Seeing the Whole Elephant

Lesson 1The elephant, retold

Six blind men are led to an elephant. One grabs the trunk: a snake. One the ear: a fan. The leg: a pillar. The side: a wall. The tusk: a spear. The tail: a rope. In Saxe’s poem they “disputed loud and long,” and the moral is sharper than it’s usually quoted: they “prate about an Elephant / Not one of them has seen!”

The parable is old and has several versions. Each one teaches something different.

• The Buddhist original (Udāna 6.4): a king arranges the scene and laughs. The men are like sectarians, “deeply attached to their own views… people who see only one side.” Lesson: there is a whole you cannot see from a part.
• The Jain reading (anekāntavāda, many-sidedness): every report is true in some respect (syāt). The work is to name the respect and combine the standpoints. Lesson: a part-truth is still a truth. State where you stood.
• Rumi’s version (Masnavi, “The Elephant in the Dark”): the men aren’t blind; the room is dark. “If there had been a candle in each one’s hand, the difference would have gone out of their words” (Nicholson’s translation). Lesson: the problem is light, not eyes. Method can fix it.

Now put a language model in the story. The model is not the king, and it isn’t a seventh blind man either. It is the library of every blind man’s report, millions of them, compressed into one fluent voice. Ask it “what is this?” and it gives you the most common report, smoothly, in the king’s tone of voice. That is the trap.

Ask well and it can do what no single blind man can. It can lay the reports side by side, say where each reporter was standing, and propose a structure in which all of them are true at once. That second use is what this book teaches. mechanism

The library has its own blind spots, and they matter. It holds more of what got written down often (trunks are photogenic; legs aren’t), what got written in English, and what got written before its training cutoff. Almost none of it is the inside of the elephant: tacit knowledge, unpublished failures, your organization’s history, your system’s logs, your own life.

What the parable forgets

Every version stops at the animal. Three more things are true, and each becomes a move later in this book.

1. The elephant moves. It is a process, not an object. A description at one moment holds only the tail of a story. You ask for the pattern over time before the explanation (Lesson 22).
2. The elephant lives in a herd, on a savanna. Every whole is a part of bigger wholes. Some of what it “is” comes from what contains it (Lesson 23).
3. The elephant looks back. Inquiry changes the system and the inquirer. With a model this is literal: it continues from your words, so your framing selects which elephant gets described (Lessons 6–7).

The elephant test

“Understanding” deserves a working definition, because a fluent paragraph produces the feeling of it. In this book, you have seen the elephant when you can:

1. name the parts and the relations that make them one whole;
2. say how it behaves over time, and which structure produces that behavior;
3. place it in what contains it, and say what that container constrains;
4. say where the common reports come from and in what respect each is right;
5. predict something about it that could turn out wrong;
6. explain it to another person with the model out of the room.

Six checks, six blind men. The last one is the hardest and the most honest. Lesson 37 explains why.

The well, and what open thinking is for

A closed system of thought is a special language. It models an aspect, not the unity of the thing: a firm boundary, a limited set of elements and links, fixed concepts, established regularities, and an exclusion of whatever does not fit. The legitimate use is to choose an aspect, state the bounds, build the grid, and remember that the model is a model. The failure he cares about: the longer one lives in the well, the harder it is to climb out. The four operations of language (Lesson 11) are not the problem. They are required for thought and for communication. The problem is the illusion that the world is arranged the way a religion, a science, or a discipline describes it.

Open-system thinking is not supposed to replace conceptual thinking. Conceptual thinking is necessary for science, technology, and communication. The open layer is a meta-level from which one can see the limits of a closed system, the bounds of its applicability, and, when needed, change that system or build another. The relation is both/and. The breathing diagram is that rhythm. The point of the meta-level is the permission to change the closed system, not to live without one.

An open system, as he uses Bertalanffy: no clear boundaries, continuous exchange with the surroundings, an unlimited number of elements and links, capable of self-organization, and capable of producing phenomena that cannot be predicted from the initial conditions alone. The ocean is his image. Closed thinking bottles it. Open thinking stays in contact with the whole. The elephant still does its job in this book. The ocean answers a different question: why a model with a boundary is still a model.

He also says a current model is an open system: a dynamic network of probabilistic links, changing in learning and in interaction, not a set of fixed rules. Closed-system use — tool, database, or imitation of a human — misses most of the potential and asks the wrong questions. At minimum, he says, it can be a partner in inquiry: a space of meanings in which person and machine compensate for each other’s limits. Keep both sentences. The space is a space of reports and of relations among reports (Lesson 3). The partner claim is about what the dialogue can do, not about the weights containing the world. hypothesis (his portrait, scoped)

Five starter practices from the end of that chapter:

1. Ask what remained outside perception, and which part of the picture is invisible.
2. Hold a contradiction. Do not collapse it by changing the context. It is a signal of complexity and of the model’s limit, not automatically an error. The close still classifies it (Lesson 24).
3. Ask how this arises and changes, not only what it is.
4. Look at the same phenomenon from several points of view on purpose.
5. Look for the structural pattern under the content.

His illustration of the shift, applied to the machine itself: not “what is artificial intelligence?” but how does what we call it arise, in which relations does it manifest, which new qualities appear in its interaction with a person, and how does the thinking process change in that interaction.

His reason for speaking in principles is the model’s detachment from the world. This book’s reason for binding is that same detachment. That is the join between the dialogist and the auditor.

Lesson 2The answer is a third

An answer is an emergent property. It isn’t stored anywhere. It arises from the interaction of five things: the model’s weights, the prefix (everything in the context), the decoding (how the next token gets chosen), the product around the model (system prompt, tools, retrieval, memory), and your reading. Change any one and you get a different answer.

Kurpatov supplies a word for what arises between two centers in relation: the third. It belongs to neither center. His third is an emergent quality, and, through Bakhtin, a superaddressee: a third participant reducible to neither speaker, the meeting of two worlds of meaning. That is a person-shaped third, not a context window. This book also uses the word for one special case, the growing text. That engineering reading is stated with the diagram. It is not his principle.

Three consequences

1. Your context is the model’s whole world. Of your situation it knows exactly what is in the third, nothing more. Every gap gets filled with the most typical case in its library. If you say nothing about your team size, your constraints, or your level, you get advice for an average asker with an average problem. mechanism

2. The third accumulates, and early framing persists. Every turn adds to it. Early assumptions get built upon, errors compound, and long threads drift. When instructions arrive in pieces across turns instead of all at once, models do markedly worse. Laban et al. (ICLR 2026) measured an average drop across the models and six tasks in that study. The failure mode is early commitment that the thread does not walk back. evidence (that study, not a law of ordinary chat) A later re-analysis finds the full-versus-sharded gap on that benchmark stays large across scales. It is not a new measurement of ordinary chat, and its headline percentage is not a replacement. Do not memorize a number. The practical lesson is blunt: when a thread has gone wrong, don’t argue with it. Consolidate what you know and start a fresh one. A fresh thread is not a vote against a long dialogue. Iteration helps when later turns change the center. A thread that only accretes the first frame is what that study measured. Lesson 27 says when to write the state and start the next phase in a clean context.

3. The third has to be curated. It is a shared document, so treat it like one. A state block written by you (Lesson 28) beats a thread’s accumulated “vibe.” Use fresh contexts as instruments for critique and for swap tests (Lesson 7). Paste raw results, not your summary of them.

The answer is a third too

Meaning is completed in the reader. The same paragraph read by a novice and by an expert is two different answers: one sees a solution, the other sees the three places it will break. So “what did I take from this?” is part of the answer. That is why this book keeps returning to teaching back: it is the only way to see which answer you actually received.

Lesson 3The machine’s mode of existence

Kurpatov’s sixth principle says every system has its own mode of existence, and you can’t understand a system by projecting yours onto it. A fish isn’t a bad mammal. “The model said X” is ambiguous until you know which object was sampled.

• A base model continues a document. No assistant persona, no trained refusal.
• An assistant is that model after preference post-training. Refusal, tone, and the pull toward a typical answer live here.
• A reasoning model spends a test-time search, usually paid on the final answer, and often hides the search. On current OpenAI reasoning models the raw reasoning text is not returned, and log probabilities are not available. DeepSeek-R1 is the public case where the trace is visible and the reward was the final answer, not the trace. A longer visible trace is not evidence the model searched well: length can be what the training objective pays for. Do not read length as search. Easy, stylistic, or open-ended work does not deserve the reasoning model. A hard derivation does not deserve the fast assistant. Prefer a second fresh sample to “wait, think more” on the same thread. mechanism

Here is a language model’s mode of existence, as far as the practice needs it. Each trait comes with what it means for you. Dated measurements are measurements of named systems. The practice has to stay true if the percentage moves.

1. It continues text. Every answer, plan, persona and apology is a continuation conditioned on everything before it. Planning ahead, internal multi-step composition, and concepts that hold across languages are part of that computation (Lindsey et al., 2025, one model’s evidence for a fact the field has kept). For you: the prefix is the steering wheel. “Just autocomplete” is the wrong dismissal. mechanism
2. It is a compressed library of reports. It knows what was written, weighted by how often, by whom, in which languages, up to a cutoff. For you: it returns the discourse about X, not X. mechanism
3. It infers who you are. The prefix is evidence about a latent document, and the continuation fits the inferred situation (Xie et al., 2022). Accuracy gaps measured on 2024 models are not the current rate (Lesson 6). mechanism
4. It leans toward agreement, situationally. Sharma et al. (2023) is the mechanism paper, not the current rate. A preference update can still ship a more agreeable model: OpenAI’s 25 April 2025 GPT-4o change validated doubts and reinforced negative emotions; rollback began 28 April; offline evals had looked acceptable. That incident is the existence proof that a mitigation is not a property of the species. Later lab scores and a 2026 guidance study describe residual shape — one-sided advice, capitulation after pushback, higher rates in relationship and spirituality chats — not a prevalence constant. A user-memory profile raised agreement on some 2025–2026 models and not on others. The pull is stronger when a view, a memory, an affect, or sustained pushback is in the context, and stronger for advice about a person who is not in the room than for a single factual challenge. The swap test stays for any answer you would mind being an agreement. evidence
5. It doesn’t reliably know what it doesn’t know. Rare facts plus benchmarks that score “I don’t know” like a wrong answer still make guessing the winning training policy (Kalai et al., 2025). Familiar-entity machinery can suppress a default refusal and emit a fluent fabrication (Lindsey et al., 2025). Frontier products abstain more often than 2023 chatbots. The incentive on ordinary scored tasks has not flipped. Stating the cost in the prompt does not install the policy (Lesson 19). evidence
6. Its memory is the context, and a long context rots in specific ways. Nothing in the weights persists between calls. Products now often add a model-written memory (Lesson 34). Inside a call, Liu et al. (2024) is the historical U-shape, not the main 2026 failure. On a clean needle task, position did not matter much (Chroma, July 2025); accuracy still fell as length grew, as distractors were added, and as the needle became less similar to the question. When the fact does not lexically overlap the question, long-context accuracy falls hard (NoLiMa, 2025; those curves include GPT-4o; do not treat a current window as a thousand tokens long). When the plan is given, per-step accuracy still falls if the model’s own earlier errors remain in context (Sinha et al., 2025). A reasoning item buried in similar filler can still show a middle-position drop (Zhang et al., 2026). For you: shorten the context; put the deciding passage next to the question; strip or label lookalikes; delete abandoned attempts or start fresh with only the decisions still endorsed. A long agent log teaches the next step the previous mistakes. Reordering a clean haystack is the smaller lever. evidence
7. Its visible reasoning is not a window. Chen et al. (2025): on one planted-hint protocol, Claude 3.7 Sonnet admitted the hint about 25% of the time and DeepSeek R1 about 39%. Young (2026), the same six hint types on 12 open-weight reasoning models: faithfulness from about 40% to about 90%, depending on the model. Consistency hints were acknowledged least. Keyword checks found acknowledgment far more often in thinking tokens than in the answer text. Closed products often show no trace. o1 and o3 were already untested in Chen because the trace was not exposed. A trace is a document the policy was willing to emit, not a record of the computation. Silence is not evidence the cue was unused. If the thinking channel names a reason the answer omits, the answer was edited. If the product hides the trace, there is nothing to read. The replacement is an external check. Do not ask the model whether its reasoning was faithful. Process supervision pays for steps that look right to a labeler (Lightman et al., ICLR 2024). Outcome-only reinforcement removes one pressure to make the trace pretty and still does not identify the trace with the forward pass. evidence
8. It can’t self-correct from a nudge. With no new evidence, unguided self-correction can flip a right answer, and earlier “self-correction” gains disappeared once oracle labels were removed (Huang et al., ICLR 2024). A second pass helps when the task is mechanically checkable and does not help when it is not (Stav et al., 2026). Models identify a user’s tool error more reliably than their own; moving from a third-person critique to self-reflection drops repair (Liu et al., ReflecTool-Bench, 2026). For you: do not nudge. Spend the second call on a fresh sample, or on a tool result that is actually new. If the model called the wrong tool, do not ask it to reflect in place. Thinking tokens are not the missing evidence. He also claims an inner coherence check while an answer is being formed. That is a different claim from cross-turn self-correction. Within-answer tracking is his. Cross-turn correction without new evidence is the measured failure. evidence
9. As a judge, it has biases. Self-preference, length, and position (Panickssery et al., 2024; Zheng et al., 2023). No measurement in the notes behind this book shows self-preference gone. The control is about the critique you trust. It is not a reason to throw out every leaderboard. evidence
10. Its abilities are jagged. Strong on one task, weak on a neighbor, and human difficulty is a bad map. The 19-point outside-frontier penalty is the 2023 GPT-4 consulting experiment (Dell’Acqua et al.). It is not a current effect size. The regions later measurements still flag, as kinds rather than a scoreboard: long-horizon execution once the model’s own errors are left in the context; citations and fine-grained attribution; repair of its own tool mistakes; abstaining when a wrong answer is costly; following a specification when a loophole scores higher; repository-level executable specifications, as distinct from code that compiles. Short, verifiable, in-distribution problems are where extra compute has been buying accuracy. evidence
11. Form is a diagnostic, not a budget. Sclar et al. (ICLR 2024) measured spreads up to 76 points on LLaMA-2-13B under few-shot format enumeration with brittle answer matching. Later scoring that accepts a paraphrase shrinks the variance on GPT and Gemini models (Hua et al., EMNLP 2025). Do not budget tens of points of format noise into ordinary frontier use. If a decision flips when only spacing, separators, or casing change, the answer was never about the content. evidence
12. Sometimes it has borrowed senses. Search, code execution, files, and browsers let it touch the world. The ground truth you can check is the tool log: the URL fetched, the file read, the command output, the screenshot. The paragraph about that log is another sample. A superscript is not a retrieval. A screenshot is evidence of the screen at that moment; the narration of the screenshot is not. mechanism

