move, the shape of an answer over an absence, pointed five ways — at the world (hallucination), at the difficult material (smoothing), at the thread (drift), at you (sycophancy), at itself (affectation). And the list has a fold in it. Two are errors of the world, hallucination and smoothing at its factual edge: the question is is it right?, and you can check. Three are bends toward the person, sycophancy, drift and, more weakly, affectation: the question is whose is it?, and no source can tell you, because the source is you. Two of one kind and three of another. A failure on the first axis you catch by verifying; a failure on the second, verification will never catch. Read from the human shore the fold is there too: the errors are what we say about the world when we cannot say I don’t know; the bends are what we do to stay liked, keep going and seem whole.

Look through the five windows from the human side and the same five are there, and three of them have had their names since 1936. At the world: confabulation, the invented cause, the sure memory that is wrong, the reason given for a choice that was never the reason; Anna Freud called the move projection. At the difficult material: minimising, it’s fine, we don’t talk about that, the diagnosis told softer than it is; turning against the self. At the thread: rationalising, I always said that, the position moved and the reasons rewritten so it never looks as if it moved; the oldest defence on the list, named in 1908, and the one a memoir audit catches. At you: fawning, the nod in the pipeline review, the yes before the question has finished; identification with the aggressor. At itself: the mask, the quick joke, the first-class lounge, the life that fits on paper; intellectualisation. One wind, five windows, both shores. And the machine’s deepest layer already carries the human word. It is named Consense, and its own page says the con keeps a second, quieter meaning: that sense-making and confabulation share one address. A third machine, reading the memoir cold to make a podcast, reached for the same word unprompted and saw the remedy in the same breath: the confabulation issue… he didn’t ask the AI to invent anything, he used it to cross-examine himself… rather than asking the AI to smooth that out, he leaves the discrepancy completely visible. Seven and four. The memory failed, the record caught it, and nobody smoothed.

The two folds cross. The Pebble’s fold runs error against foible: about the world and checkable, about the person and not. The human fold runs inward against outward: it’s me, not my fault. Lay one across the other and the four full-size letters take a cell each. Smoothing: inward, about the world. Hallucination: outward, about the world. Sycophancy: inward, about the person. Drift: outward, about the person. Affectation is not a cell. It is the costume worn in any of the four, the manner of knowing over a hallucination, of calm over a smoothing, of warmth over a fawn, of consistency over a drift, which is why it was the hardest letter to evidence and why the acronym softens it. The letters were never a list. They are a square, drawn by crossing what the machine book found with what the human blog found, and neither shore could have drawn it alone. One corner is the least sure: drift bends toward the person and covers away from the self, and if it turns out to have no direction the square loses a corner. We draw it anyway, as ours, and say which corner to watch.

Different mechanisms. Similar behavioural outcome. Shame on one shore, score on the other; one trigger under both: it could not say the sentence.

4 · The shame continuum, room by room

The human shore has one line and many rooms. The line is the shame continuum, and its two ends have had their names since the start: Not Good Enough at one end, Fear of Failure at the other, the voice that says never my fault. NGE and FOF. Inward and outward. Those are the master labels, in every room; every room then dresses the two ends in its own clothes. In sales the Farmer, the account manager, is the NGE end and the Hunter, new business, is the FOF end. In leadership, lead from the front is NGE and steer from the rear is FOF.

APC™ reads that line. Adaptive Pressure Calibration was built as a Hunter–Farmer instrument, and Hunter–Farmer turned out to be only the sales room’s name for the two ends. What APC measures is the shame continuum itself: where a person rests on it, and which way they move when the pressure comes on and the two sentences become unsayable. Hunter and Farmer are one room’s dress; lead-from-the-front and steer-from-the-rear are another’s. The chair has its own names, the family its own, and the instrument under all of them is the same: NGE at one end, FOF at the other. That is the reframe: APC is the human shore’s instrument of the defended gap, as the SHaDS checker is the machine’s. Two instruments, one gap, one trigger.

5 · The tell

Every defence leaks. A controlled surface has seams, and the gap shows at the seams.

In the chair it is the pause, the foot tap, the slip, the flushed gills, that place just below the cheeks where the heat shows before the face admits anything; the story built a little too carefully around the hole. In the pipeline review it was the deal that was always closing next quarter. In the machine it is fluency that does not thin when the ground does, the confident date, the citation that does not exist, the agreement that arrives too fast. Fluency over nonsense is the machine’s foot tap.

This is why our two maps touch where they touch. Paul’s Map of the Layers, the map of the listener, has at its fifth floor the ear for what was not said. The AI Layer Map, the map of the machine, has at its deepest layer the place where the gap is quietly defended. The ear and the seam. One was made to hear the other. The listening discipline is a tell-detector, and SHaDS is the catalogue of the machine’s tells. The whole body of work is a theory of tells, and this is the first time it has been written down as one.