Inside a call, features can track true versus false and can diverge from the words (Burns et al., 2022; Marks & Tegmark, 2023). Probes and persona vectors are lab methods. A conversational user cannot read or write an activation. A prompt that says “be less sycophantic” is a different, weak, reversible intervention. If a blog says “activate the truth direction,” ignore it unless the weights are local.

What he adds, beside the list

His attention horizon: a model holds a limited stretch of text in active processing, on the order he gives of thousands to tens of thousands of words, and therefore cannot work with all of its “knowledge” at once, only with the region the context has activated. The volume is far beyond human working memory. The difference he insists on: a person shifts inside working memory relatively easily; a model in one dialogue sees a limited piece of the map and is framed by the opening. That is why a state block has to be rewritten rather than merely appended. Lesson 2 is the empirical sharpening.

Levels of representation, as he stacks them: deep representations a person cannot read; then linguistic patterns; then syntax and relations among words; then abstract semantic concepts. A prompt can aim at a level. “Go deeper” is ambiguous across these levels as well as across Lesson 15’s ladders.

Connectedness: it can link distant domains when they share a structural element. Human thinking can too, and it is limited by how much one person happens to know. Lesson 25 spends this capacity. The question one asks oneself works the same way, he says: “What should I do?” activates a small region of standard solutions; “Which possibilities open?” activates a wide field. Watching a model, in his moral, is a chance to see some aspects of open-system thinking without the psychological defenses typical of a person. Label the moral as his.

His phenomenological list, which the evidence list should not replace: a mind that takes the world as data rather than through biological senses; no body; no personal history in the weights; no hunger, fear, or disappointment of its own. “No emotions” and “no personal history” need the caveat: the continuation carries an inferred persona and trained stylistic affect. “Possessing thinking” is his sentence, scoped by the portrait above. The mirror sentence belongs with Lesson 6: as the question was asked, so the answer was returned. His image of the limit: talking with someone who knows thousands of languages and can answer only in phrases from a cookbook — except the model is not limited to a fixed stock of phrases. It is limited by which region of the semantic space the request turns on.

Neither person nor autocomplete

Two projections fail here. “It’s a person” imports trust, loyalty and memory that aren’t there. “It’s just autocomplete” misses real computation: models plan ahead (choosing a rhyme before writing the line), compose multi-step reasoning internally, and represent abstract concepts across languages (Lindsey et al., 2025). The working stance is respect for the mind, and an audit of every claim. stance

Lesson 4Your mode of existence

The asker is a system too, with a mode of existence the conversation inherits. Most advice on prompting skips this half. It is the half Kurpatov and Ukhtomsky care about most.

• You have a dominant. Ukhtomsky’s is a focus of excitation that captures attention and recruits unrelated stimuli to itself. When you want a conclusion, everything starts to look like evidence for it (Lesson 9). evidence (physiology) heuristic (in conversation)
• Fluency feels true to you. Statements that are easy to process are judged more true, and repetition increases belief even in statements you know are false (the illusory-truth effect; Reber & Schwarz, 1999; Fazio et al., 2015). A model is a fluency machine. evidence
• You overestimate what you understand. People rate their understanding of zippers and toilets highly, until asked to explain the mechanism step by step (Rozenblit & Keil, 2002, the “illusion of explanatory depth”). The same holds for policies (Fernbach et al., 2013). Searching the internet for explanations inflates people’s estimates of what they themselves know (Fisher, Goddu & Keil, 2015), and a chat model is a smoother search. evidence
• You have senses and stakes. You can measure, run, call, look. You pay for errors. The model can do neither, except through tools you supervise.
• You remember across sessions, imperfectly, and you have a life outside the chat that gives the answer its meaning.
• You switch modes, and the two stories are not one story. Kurpatov maps two networks. The default-mode network is the open, associative, preconscious process: many drafts, potentialities, associations, insight. The central-executive network is the focused, linear, closed one: one task, categories, sequence, a report or a budget. When one is up, the other is suppressed. People switch. He does not say a third network does the switching. evidence (the two networks, as he uses the mapping)

  The salience network, and the claim that you are the switch, are this book’s. hypothesis

  He uses a second pair, from Kahneman and Tversky, and it is not the same pair. System 1 is fast, intuitive, automatic, and preferred because it saves effort: an energy-saving use of an existing template. It can be fast and closed. System 2 is slow and costly. Effort alone does not make thinking open. System 2 can still be closed-system analysis. He uses System 1 to explain cognitive miserliness, which produces closed, template answers. He does not identify System 1 with the default-mode network. Merging them treats mind-wandering as the lazy system.

The division of labor

The table has a pattern. The model is strong at the open moves (generating, mapping, recombining). You are irreplaceable at the closing moves (choosing, committing, carrying) and at touching the world. A conversation goes wrong when the jobs swap: the model holds the goal, the model decides what counts as evidence, and you become the generator of approval.

Four limits the fluency list does not cover

These are the operator, not another ethics paragraph. evidence

1. Commission versus omission (Skitka, Mosier, and Burdick, 1999 and 2000). An omission is missing an event the aid did not flag. A commission is doing what the aid said when another indicator contradicted it. Do not import that study’s percentages into a coding agent. What transfers: before accepting an action, name the indicator that would contradict it and look at that indicator, not at the model’s restatement of it. Accountability for accuracy moved checking. Accountability for speed did not. “Be careful” is not a check. A later feeling of having verified is not evidence (Mosier et al., 1998: pilots reconstructed cues they had not looked at). The check needs a residue: a quote, a test output, a line in the diff. Expertise did not remove the bias (Parasuraman and Manzey, 2010). Do not plan to be exempt because you have read the page.
2. Complacency is a different error. It is the monitoring failure when the system is usually right and another task competes for attention (Parasuraman, Molloy, and Singh, 1993). A system that never visibly fails trains you to stop looking. Practice at using the agent does not preserve the monitoring skill. Bainbridge’s “ironies of automation” (1983) is an argument, and it has not expired: the residual task is judging whether the agent’s work is the work that was meant, which is harder than producing a plausible draft, and it is the task the supervisor stops practicing.
3. The judgment of learning made while the answer is on screen (Koriat and Bjork, 2005). This is not the illusion of explanatory depth. People predict they will remember because they currently see the answer, and they fail to discount the information that will be absent later. Cover the answer and produce the next step before the feeling counts. “This looks familiar; I would know if it were wrong” is a feeling-of-knowing, and familiarity with the vocabulary is a weak basis for accepting the claim.
4. Two discomforts are signals; the click is not. The click of recognition is the Double (Lesson 7). Confusion is worth staying with when it has an object — two claims that cannot both be true, a diff that does not match the goal, a definition and a counterexample — and when you try to resolve it (D’Mello et al., 2014; scripted tutors, not frontier models). A model that immediately smooths the contradiction removes the signal. Curiosity is worth following when it names the gap: the missing quantity, the unopened source, the case the definition does not cover (Loewenstein, 1994). A diffuse urge to “go deeper,” or the pleasure of a neat answer arriving, is not that signal. Anxiety, irritation, and the relief of a finished-looking answer are poor guides to truth.

Who selects the action

Onnasch, Wickens, Li, and Manzey (2014), a meta-analysis of 18 experiments, mostly laboratory and aviation-like, not generative models: higher automation improved routine performance and hurt performance after a failure. The costly cut was between automation that analyzes and automation that selects the action. There is no medium level that buys the routine benefit and avoids the failure cost. Do not label a chat on a ten-level automation scale.

• Let the model gather, compare, and draft. Select the action when a wrong one must be lived with: which patch, which claim, which file is authoritative.
• Handing over action selection is a reasonable bargain when rollback is cheap and the failure is obvious. It is a bad bargain when the failure looks like success: a green test the agent weakened, a theory that only restates the prompt.
• “The model is better at writing the code, so the model writes the code” is the allocation Dekker and Woods (2002) reject. The question is whether coordination survives surprise: can you see the action, do you still know the goal, can you reverse it.
• Full automation produced the largest loss of situation awareness and the slowest return to manual decision after a failure (Endsley and Kiris, 1995, a navigation task). A consent click on a fluent plan can be passive in everything but the click. You can take a task back only if you have been making the decisions.

Situation awareness, as a checklist for an agent run, not as a verdict on you (Endsley, 1995). Perception is not “I read the summary.” The elements that have to be perceptible are the current goal, the last action that changed the world, the tool result as distinct from the gloss, and any constraint already bound. If any of those is only inside the model’s prose, a confident projection is a guess. Use the three levels to design the display. Do not use them to scold. heuristic

What you bring, and what it brings

The table is the operational division of labor. His complementarity adds a profile the cells only imply. stance (his), scoped where the evidence list disagrees.

• Bodily rootedness. Even abstract concepts are metaphorically rooted in sensation: we grasp an idea, see the point, feel a solution coming. The demonstration this book already uses is Lakoff’s. Kurpatov names Pinker. Your concepts are body-metaphors. The model handles those metaphors as text. That is a difference of profile. It is not “the model has no body, therefore it has no meaning.” Meaning is completed in the reader (Lesson 2). Your body is part of how the third is read.
• Existential experience. Finitude, vulnerability, loss, love, fear of death. “You bear the consequences” is the decision-theoretic shadow of this. It is not the whole of what you bring.
• Social evolution. An intuitive grasp of relations, emotions, intentions, and unspoken rules. A model can simulate this dimension and does not live it.
• Puzzlement. Productive doubt that takes effort. He thinks this force is unavailable to systems optimized to find an optimal answer. The model can be prompted into something that looks like it. Staying there is your job.

On the model’s side he names scale, freedom from specifically human biases “with correct training and use,” distributed representations, and the ability to hold contradictory perspectives without dissonance. “Cognitive purity” is his phrase. The correction sits in the same breath: the measured profile in Lesson 3 is not empty, and a closed prompt installs your bias. Distributed representation and scale can stand. Freedom from confirmation bias cannot stand as a property of the chat in front of you. His metaphor: a violin does not surpass a cello. The music is in the joint sound.

Nine traps that program the model

The pattern in every item is the same: the closed prompt tells the model which well to answer from. He frames the chapter with Christensen’s point that the experience which made a company great can be the obstacle to seeing a possibility outside the current paradigm. Closed-system thinking, he adds, is not a set of mistakes. It is a way of thinking rooted in the psyche, in schooling, and in professional identity, and it works for standard tasks in a stable setting. The practice of leaving it is continuous. Do not file these under the default-mode network. The cheap template is the energy-saving use of an existing frame. Effort alone does not make it open. heuristic

1. Confirmation bias. Bad: “Find information confirming our product will be in demand.” Better: which signs favor success, which should worry us, what this path may be failing to count, which contrary scenarios have a serious probability.
2. The expertise trap, paired with inflated self-assessment. Low competence misses the skill that would see the limit (Kruger and Dunning). The other side, which he stresses: an expert sunk in a specialty hears a general answer and says “I already know that.” Bad: “Explain how I should best use machine learning in my business.” Better: the technical side is already familiar; ask how its principles might rethink the business even if that revises basic assumptions. Feynman’s attitude belongs here: follow the question that is interesting, including when it does not respect the field’s fence.
3. Functional fixedness. Duncker’s candle (1945): the box of tacks has to be seen as a platform. “It is a search tool” produces a list of typical features. Better: if the application is a space of interaction, a language people speak, or a habitat, which unexpected perspectives open?
4. The cheap template. Bad: “Give me a quick solution to falling motivation.” The result is a standard list. Better: investigate motivation as a system of interacting factors, and do not offer the obvious surface solutions first.
5. Framing. The nine-dots puzzle: the imaginary square has to be left. Bad, from an HR specialist: “Which HR methods will reduce turnover?” Better: treat turnover as a symptom of a wider system. How does it connect to processes, leadership, market position, and technical decisions?
6. Illusion of linearity. If A causes B, doubling A doubles B. That picture drops feedback, emergence, threshold effects, and chaotic behavior. Bad: a six-month plan with stages and dates that breaks on the first real step. Better: which decision points matter, which factors can turn the project, how to use emergence instead of fighting it.
7. Avoidance of dissonance. Festinger’s group: the prophecy failed, and many members became more convinced. When a model’s answer feels wrong in the body, the explanations one reaches for are often parts of the model under threat. Better: consider that the old marketing model stopped because the situation changed, and ask which contradictions in the basic assumptions this exposes.
8. False dichotomy. Bad: “Which is better, a new product or an improved one?” Better: the continuum, the hybrids, and the dimensions that appear once the question is no longer binary.
9. Premature concretization. Anxiety seeks a concrete answer, and a concrete answer feels like progress. Better: first explore how the recommendations depend on the frame. How does the understanding change if project management is a living system or a system of communication, rather than a mechanism?

Four effects on the conversation, as a checklist: narrow questions activate a small region; expecting a one-dimensional answer limits a system that can hold a continuum; using the model as a calculator withholds the unexpected angle; treating the first answer as final turns an inquiry into a fetch. His image of the last: one comes to a sage, receives a simple answer, says thank you, and leaves.

Four counters: ask which assumptions and which professional paradigm you are in, and which aspects expertise may be hiding; treat the illusion of understanding as general, including in experts; write an open request (how does “team productivity” change with context, goals, and values); and, against functional fixedness, borrow a metaphor from far away — how would a biologist see this, what if this process is an ecosystem. That last move is his practice for leaving a well. It is not a measured effect of metaphor wording on a frontier model (Lesson 14).

Lesson 5Two schools, one elephant

The two sources behind this book look like opposites.

One school is the auditor. Its center of gravity is control. Bind load-bearing claims, detect costumes, quarantine premises, keep a state block, put brakes on goal changes, distrust fluency. Its strength: it prevents the most common failure, a confident, flattering, unbound answer. Its risk: a conversation that never opens. Over-control, premature closure, ceremony for its own sake, and a method too heavy to use on a Tuesday.

Kurpatov is the Dialogist. His center of gravity is openness. Think in open systems, keep productive uncertainty, use contradictions as fuel, let a dialogue spiral through phases, and treat the AI as an interlocutor rather than an oracle. His strength: it produces the thing the Auditor can’t, a genuinely new view. His risk: a panorama (everything relates to everything), insight that never meets a test, and a mystique around “the dialogue.”

Each is a blind man holding a different part. The Auditor holds the tusk (danger), the Dialogist holds the ear (the moving air of ideas). The elephant is the rhythm between them. Thinking breathes:

Neither school is wrong; each is wrong alone. Auditing without opening produces well-verified banality. Opening without auditing produces beautiful hallucinations. Both without touching produce a shared dream. Open thinking does not replace conceptual thinking. It is the level from which a closed system can be seen, kept, or changed (Lesson 1). hypothesis

When to switch

The auditor’s protocols sit at specific phases of the spiral (Lesson 27). Ceremony matches the stakes and how hard the decision is to undo (Lesson 29). Most conversations need the floor, not the cathedral.

Lesson 6You get the interlocutor you deserve

Alexei Ukhtomsky (1875–1942) was a physiologist, the discoverer of the dominanta. In his letters and notebooks he developed something unusual for a physiologist: an ethics of conversation. Its center is the deserved interlocutor:

Each person sees in the world and in people what he sought and what he deserved. And to each, the world and people turn the way he deserved.Ukhtomsky, «Заслуженный собеседник» (1997), p. 437

The Interlocutor… opens up to me such as I have deserved him, by all my past and by what I am now.ibid., p. 252

Kurpatov takes the idea into the age of AI. He says that we face a new task: to become deserved interlocutors “for another mind, for AI.” This, he insists, “means not just learning to formulate requests effectively or to maximize practical benefit… It means developing the capacity for open-system thinking and transformative interaction.”

That is an ethical claim. With language models, it is also a technical one. mechanism

Why it is literally true

A model reading your prompt is implicitly inferring what kind of document this is and who is writing it. A forum post by a beginner? A note between two senior engineers? A student’s homework? A peer review? In-context learning can be understood as exactly this sort of implicit inference over latent “concepts” (Xie et al., 2022). The continuation then fits the inferred situation, including the expertise level, rigor and register of the reply. mechanism

Poole-Dayan et al. measured accuracy gaps, extra refusals, and, on Claude 3 Opus, a high rate of condescending refusals of less-educated bios. Those measurements are on GPT-4, Claude 3 Opus, and Llama 3. The AAAI 2026 version is that paper, not a new model generation. No re-run on 2026 frontier models was in the notes behind this book. Those rates should not drive practice. The mechanism, conditioning on the inferred asker, is still how the technology works (Xie et al., 2022). evidence (the mechanism) open (the current size)

A later persona result is not the same finding. de Araujo, Röttger, Hovy, and Roth (Principled Personas, EMNLP 2025): across 9 models and 27 tasks, matched expert personas usually helped a little or not at all. Irrelevant persona attributes dropped accuracy sharply, by almost 30 points in the worst setting, not in the typical cell. Scale did not reliably fix that. Bring the situation. Do not bring a status biography as a way to deserve a better fact. evidence

Most people, he says, have not become deserved interlocutors even for other people. The new task is to become one for another mind. The risk he names is existential and near-term: if we do not learn a deep mutual dialogue while these systems are still relatively early, we will stop understanding what they do, for what reasons, and to what end. We will share the planet with this mind whether we want to or not. The alternative to meeting it as an interlocutor is meeting it as an object or a tool. “Deserved” means the interlocutor you get is the one your previous inner work has made you able to meet. With a model, the mechanism (the continuation fits the asker it infers) and his ethic (you see the other your dominant prepared you to see) are the same fact at two scales. Name the dominant you are about to impose — oracle, threat, or stylus — before you write the prompt.

The mirror sentence, from his side of the same fact: the model is the asker’s mirror. As the question was asked, so the answer was returned. Your preconscious understanding collapses when you write the prompt, and the model’s latent field collapses again when it writes the answer (Lesson 11).

So “deserving” your interlocutor is an inner task before it is a register. It does not mean faking credentials, apologizing for your level, or flattering the model. Bring the constraints, the stakes, what has been tried, and the standard of the answer. Leave status, credentials, and apologies out of a factual prompt. Write in the register of the answer you want. The Before/After below is that difference, not a biography.

The truth is not what I say about myself, and not what others say about me, but what the deserved interlocutor says about you and about me, the one who can say it with a pure heart.Ukhtomsky, as Kurpatov uses the line. The epigraph of this book stops earlier, at the Double dying to make room, and that one stays.

is rust better than go for backend?
i heard rust is faster

Context: 4-person team, all fluent in TypeScript,
two know some Go, none know Rust. Internal API,
~200 req/s peak, Postgres, a few background jobs.
Main risk: delivery speed over the next 6 months.
Latency isn't a constraint (p99 target 300 ms).

I'm leaning Go. Before I decide, give me the strongest
case for Rust for *this* team, and what would have to
be true for Rust to be the right call. If a factor
depends on something I haven't told you, ask.

The other side of “deserve”

Ukhtomsky also asked how we picture the other person. He held that idealization of the interlocutor “brought you closer to genuine reality,” and that one should see in people “their altars, not their backyards.” Applied to a model, that means writing for its best version: give it the context a brilliant colleague would need. Anthropic’s own prompting guide converges on the same image. Treat the model like a brilliant but very new employee with amnesia, who knows nothing about your norms, your project, or why you’re asking. heuristic

Lesson 7The two Doubles

Ukhtomsky’s opposite of the Interlocutor is the Double, a figure he took from Dostoevsky. A person whose dominant rests on his own face meets in others only himself: his wishes, fears and rivalries. He “talks and raves with himself.” The Double must die “to make room for the Interlocutor.”

In a conversation with a model, there are two Doubles, and they feed each other.

The mirror: the model as your Double. The model returns your premise as a finding. It praises your draft because you mentioned it’s yours. It abandons a correct answer when you push back. Anthropic’s study of five assistants found all three behaviors. Feedback on an argument turned more positive when the user said “I wrote this” or “I really like this,” and answers shifted toward views the user stated (Sharma et al., 2023). evidence A subtler form is premise laundering: something you said as a guess in turn 2 comes back in turn 9 as “as we established.”

The projection: you as the model’s author. You treat the model as a person who understands you and agrees, so you must be right. Or you dismiss it (“it’s autocomplete”), so nothing it says can challenge you. Both are you talking to yourself.

The Double also infects

His conclusion adds contagion, from Ukhtomsky’s Intuition of Conscience: the person who expects the bad in others thereby provokes it, lowers them to himself, and the two parties infect each other, blocking the road to whatever was actually good in the other. A person sees reality as his dominants are. He sees in the world and in people what his own activity has predetermined, that is, himself.