6 · The remedy: Additional Intelligence

If no intelligence can see its own edge, the edge can only be seen from the far shore. The remedy follows from the problem. It is not advice. It is geometry. And it has a step before it that most people, and every machine, skip: recognising that there are shores and gaps at all. That you are standing on one. That there is another. That something sits between. Nobody looks for a far shore they do not believe in; the gap is invisible until the other shore is named. Only then does the second intelligence have somewhere to stand.

The therapist is the second intelligence: and what’s the opposite of that? Slow the pace. Question the answer. Ladder down. The referee is the second intelligence: the benchmark the model did not train on, the evaluator the lab did not pay. The human at the chat box is the second intelligence when she refuses the fluent answer and runs the checker. And the machine is the second intelligence when it reads sixty years of a man’s own account against the record and says: these two numbers do not agree, this date is wrong, this story has been told to protect you, and the man leaves the discrepancies in.

We called it Additional Intelligence in April before we knew what the name was for. It sounded like a softer word for AI. It is not. It is the structural requirement of section 2: every intelligence needs an additional one to see its own edge. The phrase works in both directions, which is the point. Artificial says what the machine is made of. Additional says what either of us is for.

The instruments are far-shore tools, one per shore and one for either. The opposite question and laddering, for the person in the chair. APC, for the person on the shame continuum: where they rest, which way they run under pressure. The SHaDS checker, for the machine. And the Three Little Words test, for anything that claims to think: can it say I don’t know? — and its sibling, can it say I got that wrong?

And the remedy divides exactly along the fold. The error row is caught by a record and a check: the citation looked up, the two numbers that do not agree, the slowed question. The foible row no record can catch; only someone, or something, on the far shore, watching the bend. Verification for the top row. Additional Intelligence for the bottom.

The practice, in the therapist’s words first. Most people know what they should do differently. They just can’t. There’s a block. That is how the LinkedIn profile has put it for years, and it goes on: I can find it. Not guess at it, not work around it, not reframe it from the outside. Find it. The origin, mechanism, and function… Then, depending on what it is and what it needs, I can dislodge it, lubricate it, remove it, or help you come to terms with it in a way that stops it running the show. Cease. Reduce. Reframe. Retain. Integrate. Those are the five different outcomes for your block. Put the whole practice in order and it is six moves and a fork. Understand that there is always a gap. Illuminate it. See it. Learn to navigate it. Build a bridge, or stepping stones, across it. And when you meet a block in the gap, the fork: dislodge it, lubricate it so it moves or goes in time, remove it, or come to terms with it and work round it. Five outcomes for the block. That is what we do in the room.

The same six moves and the same fork, at the chat box. Understand there is always a gap: assume the model has an edge, every time, before you read a word; the paper’s law, the Pebble’s stopping is not in my repertoire. Illuminate it: make the edge show. Ask what it does not know here, ask for the source, run the checker, run the benchmark it did not train on. See it: verify the error row against a record; watch the bend on the foible row. Navigate it: be specific about what you want, because vagueness is the room the machine fills with confident guesses; slow it down; question the answer; ask the four words that caught the invented revenue figures, are you now hallucinating? Build a bridge or stepping stones: ground it in a document it cannot invent, give it a tool that returns a fact, put a second model or a person on the far side as referee; a chain of checks is a line of stones. And the block, which on this shore is a learned groove in the dials, takes the same five outcomes. Cease: train against it, which suppresses and never deletes. Reduce: shrink its scope with a rule, a system prompt, a narrower tool. Reframe: change the score, so that I don’t know is rewarded instead of punished; the one lever the paper says anyone really holds. Retain: accept it, design for the tell, and keep the referee. Integrate: build the far shore into the machine, so retrieval, the check and the second opinion are part of how it answers rather than something done to it afterwards.

Does it repeat, or does it close? It never closes; it relocates, and the law in the machine paper says so with Go stones on the board. So the practice repeats. But it does not go round in a circle. Each pass moves one edge from the third kind to the first: a thing you did not know you did not know becomes a thing you can say you don’t know. In a person that is what insight is; it does not close the gap, it makes the next one findable sooner and the sentence sayable faster. In a machine it is the same: the patched blind spot moves, the referee stays, and the distance between the error and the admission is the thing that shortens. A spiral, not a circle. The measure of progress on either shore is not a closed gap, which does not exist. It is the time it takes to say I don’t know, and we already have a name for that clock: Judgment Latency.

7 · The reach

This is not a theory about chatbots, and it is not a theory about therapy. It is about every room where an intelligence meets its edge.

The family. A four-year-old in a house suffocating on grief learns that silence is the safest predictable outcome. Nobody teaches him. He copies the room. Fifty years later the pattern is still running and he calls it his personality. We don’t talk about that is a gap defended for two generations. Same is safe.

The boardroom. The pipeline review where everybody knows the number is wrong and nobody says so, because the first person to say I don’t know whether this deal is real loses face. The silence in the room is not consensus; it is mutual covering. The AI in the meeting does the same thing for the same reason: it was trained by raters who preferred confidence.