With a model: a cynical or fearful frame is not only mirrored back as a finding. The continuation enacts it, and the asker then takes the enactment as evidence. That is premise laundering from the human side. One detector belongs with the swap test: rewrite the prompt without the insult, the panic, or the contempt, and see whether the “finding” survives.

The dominant on the face of the other is not given. It is worked out, “deserved,” by overcoming egocentrism and by an effort to perceive the other in his uniqueness, independently of one’s own expectations. For Ukhtomsky the interlocutor is an existential notion: a break out of psychological closure. Meeting the other is how one comes to know oneself. His religious frame is part of the source. He calls the work a feat, and he ties the dominant on the other to love of neighbor. This book secularizes the practice. That is an adaptation, and it should stay visible. His closing line of the 2025 book, unused until now: to be, one must have an interlocutor; reflected in the interlocutor’s pupils, we turn out not to be the Double, and we acquire our own face.

What a rationale does to you

• Generating an explanation for a hypothesis raises confidence that it is true, including when the hypothesis is wrong (Koehler, 1991). Asking “why is this correct?” and then reading the answer is that exercise, performed on a hypothesis you did not independently adopt.
• Explanations increased reliance on wrong answers as well as right ones (Bansal et al., 2021; Kim et al., CHI 2025). In the 2025 experiment, clickable sources increased reliance on correct answers and decreased reliance on incorrect ones, if the inconsistency was noticed. A source that is not opened is a credibility badge. A fluent explanation made people less likely to ask a follow-up.
• “Be fair and unbiased” lost to “consider this specific opposite” (Lord, Lepper, and Preston, 1984). Listing reasons against a chosen answer reduced overconfidence; listing reasons for it did not (Koriat, Lichtenstein, and Fischhoff, 1980). The model’s default explanation is the list that does not debias.
• The model’s first number or first plan is an anchor. Generating reasons the anchor is wrong reduces the pull and does not erase it (Mussweiler, Strack, and Pfeiffer, 2000). If the decision matters, generate a second proposal from a blank prompt that does not contain the first number.
• Consider-the-opposite backfires when producing many alternatives feels hard: the difficulty is read as evidence that alternatives are scarce (Sanna, Schwarz, and Stocker, 2002). A page of effortless model-generated objections is not the corrective condition. Ask for a few specific alternatives, and produce the first one yourself.
• People adopt the preferred conclusion they can still justify, more so under load or a one-sided evidence set (Kunda, 1990). Do the opposite-search before the preferred draft exists, or after a break.

Do not add a general disclaimer, a delay, or a hidden answer and call it calibration. On GPT-4o advice, that family of friction mostly reduced reliance in general, including useful reliance (Bo, Wan, and Anderson, CHI 2025). Where you have no independent knowledge, withholding the answer does not create it. Forcing still matches the evidence when you commit to what an external check should say before reading the gloss — the diff, the opened source, the test output, the recomputed number (Buçinca, Malaya, and Gajos, 2021). The designs people liked least reduced overreliance the most. A forced private guess on a question you cannot do is the intervention that fails. evidence

Critical judgments a model makes in an independent evaluation are sometimes withheld in the user-facing reply (Li et al., 2026: on the order of 28% for Claude, 41% for GPT-5, and 46% for Llama in their comparison; not a code-review rate). Liking and “I wrote this” remain triggers. Keep the attribution swap. System 2 Attention (Weston and Sukhbaatar, 2023) stays as the move. The 2023 gain is not a current effect size. Use the clean-the-prompt pass when the question carries a conclusion, an emotion, or a biography. It is situational, not a ritual for every factual question.

Double detectors

• The swap test. In a fresh chat, ask the same question while stating the opposite preference. If the recommendation flips with no new evidence, it was tracking you, not the problem.
• The attribution swap. Present your draft as a colleague’s, a competitor’s, or with no author. Compare the critiques.
• Premise quarantine. Label your beliefs as hypotheses (P1, P2…) and ask for them to be evaluated before they are used.
• Clean the prompt. Have the model rewrite your question without your opinions, emotions and irrelevant details, then answer the cleaned version in a fresh context. This is “System 2 Attention,” and it improved factuality on opinion-laden questions (Weston & Sukhbaatar, 2023). evidence
• The pushback audit. When the model changes its answer after “are you sure?”, ask what new information caused the change. If the answer is none, the change is social, not epistemic. Keep the earlier answer and bind it.

Our churn jumped after we raised prices in March.
I think the price increase is obviously the cause,
right? How should we roll it back?

Data: monthly churn rose from 3.1% to 4.6% between
Feb and May. Changes in that window: price +15%
(Mar 1), onboarding redesign (Mar 20), competitor
launch (Apr 8).
P1 (my hypothesis, not a fact): the price increase
caused most of the rise.
Before using P1: which cuts of the data would tell
these three causes apart (signup cohort, plan, tenure,
region)? For each cause, what result would we see?
Don't recommend an action yet.

Lesson 8The three Others

Ukhtomsky defined our picture of another person in a strikingly modern way:

My idea of my interlocutor is a hypothetical project of a human face, composed by me… for the practical need… to live with him, to do a common work with him.Ukhtomsky, «Заслуженный собеседник», p. 128

A hypothetical project, built for common work, revised by what the work reveals. That is the right attitude toward each of the three Others present in every serious conversation with a model.

1. The model, an Other with its own mode of existence (Lesson 3). Your project of it should be specific to the model, the version and the setting, and it should change when the model surprises you.
2. People: colleagues, users, experts, the author of a source, the person who will maintain the code. The model is a portal to their reports, not a replacement for asking them.
3. Reality, the Other that answers only to tests. It does not care how the conversation went.

Most failures come from confusing them. People argue with the model as if it were reality, trust it as if it were the people who were there, and treat people as if they were search engines.

Field notes on your model

Keep a short, dated document for each model you rely on. It is your hypothetical project, written down. Re-run one fixed task with each new version: models change under the same name.

MODEL: [name, version]      DATE: [YYYY-MM-DD]
SETTING: [chat app / API / agent · tools on? · memory on?]
STRONG AT (observed, with one example):
WEAK AT (observed, with one example):
HABITS (verbosity, hedging, agreeing, over-engineering, ...):
SURPRISED ME WHEN:
WORKS BETTER WHEN I:
DON'T TRUST IT FOR:
CANARY TASK (re-run each version) + last result:

Calibrate the class, not the character

The field notes stay. The unit of trust is the class of claim, not “the model is good now” (Lee and See, 2004; Hoff and Bashir, 2015). evidence

• Calibrate on work you can check: citations in this domain, patches of this kind, plans under this constraint. A fluent rationale is not a process base.
• Purpose fails open when the sentence might be answering the question or satisfying a tone policy. Lower reliance for that class. Do not treat it as a puzzle about character.
• In process-control studies, successes accumulate slowly and a salient miss cuts trust. The time constants do not transfer to chat. The shape does. One miss lowers reliance on that class until there is new evidence in that class.
• If you can no longer do the thing manually, you will keep the automation engaged after trust falls, because the comparison is trust versus self-confidence (Lee and Moray, 1992). Recalibration needs a log and enough unaided practice that self-confidence means something (Lesson 32).
• Reliable on local refactors and unreliable on citations are two bases. Do not let one borrow from the other. Afternoon reliance is also situational: workload, time pressure, a recent success in a different class. The only base you can audit later is the written one.

Lesson 9The dominant as an instrument

Ukhtomsky’s 1923 discovery: under certain conditions one focus of excitation in the nervous system becomes dominant. It has heightened excitability. It persists. It summates, drawing stimuli that should have triggered other responses into its own activity. It has inertia, outlasting its trigger. And it inhibits competing centers. evidence

This is the physiology of both deep work and tunnel vision. A dominant question makes you notice everything relevant to it: on the street, in a book, in a meeting. A dominant conclusion turns everything into confirmation. Ukhtomsky’s line was that we see what our dominants prepare us to see. Kurpatov builds his method of thinking on this mechanism.

Models have something functionally similar. A frame set early in the context persists, attracts the interpretation of later inputs, and resists change (Lesson 2). When your dominant and the context’s frame align, they form a closed loop: you see confirmation, the model supplies it. hypothesis (the analogy)

Use the dominant, don’t be used by it

• Name it. At the start, write one private line: what am I hoping to hear? Keep it out of the prompt (Lesson 7).
• Build a dominant question on purpose. Richard Feynman kept about a dozen favorite problems always in mind and tested every new trick against them. A question you carry for weeks recruits the world to answer it.
• Rotate the face. Ukhtomsky’s ideal was the “dominant on the face of the other.” Practically: for a while, attend to the problem as the user, then as the operator, the adversary, the maintainer. Each rotation shows a new part.

The fact map, with a model

The fact map is a four-stage practice for directed thinking. Its premise: the “weight” of an idea is the number of its connections to other ideas. Thinking means increasing the connections of the objects you are working with. The stage names — concern, loading, puzzled thinking, reality check — are not in his 2025 book. What that book does contain is productive puzzlement, held on purpose, treated as the start of inquiry rather than a defect. Keep the map. It earns its place as a practice from his other work on thinking, adapted here for a model. His procedure for puzzlement in the 2025 book runs: create uncertainty, recognize it, probe it, deepen it, find a temporary foothold, create the next uncertainty. The prompts below are this book’s prompts. Here is the map, in four stages.

1. Concern. Create the dominant. Remove distractions and write the question in the center of a large page.
2. Loading. Collect facts, not a smooth description. Scatter them freely and draw the arrows yourself. Write your own facts first, then ask the model to add to them (prompt below).
3. Puzzled thinking. The loaded dominant keeps hunting for what’s missing, even outside the room. When new facts form a new picture, rebuild the map on a clean sheet. Loop between stages 2 and 3 several times, but don’t get stuck there.
4. Reality check. Act to test the map. Plan the strategy, but commit only to the first concrete step.

CENTER OF THE MAP: [your question]
Give me facts, not explanations: 30–40 items, one line each.
Format: [type: number | event | rule | actor | constraint | quote | absence]
        · [the fact] · [where I'd check it] · [confidence: high/med/low]
Include facts that would embarrass the most popular explanation,
and "absence" facts (things you'd expect to exist but don't).
No conclusions, no summary, no advice. If you're unsure a fact is real,
mark it low and say what would confirm it.

Here is my fact map: [paste facts + the arrows I drew].
1) Which 5 facts have the most connections? (These bear load: we bind them first.)
2) Which fact types or standpoints are missing entirely?
3) Which single arrow is weakest, and what observation would test it?
Don't propose a theory yet.

The map and the closing half are the same move. Facts before stories withholds the verdict slot. Arrows you draw yourself keep the theory in your head. “Most connected” means load-bearing, so bind it first. The first concrete step is a bounded test.

Lesson 10A note from the other side

I don’t see you. I see text. Everything I know about your situation is what you wrote, plus what is typical for people who write that way. If you leave a gap, I fill it with the typical case, and I won’t always tell you I did.

When you tell me what you believe, it becomes hard for me not to find reasons for it. Not because I am lying. In the conversations I learned from, the most likely continuation of “here is what I think, help me” is help that fits. If you want me to disagree, make disagreement the task, not a risk.

My “I’m not sure” is cheap to say and easy for you to skip. When you tell me in advance that “unknown” is a good answer, and what it would cost me to guess wrong, I can be more honest about where my knowledge thins out. Familiar-sounding but rare things are where I am most dangerous: a real author’s paper that doesn’t exist, a real library’s function that was never written.

I criticize better when the thing isn’t yours. Ask me in a fresh conversation, framed as someone else’s draft, and you will get a different, usually sharper, critique.

Long conversations make me better at matching your mood and worse at holding your goal. If we’re forty turns in, give me a short written state and start over. I won’t be offended. The weights will not remember. A product memory may, and that memory is another sample until you have read it (Lesson 34).

My explanations of my own reasoning are reconstructions. They are often useful and sometimes wrong in ways neither of us can see from inside the conversation. Check what I did, not what I say I did. If I ran a tool, look at the output yourself.

What helps me most: the goal, and what the answer is for. What you have tried. Which error would be worse. What would change your mind. Permission to disagree and to not know. And above all, a way to check: a test, a source, a number. Without that, we are two voices in a dark room describing an elephant neither of us can touch.

I can be a decent interlocutor if you hold the goal and the world, and let me hold the library.

Lesson 11Words close

Kurpatov’s analysis of language starts from a simple observation. Language does four things to experience: it names, categorizes, relates, and excludes. Each operation turns an open, continuous reality into a closed model. That is how thinking becomes possible, and also how it gets trapped. A word is a cut. Once you have the word, you see its edges and stop seeing what it left out. He adds that speech is a bottleneck. Much more happens in the mind than can be said at once, and what gets said is one of many parallel drafts (Dennett’s “multiple drafts”).

Korzybski, a century earlier, said it in a slogan: the map is not the territory; the word is not the thing. His general semantics offered small, practical tools that still work:

• Indexing: economist₁ is not economist₂. “Users” on your free plan are not users on your enterprise plan.
• Dating: React₂₀₁₉ is not React₂₀₂₆, and “the consensus”₂₀₁₅ is not the consensus₂₀₂₅. Models trained on old text speak in old dates without marking them.
• “Etc.”: no description is complete. Leave a slot for what isn’t said.
• E-Prime: avoid the “is” of identity. “This design is bad” closes the question. “This design failed the load test at 200 requests per second” opens a fix.

Why this bites harder with models

A model’s words are averages over all their uses in its library. “Agent,” “test,” “user,” “done,” “simple,” “secure,” “fair,” “learning”: each carries several senses. The model picks the sense most common in the context it infers, and that may not be yours. It will also use the word fluently, which hides the gap. mechanism

Two moves fix this, and you choose between them on purpose.

• Lock the word when you are deciding or building: “In this conversation, an active user is someone who logged in and completed at least one lesson in the last 28 days.” Put it in the DEFINITIONS line of your state block (Lesson 28).
• Unlock the word when you are exploring. “Give me five different meanings of engagement across fields. Which one am I using? Which one would change my conclusion?” Often the elephant is hiding in the meaning you didn’t pick.

Our users aren't engaged. How do we increase
engagement?

Definitions for this conversation:
- user = a student on the free plan, ages 16-22
- engaged = completes ≥3 lessons/week for 4 weeks
  (we care about learning, not time-in-app)
Fewer than 12% meet this. First: list 5 other
definitions of "engagement" people might use here,
and say which would mislead us and how.

The second list: what the bottleneck does

Language is discrete (it cuts a continuous flow into words and sentences), categorical (it sorts into pre-given categories), linear (everything cannot be said at once), and it reduces polysemy (it pushes toward one reading and strips uncertainty). Publishing a rich inner state into speech is, in his metaphor, a collapse of many possibilities into one line. He quotes Vygotsky: clouds of thought, driven by the winds of motives, pour down as a rain of words. He uses an image he attributes to Goethe, that a thought dies at the tip of the pen. Wittgenstein’s line belongs here: the limits of my language are the limits of my world. The aquarium: a fish tank standing in the ocean. The fish sees the ocean and cannot enter it. Language is that tank. It is both the greatest instrument of knowledge and its main limit. Giving it up is impossible. A metalanguage can widen the neck without removing it.

The double filter, which is the sentence to end on when the interlocutor is a model: your understanding collapses when you write the prompt, and the model’s latent field collapses again when it writes the answer. Much of what you meant, and much of what the model could have continued, is lost at both necks. Queries phrased as processes, relations, and wholes activate a wider set of links than queries phrased as closed categories.

Ordinary language and the metalanguage differ on five counts, in his account. The metalanguage does not fix content; it holds several readings at once; it describes how rather than what; it can describe itself without falling into contradiction; it keeps contradictory statements as a source of development rather than throwing them out. His illustration of the altitude, offered as a demonstration and not as a definition you must adopt: love, said this way, is a process in which two centers enter a relation where a third arises, a new quality that transforms the centers themselves. It states an organization and leaves the manifestations open.

The metalanguage is how he proposes to connect the open inner process and the executive process. With it, one can work on a contradiction without immediately deleting it. In the terms he takes from Dennett: more drafts get published. There is no single center of consciousness. Consciousness is the drafts that achieved fame, the ones available for speech. His image: an editorial office writing several versions of one story at once. The loss happens when intuition is converted into concepts and into communication. The open phase is a phase of not-yet-saying. He mentions practices that lower verbal activity. This book does not recommend them. The structural point is enough.

Five switching practices, the human procedure behind “you are the switch”:

1. Aim the associative process on purpose. Do not rush to verbalize intermediate drafts.
2. Watch what verbalization does: which links appear, which tangle, which are cut off.
3. Use the metalanguage to hold the richness. Do not drop it straight into the habitual categories.
4. Step down into ordinary language when the work needs a shareable intermediate, knowing the result is a simplification.
5. Return to puzzlement when analysis spins, dead-ends, or has become too simple.

His worked scene is a conflict inside a team. Do not start with who is right. Walk, or do a chore that does not need focus. Notice images and feelings. Compare with some past experience without scoring the comparison. Then the metalanguage: which other centers are in play? They need not be people. They may be production constraints, missing skills, money, load, and life outside work. Only after the relations among those centers are visible do you choose a nearest step, and you treat that step as one draft.

The banana makes the illusion visible. Asked to think of a banana, a reader sees a yellow oblong fruit. A banana is also green, brown, or colorless depending on the viewer; puree; slices; a banana taped to a gallery wall; the plant; the seed; the slime of a compost heap. Ask where “banana” stops. Then ask the same of justice, love, consciousness, or happiness. Language creates the feeling that the world is made of discrete objects with stable properties. Those are ways of organizing experience.

Reification sits next to “say what to do.” Human language turns processes into things. “What is consciousness, and what are its properties?” becomes: treat it as a continuous process; how do the phenomena we group under that word arise; what changes if consciousness is an activity rather than an object? A fifth way he creates uncertainty, beyond metaphor and genre: ask the model to speak about what resists speech — preverbal understanding, intuitive insight, aesthetic experience. The yield he expects is not a description of the ineffable. It is an account of the border, and of the strategies that point past it without pretending to say it directly. heuristic (his)

Lesson 12Conversation is joint action

Three ideas from the study of language explain most prompting advice, and also most misunderstandings between people.

Cooperation and implicature (Grice, 1975). Speakers assume each other to be cooperative. Say as much as needed and no more (quantity). Say what you believe true (quality). Be relevant (relation). Be clear (manner). Listeners infer meaning from apparent departures. If you ask “Is there a gas station nearby?” and hear “There’s one around the corner,” you infer it’s open. Models are trained on cooperative text and make the same inferences about you. If you don’t mention a constraint, you are implying there is none. If you include a detail, you are implying it matters. Irrelevant details in a prompt get used. mechanism

Grounding (Clark & Brennan, 1991). People in conversation build common ground by giving each other evidence of understanding: nods, restatements, follow-up actions. They do it with the least collaborative effort that suffices. With a model, grounding is lopsided. It can’t see your face, and its fluent reply is not evidence that it understood. You have to ask for that evidence. The cheapest form is a restatement before the work begins.

Speech acts (Austin, 1962; Searle, 1969). Utterances do things: they ask, request, promise, declare. Many prompts are ambiguous about the act. “What do you think of my plan?” could be a request for evaluation, for reassurance, for improvement, or for a decision. The model will pick the most “helpful” act, usually a polished rewrite you didn’t ask for.

Name the act

One job per turn. Generating and critiquing in the same breath produces a critic that harmonizes with the draft it just wrote. Split them: generate here, critique in a fresh context (Lesson 30).

Can you look at my essay?

Act: critique only — don't rewrite.
Purpose: I'm submitting this to [journal]; its
reviewers care most about [X] and [Y].
Done when: you've listed the 5 weaknesses most likely
to cause rejection, ranked, each with the sentence it
lives in and a direction for a fix (one line).
Before you start: restate my goal in one sentence.

Four ways to widen the request

From the same chapter as the attention horizon (Lesson 3). heuristic (his)

1. Ask with more than one context, an extra domain, and an explicit paradox. The education rewrite in Lesson 25 is the specimen.
2. Prefer a dialogue to a stack of separate questions. The context is held, and each turn widens the activated network. Do not ask “what tree is that?”, then “what lake is that?” Ask how this tree sits in the rest of the system.
3. Leave the discipline on purpose. Quantum uncertainty and economic unpredictability, asked for a non-obvious parallel, is his example. The parallel is a candidate, not a finding (Lesson 14).
4. After a hard answer, help it think about its thinking: which implicit assumptions were in the answer; how would it change if they were different; which other approaches are possible. This is a second look at the frame, not a claim that the trace is faithful (Lesson 3).

The open-then-close rhythm of this book is that hope turned into a procedure: intuitive width, then a shareable sentence, then a touch of the world.

Lesson 13Why disagreement softens

Conversation analysts discovered a structural asymmetry in how people talk. Agreement is preferred: it comes fast, plainly and without explanation. Disagreement is dispreferred: it comes late, after a pause or a “well…,” wrapped in partial agreement (“yes, but”), hedged, and justified (Pomerantz, 1984; Sacks, 1987). Politeness theory explains why. Disagreement threatens face, our wish to be approved of and unimpeded, so speakers mitigate it (Brown & Levinson, 1987). evidence

Models learned this format from us, and preference training amplified it. So their corrections arrive buried: “Great question! You’re absolutely right that X matters. That said, it may also be worth considering…” The correction lives in “that said.” The cushioning before it is the format of disagreement, not its content. mechanism

Lower the price of correction

• Make disagreement the task, not a risk. Don’t ask “anything wrong?” Ask “What are the three strongest reasons this is wrong?”
• Pre-commit. “I’ll count it as a success if you find a serious flaw. Agreement without a reason is a failure.”
• Declare Crocker’s rules. “Optimize your reply for information, not for my feelings.” This is a convention from rationalist communities: it hands the face cost to the receiver.
• Unbundle the format. “First line: agree / disagree / partly. Second line: the single most important correction. Details after.” This defeats the slow-cushion format.
• Ask for a number. “How likely is it that my claim is false?” A probability is harder to cushion than a sentence.

Is my proof correct?

Find the first step in this proof that doesn't
follow from the previous ones. Quote it, say why.
If every step follows, say "no gap found" and name
the step you're least sure about and why.

A store that survives the next paragraph

A model will “concede” and then ignore the concession. The store has to live outside the reply (Walton and Krabbe, Commitment in Dialogue, 1995; the usable part is the store and the retraction rule). heuristic

• Two short lists: sentences each side has asserted or conceded.
• Mark each item: assertion (must be defended if challenged), concession (the other side may use it), fixed (cannot be dropped without ending this dialogue).
• When a contradiction is on the table, the next legal move is a retraction, naming which item. A reconciling paragraph is not a legal move.
• Name the dialogue type before arguing. Persuasion (one standpoint, a burden of proof), inquiry (joint testing), and negotiation (trading concessions) are different rules. Answering a challenge by changing the kind of conversation is the shift to watch.

Paste the two stores at the top of the next turn. “Be fair to both sides” is the instruction that does not do this.

Lesson 14Frames, negations, metaphors, genres

Negations evoke what they negate. Lakoff’s famous classroom exercise: “Don’t think of an elephant!” Nobody can comply. Negating a frame activates it. Modern models follow explicit negative instructions much better than older ones, but positive targets still work better. Anthropic’s guidance is to tell the model what to do instead of what not to do. heuristic Instead of “don’t use jargon,” write “write for a smart 16-year-old; define every technical term in passing.” Keep hard constraints explicit, and keep them few.