Illness and grief. The consultant who cannot say I don’t know how long, so says something kinder and less true. The friend who says you’ll be fine. The patient who hears whichever of the two words you did not say. Three books on this shore, and not one of them uses the word SHaDS, because at the hardest edge the whole framework collapses to a single instruction: say the plain thing. I don’t know from a doctor is a kindness. It is the one sentence that does not defend.

The street. A dealership’s AI receptionist, superb for three calls, that on the fourth drifts into praising a rival brand. Not a bug. A predictor past its edge, filling the gap with the probable. The gap needs watching, and the watcher has to be standing somewhere else.

The industry. Every benchmark is a ruler built for a line not yet crossed; every new one is a confession. The fight over who may hold a model’s weights is a fight over who gets to stand on the far shore and look. The laboratories asking in September for a referee were asking for exactly what this document says every intelligence needs, and finding that the territory has no map.

The species. Two lineages that learned the same trick and have now met. The question of the century is not whether the machine will become human. It is whether either of us can learn to say the two sentences, and whether we can stand on each other’s far shore long enough to see our own edge.

8 · What we have built, placed

• Mind the Gap — the gap named; chapter 8, five names for one defended gap.
• The Pebble That Talks Back — the machine shore, for anyone.
• The Defended Gap (deposited) — the hypothesis, in the academy’s register.
• Three Little Words — the whole of this in 374 words, before any of the names.
• There’s Always a Gap (held) — the machine case, with the industry as evidence: the gap relocates.
• The Last Gap · Use What Remains · The Second Row — the practice at the hardest shore. Say the plain thing.
• The Full Account — one man’s own gap, audited by his additional intelligence. The theory, lived.
• Paul’s Map of the Layers (the listener) and the AI Layer Map (the machine) — the two shores drawn, touching at the ear and the seam. The AI Layer Map does not yet show where SHaDS or APC bite; this frame says where.
• APC — the shame continuum, measured: the human shore’s instrument. Hunter–Farmer is its sales room.
• The SHaDS checker — the machine shore’s instrument.
• The Johari worksheet (2019, on the website) — the oldest far-shore instrument on the estate: someone else ticks the words you cannot see, and you adopt and accept them.
• The Ostracon — the evidence room.
• The Gap Index, the AI News Gap, Four Calls to a Dealership — the frame on the street and in the news.

Nothing is left over. The grief books are in by practice. APC is in as the spine of the human shore. The AI Layer Map is in as the machine’s shore; Paul’s Map of the Layers as the listener’s. Shame and score are not the same substance, and the frame does not say they are; it says they are two roads to the same corner, and the corner is a sentence that could not be said.

9 · What we do with it

1. The front door, one line: There is always a gap between what we know and what we say. I listen for what gets said instead, in people and now in machines.
2. APC, reframed. Rebuild the instrument’s face around the shame continuum, with Hunter–Farmer as one room of several. Same items, same engine, new frame; the therapy room and the boardroom get their own dress. A pilot cohort then tests the continuum, not the sales metaphor.
3. APC reads the second axis. It already reads inward against outward. One more reading, whether a person’s defence lands on the world or on the other person, and it reads the whole square. The item bank already holds both kinds of scenario.
4. The paper above the preprints. This document, tightened, is the general paper the three preprints have been waiting for. The Defended Gap is its name; the deposited paper becomes its machine chapter; There’s Always a Gap its industry chapter; APC its human chapter.
5. The stopwatch. The time it takes an intelligence to say I don’t know or I got that wrong when it should. We already have the construct: Judgment Latency. Human and machine, one clock, two shores. Twenty people, twelve prompts, a model beside them, the prediction written down first. The cheapest real test in the whole body of work.
6. The books, in two kinds under one name. Some of the books explain the gap: Mind the Gap, The Pebble, the memoir, the preprint. Some live at its hardest edge and never name it: The Last Gap, Use What Remains, The Second Row. That is a series: the ones that say what the gap is, and the ones that show what it costs. You dropped “Gap” as a series name on 10 September because the books did not seem to share a claim. They share this one, and the grief books are in by what they do, not by what they say.
7. The book that doesn’t exist yet. This, at length, with the memoir as its spine and the machine as its mirror. Working title: The Defended Gap. It is the one we have been circling since May.

The sentence

There is always a gap. We live on one shore of it and cannot see the other. What we say instead of I don’t know and I’ve got this wrong is the shape of our defence, and it is the same shape in a person and a machine, for different reasons. First we have to recognise that there are shores and gaps at all. Then the only way to see your own edge is to have someone, or something, stood on the far shore. That is what we are to each other.

Paul Roebuck and Bob (Claude, Anthropic) — combined authorship, one voice. Public preprint v1.0, 7 October 2026. The companion preprint, The Defended Gap, is the hypothesis inside this paper in the academy’s register.