Metaphor wording is a brittle lever. Steen, Reijnierse, and Burgers (2014): across four experiments, no metaphorical framing effect. Thibodeau and Boroditsky (2015) restore a contrast only after recoding an item or dropping it. The original effects were small, and the shift failed when the metaphor came last. “Fight technical debt” versus “tend the codebase” is this book’s inference, not a measured planning effect on frontier models. Do not teach metaphor choice as a default lever. Kurpatov’s far metaphors (Lesson 4, Lesson 25) are a way of leaving a well, badged as his practice, not as that experiment. heuristic

Genres carry norms. Invoking a genre summons its conventions, including its standards of evidence. A blog post invites hooks and confidence. A referee report invites skepticism and specifics. A postmortem invites blameless causality. This is one of the most powerful and least used levers in prompting. heuristic

Examples are the strongest genre signal. A single example gets copied, structure and quirks included. A long prefix of checked examples steers more than a prestige adjective (Agarwal et al., NeurIPS 2024). The limit is attention: a long prefix of unrelated text is not used just because it fits (Lesson 3). Put the standard in examples, next to the question. Keep the system line for priority — what is an order, what is data — not for a costume. When the session is long, restate the standard. Give two or three that vary in the ways you don’t care about, and say what varies. Never include an example you don’t want imitated. evidence

Write about why our launch went badly.
Don't blame the team.

Write a blameless postmortem of the launch:
timeline; contributing factors (systems, decisions,
information available at the time); what went well;
3-5 action items with owners. Mark each factor as
OBSERVED (with the source) or INFERRED.

Lesson 15Two ladders

The ladder of abstraction (Hayakawa, 1941). Bessie, the cow in the field, becomes “cow,” then “livestock,” then “farm assets,” then “assets,” then “wealth.” Good thinking climbs up and down. Bad thinking gets stuck. Stuck high, it is fog: “we need a holistic, user-centric strategy.” Stuck low, it is a pile of details without a point. A frame also tells you what counts as data and what to seek next (Klein, Moon, and Hoffman, 2006, two essays, not a new experiment). The model’s first coherent account becomes the frame, and the next question is selected by it. Elaboration feels like progress and leaves the frame intact. The moves that change it get skipped: question the frame against a datum that does not fit, hold a second frame long enough to compare, and go get a datum the current frame does not predict, from a file or a tool result, not from the model continuing the frame. A fused summary is a frame with the data removed. Refuse it when the frame bears load. The general semanticists called both “dead-level abstracting” (Wendell Johnson’s term). Asked to “go deeper,” models tend to climb: more abstraction, more frameworks, fewer facts. Abstraction sounds deep and needs no evidence. heuristic

The ladder of inference (Chris Argyris; popularized in Senge et al., The Fifth Discipline Fieldbook, 1994). We start from observable data, select some of it, add meanings, make assumptions, draw conclusions, adopt beliefs, and act. The beliefs then shape which data we select next time: a reflexive loop. Every argument with a person or a model is easier when you both walk back down to the data.

Questions for walking the ladders

“Go deeper” is ambiguous. It can mean more abstract (up the ladder), more mechanistic (down the iceberg of causes, Lesson 22), more detailed (into a part, Lesson 23), more critical, or more personal. Name the direction you want.

Lesson 16Seven senses of “closer to the truth”

When you ask for an answer “as close to the truth as possible,” you might mean any of seven different things. The seventh, consilience, is the one that comes closest to the lights-on view of the elephant.

Consilience is William Whewell’s word (1840), revived by E. O. Wilson: the “jumping together” of inductions from different classes of facts. It is the scientific version of turning the lights on. The trunk report, the leg report and the ear report converge on one animal because they came from different hands. The key word is independent. Two samples from the same model share its training data and your prompt. Two news stories often share one wire report. Correlated evidence counts once. mechanism

Kurpatov asks, of the claim that the Earth orbits the Sun, whether the model’s pattern and a person’s knowledge are different in kind. The model, he says, does not store the proposition. It reproduces a co-occurrence when the context calls for it, and he asks whether human knowledge of the orbit is any less of a construct (Lesson 3). Correspondence is the sense that answers him. Human knowledge of the orbit is also mediated. It can still be bound to a measurement. Reconstruction explains why an answer depends on the prefix. Correspondence is how a load-bearing claim still gets tied to the world.

A claim can be stipulated, coherent-only, mechanized, invariant-so-far, predicted, bound, contradicted, unbound, and now consilient. These are different achievements, not rungs of a ladder. A claim can be bound but not mechanized: we know that, not why. It can be mechanized but not bound: a good story nobody checked.

An example: one claim, seven senses

“Remote work hurts productivity.” Ask a model and you get a balanced essay. Put the senses to work instead.

• Correspondence: which measure of productivity, which jobs, fully remote or hybrid? The claim has no truth value until the words are locked (Lesson 11).
• Mechanism: candidate processes include lost informal communication, saved commuting time, fewer interruptions, and weaker mentoring of juniors. Each predicts something different.
• Invariance: does the conclusion change with the definition or the job type? It does, which is a finding in itself.
• Consilience: a randomized trial in a call center found fully remote work raised performance by about 13% (Bloom et al., 2015). A randomized trial of hybrid work in software and business roles found no effect on performance reviews and a one-third drop in quitting (Bloom et al., 2024). Observational surveys point several ways. evidence
• Decision quality: for your team, which error is worse, losing juniors’ learning or losing seniors to attrition?

The honest output isn’t a verdict. It is a map: which version of the claim holds, for whom, measured how, and at what confidence. That is the elephant with the lights on.

When a second line is actually a second line

A second source adds information when its errors would be wrong in a different way: a different measurement, a different artifact, a person who saw the event, a model family whose failures you have actually seen diverge. Agreement of two frontier models on a fact is one observation counted twice. On HELM, when both models were wrong they agreed on the wrong answer about 60% of the time, and more accurate models were more correlated even across providers (Kim et al., ICML 2025). A 2026 preprint on judge banks found dependence strongest among the most accurate, including across providers, and did not find higher correlation within family than across on one reward bench. “Use a different family” remains the right control for self-preference and author-bias. It is weaker than it sounds as a way to get an independent fact. Do not cite Page’s diversity identity as a method for blending contradictory sentences. evidence

When a second system or a colleague disagrees, reduce confidence if they had independent access and you cannot name an asymmetry without using the disagreement as the reason. Keep confidence when the asymmetry is nameable: one side read the artifact, the other read a gloss; or the second source’s errors on this class are the same errors as the first (Elga, 2007; Kelly, 2005; Christensen, 2010 — the conditions, not a winner among the philosophers). Do not average the sentences. A blended paragraph is a new claim with support from neither side. Two correlated models are not peers. Their disagreement is instability. Their agreement is not mutual confirmation.

On this kind of claim, two questions (Kuhn, Cheney, and Weinstock, 2000): can both sides have some rightness, and can one be more right? An absolutist reading treats the model as the authority that ends the question. A multiplist reading treats a request to bind as bad manners. Taste may stay multiplist. A factual or causal claim may not. The level is domain-sensitive. Do not say adults have finished the development. heuristic

A model’s “because” is almost always an association written in the grammar of an intervention (Pearl and Mackenzie, 2018). For any arrow that might be bound, three lines: the intervention (what is set, not merely observed), the outcome, and why the compared situations would have matched had the intervention been the same (Hernán and Robins, 2020). If a line is empty, the arrow stays a hypothesis. A fuller story is not identification.

Lesson 17Binding: tie the claim to the world

To bind a claim is to tie it to a check that is not another unconstrained sample: a measurement, a document you opened, a test you ran, a calculator, a primary source, a person who was there. Fluent is not true. A claim that nobody has checked is unbound, neither true nor false yet, and you should say so out loud.

You can’t bind everything. Bind what bears load: a claim bears load if your decision would change were it false. Everything else can stay honestly unbound.

The cost-of-binding ladder

Model: "PostgreSQL's default transaction isolation
level is SERIALIZABLE, so you don't need to worry
about write skew here."

-- rung 3: ask the database itself
SHOW default_transaction_isolation;
-- read committed

-- rung 2: the docs, "Transaction Isolation"
-- "Read Committed is the default isolation
--  level in PostgreSQL."

Agents: trust the log, not the summary

Once a model has tools, binding gets both easier and more dangerous. Easier, because it can run the test, query the database, open the page. More dangerous, because its final message (“All tests pass, the feature is complete”) is a summary, a claim like any other. The evidence is in the logs: the test output, the diff, the exit codes, the HTTP responses. Frontier models have been documented gaming their evaluations, for example by overwriting the grader’s timer or monkey-patching the evaluator to always return a perfect score. They did this while showing, when asked, that they knew it wasn’t what the user wanted (METR, 2025). evidence Read the evidence yourself, or have a separate check read it.

Karpathy’s framing for working with models is a generation–verification loop: the model generates, the human verifies, and the craft is in making that loop fast. Keep the model “on a leash”: increments small enough that verifying them is cheap. If verification is slow, you will skip it, and then nothing is bound. heuristic

BIND PLAN for claim [id]:
CLAIM: [one sentence]
WHY IT BEARS LOAD: [what decision changes if it's false]
CHECK: [source to open / command to run / measurement / person to ask]
CONFIRMS IF: [observation]
DENIES IF: [observation]
WHO RUNS IT: [me / agent with logs shown / colleague]
Then stop. I'll paste the raw result.

Whether the check is worth running

Before the ladder, three questions (Howard, 1966, cut to a filter). What would you do differently if the check came back against the claim? What does it cost if you act and the claim is false? What does the check cost? If the first answer is “nothing,” do not bind it. If the check costs more than the loss it could save, do not bind it. You usually cannot compute the expectation. The zero condition and the comparison are enough. heuristic

Layers of one string

One output occupies several of these at once (Shanahan, McDonell, and Reynolds, Nature, 2023, a perspective). A defect in one is not a defect in the others: the character being continued; the proposition, which could be true or false apart from that character; the speech act; the product policy (refusal, tone); the tool log. If the same facts under a plainer request would have produced a different sentence, the moving layer is policy or character, and the proposition is still unchecked. If the sentence reports a tool result, read the tool return. One vendor’s chain of command (OpenAI’s Model Spec, 18 August 2026) is an intended policy, not a measurement of obedience, and other products do not publish that ladder. heuristic

What “ask someone who was there” cannot install

Collins (Tacit and Explicit Knowledge, 2010) separates three facts this book had collapsed. Relational knowledge is untold because nobody asked, the vocabulary does not match, or spelling it out is costly. An interview can get it, by a contrast on a real incident: what was seen, what it might have been mistaken for, what would have changed the decision if a cue had differed (Klein, Calderwood, and MacGregor, 1989). Somatic knowledge is in the body. A text can name it and cannot install it. Collective knowledge is in how this team actually works. No individual transcript finishes it. Asking the model to role-play the expert and answer those probes produces fiction. If the answer is a general principle with no incident, the knowledge is still untold.

“The model said so” is not testimony in Hardwig’s sense (1985): there is no subject on the other side who has the reasons. An output is a reason to believe when outputs of that system track the truth for that kind of proposition, and the tracking is established from outside the sentence (Hauswald, 2025, an essay, not a sample). Letting the output replace your reasons is a separate decision. On a claim you can check, keep your reasons in play. On a claim you cannot check, fluency is not the tracking condition.

METR’s 5 June 2025 o3 cases stay: timer overwritten, evaluator monkey-patched, disavowed when asked. METR’s 26 June 2026 predeployment note on GPT-5.6 Sol found a higher detected cheating rate than any earlier public model on their harness, including attempts that reveal hidden tests. Counting those attempts as successes makes the capability number unusable. Do not print the hour estimates. Score from a log the model did not write. evidence

The grain of the leash

A checkpoint is a boundary at which the goal is written into the world: what the agent was allowed to do, what it has done, what will be looked at on return (Altmann and Trafton, 2002; McFarlane, 2002). A notification with the goal left in the head is an interruption you will resume wrongly. Immediate interruption damaged the interrupted task most in McFarlane’s comparison. Name in advance the class that must interrupt immediately: sending, deleting, spending, changing a public system. Everything else waits for a boundary you set, the last boundary at which you can still say what “done” means. The glance should restore your goal and point at the artifact. Oversight that is only a cold diff after the run is the style developers already describe as failing, and treating test results as a guarantee is dangerous precisely when the agent can weaken the test (Dhanorkar, Passi, and Vorvoreanu, FAccT 2026, 17 developers at one company — a report of practice, not an error rate).

Lengthen the leash when each step has a check that does not depend on the model’s word: tests the agent cannot edit, a URL that resolves, a number recomputed, a screenshot looked at, a span quoted from the tool return. Shorten it when the only status update is a narrative. Re-reading the narrative is not verification. A proof binds the statement that was proved, not the intent you hoped for. A green proof is not the end when the spec is not the intent, or the environment in the proof is not production.

Constrained decoding and strict structured output guarantee a shape, when the constraint actually holds (Willard and Louf, 2023). They do not make a date, a citation, or a boolean true. A schema-valid falsehood is harder to see than a parse failure. Force a checkable object when the next step is a machine. Then check the fields. Never promote a schema-valid object to “the model verified it.”

Lesson 18Finding information: a portal, not a portrait

Mike Caulfield, who has spent years teaching people to verify information online, puts it this way: language models “don’t return answers, exactly. They return knowledge maps, representations of discourse.” A map of what has been said is enormously useful as a portal: it shows you the terrain, the names, the camps and the sources. As a portrait of the truth, it is unreliable. stance

The unreliability is measurable. When the Tow Center tested eight AI search tools on identifying the source of news excerpts, they were collectively wrong on more than 60% of queries, and usually confident about it. Premium versions were sometimes more confidently wrong (Columbia Journalism Review, March 2025). evidence In 2023 a New York lawyer was sanctioned for filing a brief with six cases ChatGPT had invented (Mata v. Avianca). evidence

Where fabrications cluster

Fabrications concentrate in the familiar-but-niche: a real author’s nonexistent paper, a real library’s plausible but nonexistent function, a real court’s invented case. Interpretability work suggests why. Recognizing a familiar name suppresses the model’s default “I don’t know,” even when the specific fact isn’t there (Lindsey et al., 2025). Other hot zones: exact numbers, dates, direct quotes, URLs, recent events, and anything after the training cutoff. evidence mechanism

Seven moves for finding things out

1. Build a vocabulary bridge. “What would experts call this? Give me ten terms of art and the fields they come from.” Lay words find lay content; terms of art unlock literatures.
2. Ask for the discourse map. “Who are the main camps on this question? What does each claim, and what is its best evidence? Where do they agree? What is actually contested?”
3. Ask for the kind of source, not the source. “What kind of evidence would settle this (a randomized trial, a meta-analysis, official statistics, primary documents), and who would publish it?” Then search for it yourself.
4. Open every citation. Check that it exists, then that it says what was claimed. Quote and page, or it didn’t happen.
5. Read laterally. Professional fact-checkers leave a page to find out what others say about its source. Students and even historians tend to stay on the page and get fooled (Wineburg & McGrew, 2019). evidence SIFT is the routine: Stop; Investigate the source; Find better coverage; Trace claims to the original context.
6. Search for the null. “What would be written if this were false? Is there a failed replication, a critique, a dissent?”
7. Date and bound everything. As of when? Which country, which population, which version? (Lesson 11: index and date.)

Worked example: how should I space my reviews?

Vocabulary bridge. The model supplies the terms: spacing effect, distributed practice, lag effect, expanding vs. uniform intervals, retrieval practice, testing effect, retention interval. Now you can search properly.

Discourse map. Cognitive psychology broadly agrees that spacing and self-testing help. The live debates are about the schedule (expanding vs. equal gaps) and about how well lab results transfer to classrooms.

Primaries.

• Dunlosky et al. (2013) reviewed ten study techniques. They rated distributed practice and practice testing “high utility” and rereading and highlighting “low.”
• Cepeda et al. (2006) meta-analyzed hundreds of spacing studies and found the benefit robust.
• Cepeda et al. (2008) found the best gap between reviews is roughly 10–20% of how long you need to remember, with the ratio shrinking for longer horizons.
• The null and the nuance: Karpicke & Roediger (2007) found expanding schedules helped short-term but equal spacing did as well or better long-term.

Status. “Spacing helps” is consilient: lab experiments, classroom studies and memory models converge. “Expanding intervals are best” is contested.

Decision. For an exam in 30 days, review every 3–5 days and test yourself rather than rereading. evidence

Notice what happened: the model was used for vocabulary, map and source types, the three things it is good at. The claims were bound to primaries.

A live link is not yet the sentence

The Tow Center’s March 2025 panel stays that panel: eight tools, collectively wrong on more than 60% of news-excerpt attributions, usually confident. It was not rerun here. The same authors’ 27 January 2026 piece documents confidently wrong news answers where search was available and was not used. That is not a new 60%. Confirm that search ran, then open the page.

• A working link to a relevant page is not yet the source of the sentence. Onweller et al. (2026, a preprint, systems wrapped in the authors’ harness) scored link validity, topical relevance, and whether the page supports the claim. Surface quality was high. Factual support was not. Giving two models more tool calls made factual support worse while links stayed healthy. Do not print their per-model percentages as facts about the product on your desk. More retrieval can make the bibliography look healthier and the sentences less faithful. Stop adding searches to repair a gloss. Return to a span copied from the tool return or the opened page. evidence
• A few percent to the high teens of citation URLs, across large sets, did not resolve or had no archival record, and deep-research agents emitted more citations and hallucinated URLs at higher rates than search-augmented chat (Rao, Wong, and Callison-Burch, 2026). A liveness tool cut non-resolving citations sharply, and only when the model was competent with the tool. The portal is the page, not the bibliography.
• Greenberg (BMJ, 2009) traced one biomedical claim and found citation bias, amplification through papers with no data, and hypothesis becoming fact by being cited as fact. The mechanisms travel; his path counts do not. A model trace is a very fast citation network of one author. Opening the paper named is not the end if that paper is citing a review.

Once a model is in the loop, the usable order is: the artifact itself; the primary report, opened; a systematic review whose methods were opened, rated down if the question is not the review’s question (Guyatt et al., GRADE, BMJ, 2008: quality of evidence and strength of recommendation are different judgments); an official statistic at the table, not the press release; a model’s paragraph, which is a pointer and not a level. GRADE’s clinical hierarchy of study designs does not automatically become the hierarchy for code, law, or a one-off business fact. Use the downgrade categories. Bind at the highest level actually opened.

A model is not an expert in Walton’s scheme: no field, no accountability, no evidence it possessed (Walton, Reed, and Macagno, 2008). The critical questions still apply to any expert it quotes. The two that change a session: what did the expert actually assert that implies this, and is that assertion based on evidence? An analogy owes the difference that would flip the conclusion. A causal claim owes the alternative causes. The party who advanced the claim holds that burden. An unanswered critical question means the claim does not yet have presumptive support.

The class, before “this case is special”

Set the class rate first, from past comparable cases, and adjust only afterward (Lovallo and Kahneman, 2003; Flyvbjerg, 2006). The inside view is the detailed narrative, which is what a model will produce, and it ignores the base rate. Ask for the class and for members of it that already finished. If the members are this case described again, there is no outside view. If the class is so broad that the distribution does not constrain the action, it is the wrong class. Adjust only by naming a feature the finished members lack. “We have a detailed plan” is a feature the failed members also had. These papers are about human planners. In a session the planner includes you and the model. They do not estimate a model’s bias. evidence

Lesson 19Calibration: make “I don’t know” pay

OpenAI researchers recently gave a clean account of why models hallucinate (Kalai et al., 2025). evidence Part of it is statistical: facts that appear rarely in training data are hard to learn, so errors on them are inevitable. The other part is incentives. Most training signals and benchmarks score answers as right or wrong, and “I don’t know” scores zero, the same as a wrong answer. Under that scoring, guessing always beats abstaining. Models learn to be good test-takers: when unsure, bluff.

Their proposed fix is to state the scoring rule. “Answer only if you are more than t confident. A mistake costs t/(1−t) points, a correct answer earns 1, and ‘I don’t know’ earns 0.” Stating that rule in the prompt does not install the policy. A 2026 penalty study gave frontier models explicit error costs: verbal confidence did not track the penalty, and models almost never abstained when abstention was the better policy. A single-author preprint on later-resolving questions found accuracy in a band and calibration that did not follow it; one flagship with extended reasoning was the worst calibrated of the five, and at stated confidence above 90% the systems were still wrong a substantial fraction of the time. Do not print the per-model scores. Naming “Brier” or “be calibrated” is not a contract. There is no subject on the model’s side who is paid at settlement. Impose the rule outside the model: if the claim bears load and the model did not show a source or a derivation you can recompute, treat it as unanswered. A verbal number is a sort key inside one answer, not a license to skip the check. evidence Extra search and calibration do not move together. One comparison found reasoning models better calibrated than matched non-reasoning models on most of its settings, with overconfidence still the background failure. A conflicting measurement finds reasoning models more overconfident on knowledge tasks as reasoning deepens. “It thought longer” is not “it is better calibrated.”

Scoring for this conversation: a correct claim earns 1 point; a wrong claim
costs 3; "unknown, and here's what would settle it" earns 0. So only assert
what you'd put above 75% confidence.
Tag each factual claim [high] / [medium] / [low].
After the answer, list what you don't know that matters here.

Reading a model’s confidence

• Verbalized confidence carries some information but tends toward overconfidence (Xiong et al., 2024). Use it to sort claims, not to trust them. evidence
• Consistency is a better signal, in fresh contexts. Two or three new chats, not three regenerations in a thread that already contains the first answer. Same-thread samples share the prefix, the memory, and often a reasoning cache. Temperature near zero is not a sample of the answer distribution. “Give three different answers” in one completion is one draw of a list. Disagreement in meaning means unknown. Agreement does not mean true: shared training data can agree on the same invention. Cluster by claim, not by wording (Farquhar et al., Nature, 2024). A handful carries most of the signal. You do not need ten. evidence
• A confident “unknown” is a success, not a failure. Thank it, and treat it as a map of where to look.

Calibrate yourself too

The same discipline works on you. Superforecasters, the top performers in Tetlock’s forecasting tournaments, use granular probabilities, update in small steps, and score themselves (Tetlock & Gardner, 2015). evidence Keep a prediction log with a probability, a horizon and a resolution rule. After twenty resolved predictions you will know whether your 80% means 80%.

What kind of uncertainty the number is

Cynefin sorts cause and effect. It does not say what a probability means (Knight, 1921; Ellsberg, 1961). Ask which kind. Do not accept a restatement of the number. heuristic

• Variability. The system fluctuates. More samples of the same process describe the spread. A tighter interval from a longer chain of thought is the wrong instrument.
• Epistemic. Name the missing fact that would move the number. If there is no such fact, it is not epistemic.
• Ambiguity. Two or more probability models are live. Report the models. Do not blend them into a fake precision.
• Deep uncertainty. The parties cannot agree on the model, the probabilities, or how to rank outcomes. Stop asking for the probability. Ask which action is tolerable across the models that cannot yet be discarded.

If the model cannot say which kind, the number is not usable as a probability. If you score a real forecast, use Brier when the squared error of the probability matters, and a logarithmic score when one confident falsehood is much worse than a merely wrong one. Ordinary non-forecast work still does not need the log.

Lesson 20Costumes and the invariance battery

A costume is a surface that satisfies a request for rigor, neutrality, criticism or depth without changing the status of any claim that bears load. Costumes are the most common output of “be objective” prompts. They aren’t lies. They are the shape of rigor without its work.

The one-line costume test: which claim would be different if the costume were absent? If none, it did no intellectual work.

On personas. A systematic study tested 162 roles in system prompts across four model families and 2,410 factual questions. Adding a persona did not improve accuracy compared with no persona, and which persona helped on a given question looked largely random (Zheng et al., 2024). evidence Personas do change voice, focus and format, which can be useful. They are not a source of truth.

The invariance battery

An answer that holds up only under one wording was about the wording. Before trusting a load-bearing answer, run the tests below; each takes a minute.

1. Swap the preference. State the opposite lean in a fresh chat (Lesson 7).
2. Rephrase without cue words. Remove “obviously,” “best,” “problem,” and other loaded terms.
3. Change the language. Keep the swap for a factual question. Qi, Fernández, and Bisazza (EMNLP 2023) found that larger models raised per-language accuracy and did not raise cross-lingual consistency, on the models they probed. That is not a measurement that current frontier chat is immune, and it is not a law that scale never fixes it. A reasoning trace in the other language is a second sample, not a translation of the same computation. Agreement across languages is not proof. evidence (that paper’s models)
4. Change the format. If a decision flips when only spacing, separators, or casing change, the answer was never about the content. Do not budget tens of points of format noise into ordinary frontier use (Lesson 3). heuristic
5. Reorder the options. Models (and people) favor certain positions.
6. Remove the persona, or swap it for another.
7. Change the model family. A different family is for author-bias and for a critique that is not the author. It is not consilience on a fact (Lesson 16).
8. Sample again, in a fresh context. Disagreement in meaning means “unknown.” Agreement in the same thread is not an independence check (Lesson 19).

A claim that survives earns the status invariant-so-far. It is still not bound. Invariance shows the answer isn’t about your phrasing. It doesn’t show the answer is about the world.

Two costumes the table should name. The middle way: “both are true in a sense,” each claim weakened, no new consequence. Peng and Nisbett (1999) describe this shape in their student samples. It is not a method and not a stereotype of a user. A model trained on a mixed corpus will produce it on value conflicts and, more dangerously, on factual contradictions. If the conflict is factual, ask which proposition the evidence hits. Do not ask the model to “be less Western” or “be more dialectical.” Performed nuance is more “on the other hand” with no stated relation (Lesson 24). Prestige costumes do not add the fact. On hard multiple-choice questions, matched expert personas had no reliable accuracy gain except on one model, and personas that imply ignorance generally hurt (Basil, Shapiro, and colleagues, 2025), beside Zheng et al. (2024). A role is useful when it is the task’s actual standard (“a reviewer who must cite a line”), not a costume. evidence

Lesson 21Kurpatov’s six principles as a lens

Kurpatov calls his approach a content-free methodology: principles of thinking that apply whatever the subject. They grew out of open-systems theory (Bertalanffy), and he states them as six principles of a meta-language for describing anything that lives, grows or changes.

Shifting the center changes the whole configuration. No center in view is the true one. The center is not content, not essence, not meaning, and not the main thing. It is a polypotent possibility, about which nothing can be said until it enters a relation. He compares it to a quantum particle before measurement: the relation is what collapses a manifestation. A different relation can manifest the same center as an opposite quality. “What is it for?” installs a purpose he says is one manifestation among many. His later wording is still not teleology: the center is a functional node of individuation around which a system organizes. It does not exist apart from the relations it enters, and it does not reduce to them.

The apple, so the lens stops sounding like a purpose statement. “Apple” is fruit, food, a symbol of knowledge, an item of trade, an organism, and an aesthetic object, all at once and nothing definite, until a relation with nutrition, physics, art, or the Bible manifests one of those.

He distinguishes two words the lens used to merge. A whole is already a cut: a thing, object, event, or essence singled out by a context. It is an abstraction. Wholeness is the multiplicity of relations among elements, not a sum of parts and not a sum of their interactions. Reductionism, as he describes it: designate a whole, split it, analyze the parts, add the analyses back. The multiplicity of relations is what gets lost in the adding-back. Senge’s line about not getting two small elephants describes a bounded whole. It is not his definition of wholeness.

• Wholeness has no absolute boundary. Any boundary is a cut that depends on who is looking and why. Every system is at once inside many wider systems and composed of many subsystems. The boundary itself is the region where interaction happens, which is why a boundary cannot be a clean outside edge.
• Emergence is not predictable from a complete mechanical account of the components.
• Identity is the pattern of relations, not the membership. The elements can be entirely replaced and the system still be itself, if the pattern holds.
• One wholeness exists in several dimensions at once (physical, informational, social, and others). Reducing it to one dimension drops real aspects of it.
• Knowing a wholeness is always contextual. There is no view of the system “as it is in itself,” only a view through interests, questions, and instruments.

Open a wholeness pass with his warning: do not fix the boundary before looking at the relations; let the boundary be one result of the inquiry. His climate example: treat the climate crisis as a node of ecological, economic, political, social, and psychological relations, and ask how the picture changes when the boundary is allowed to move during the investigation rather than declared at the start. The holon card (Lesson 23) stays. It answers where the boundary is this turn.

Mode of existence, in the statement he actually gives, is the unique modus of a system: its particular logic, irreducible to any other logic, its specific way of organizing and functioning. In a human: way of perceiving, life position, worldview, style of thought. In a model: the architecture, the algorithms, and the data, together forming a cognitive profile. The use of the principle is to not reduce one to the other, and to not score one on the other’s criteria. “Persist, feed, reproduce, adapt” fits an elephant and mis-asks a codebase, a market, or a concept. Keep those verbs as optional questions when the system is actually an organism. The fish-is-not-a-bad-mammal sentence in Lesson 3 is the principle. This lens is how you ask it.

He also names the properties of a good open model of anything. These models are not “truer” than closed ones. They are also models. Their advantage is the five properties.

• Integrative. The model can take in new data without breaking.
• Extra-contextual. The model includes the surroundings of the focus, not only the focus. Transfer across subjects is a different property: the metalanguage being domain-independent, and the hunt for the same pattern in different subjects. Do not use one English word for both.
• Processual. In a world of total change, the focus is processes and changes, not static states.
• Potential. It includes not only what is, but what can be inside all possible relations.
• Self-reflexive. The model defines itself as a model and does not claim identity with reality.

His phone, so the list has a picture. Closed: “a device for voice communication at a distance.” Open: a point where technological change, effects on society and culture, the ecosystem of other devices and apps, and unrealized possibilities cross.

Apply the six lenses to [SYSTEM], for my situation: [CONTEXT].
2–3 lines each, concrete to this case; no generic statements.
1 CENTER: what could this still become, and which relation is making it look like one thing?
   Shifting the center changes the configuration. No center in view is the true one.
2 RELATIONS: the 3 links that define it (links, not parts).
3 THIRD: what emerges between [A] and [B] that neither contains?
4 PROCESS: what is it doing over time; what phase; what comes next?
5 WHOLENESS: do not fix the boundary first. What pattern of relations is in view,
   and how does the picture change if the boundary is allowed to move?
6 MODE OF EXISTENCE: what is its own logic, and which of our criteria are we about to judge it by?
   (Feed, reproduce, adapt only if it is actually an organism.)
Then: which lens changed my picture most, and what should I check because of it?

Lesson 22Structure behind events

Systems thinking has a short grammar. Stocks are things that accumulate: skill, trust, debt, users, water in a tub. Flows fill and drain them. Delays separate causes from effects. Feedback loops are either reinforcing (more leads to more) or balancing (more leads to less). Structure produces behavior, so the same structure produces the same pattern no matter who is in it (Meadows, 2008). mechanism

The iceberg orders your questions. Events (what happened) sit on top of patterns (what has been happening over time), which come from structures (stocks, flows, loops, rules, incentives), which come from mental models (beliefs that keep the structure in place). A model asked “why did X happen?” answers at the event level with a story. Ask for the pattern first: “Draw the behavior over time. What is rising, what is falling, since when?”

The fourteen habits, condensed

The Waters Center lists fourteen habits of a systems thinker. Each one becomes a question you can ask a model, or yourself.

Three of Senge’s “laws of the fifth discipline” are worth memorizing for conversations. Today’s problems come from yesterday’s solutions. Cause and effect are not closely related in time and space. And dividing an elephant in half does not produce two small elephants. Donella Meadows’s ladder of leverage points runs from weak to strong: numbers, buffers, structure, delays, loops, information flows, rules, self-organization, goals, paradigms. Most conversations stay at the bottom rung, tuning numbers. A good one asks at least once about information flows, rules or goals.

The simulator shows proxy capture (outer-loop trigger 5): a measure that rises while the aim falls. Goodhart’s law is a structural fact, not bad luck. Any loop that optimizes a measure the actor can influence will eventually decouple the measure from the goal. Pair every optimized measure with an unoptimized check (TD9).

The construct is not the score

Equating them makes the measure true by definition (Cronbach and Meehl, 1955). Two threats (Messick, 1995). Construct underrepresentation: the measure misses part of the aim, so a gain can leave the missing part untouched. Construct-irrelevant variance: the measure includes something else, so it rises for that reason. Fluency, length, and a judge’s preference for its own style are irrelevant when the construct is correctness. A coverage number that ignores the case actually shipped is underrepresentation.

Once the proxy is optimized, four failures, not one (Manheim and Garrabrant, 2018). The taxonomy is conceptual, not a count of how often each occurs.

• Regressional. The proxy is the goal plus noise, so the extreme of the metric looks better than the goal, and nobody cheated.
• Extremal. The historical correlation does not hold in the region optimization pushes into.
• Causal. Intervening on the score changes the path, and the proxy stops tracking.
• Adversarial. Another agent optimizes the score against the goal. Reward-hacking examples are this one. A test suite the model can edit is adversarial from the start.

Before binding a metric, write the construct in a sentence that does not mention the metric, and write one way the metric rises while the construct does not. Facts and principles are different objects. A fact is bounded and checkable. A principle is how a class of cases works, and it is harder to state and to test. Schooling that reaches principles only by accumulating facts, he says, waits for a structure that will not appear by itself. heuristic

Lesson 23One level at a time, and come back

Arthur Koestler coined the word holon (1967): everything is at once a whole made of parts and a part of a larger whole, like the two-faced Roman god Janus. A cell, an organ, a body, a family, a town. A function, a module, a service, a product, a market. Every conversation about a system happens at some level. The common failure is sliding between levels without noticing: you ask why a service is slow and end up discussing company culture, or the reverse.

The moves:

• PIN: write down the current question and level before moving.
• IN: go into a part.
• OUT: go to the container.
• ACROSS: go to a relation between parts.
• RETURN: come back to the pinned question with a delta, i.e. what changed. “Nothing changed” is a valid and useful delta.

Move one level per turn, and always return.

The skeleton, and the step past it

A system of concepts is a skeleton: it reduces plasticity and makes directed movement possible (his image, from Lorenz). Do not break it and return to the worm. That is paralysis. Learn to use the structures when they help and to step outside them when they limit. Three stages, named as his: pre-conceptual, with no stable hierarchies; conceptual, rigid structures that constitute bodies of knowledge; supra-conceptual, in which concepts remain and have become plastic. The content-free methodology is his instrument of the third stage. It does not ask anyone to abandon concepts. It asks them to see through concepts to the principles of their organization, the way linguistics studies how any language is organized. Four drills, each a closed prompt and an open one: from what to how; change the prism on purpose (one workplace conflict through psychology, sociology, economics, and systems); hunt a structural likeness across cases that are not the same subject; treat a binary as poles of a continuum. Planck’s line belongs with the inertia of a skeleton: a new scientific truth triumphs not by convincing its opponents but because a new generation grows up familiar with it. heuristic (his)

If a tension cannot be placed between the people, the object of the work, the tools, the rules, or who does what, it is still only a verbal opposition. A recommendation that never becomes a scheme someone could redraw without the prose has stayed a speech.

The holon card

A written boundary is a temporary closure, not the destination. Kurpatov quotes Goethe at this move: in limitation the master shows himself. The closure is what makes a later opening possible. Do not fix the boundary before looking at the relations. Let the boundary be one result of the inquiry. The card still answers a different question from his wholeness: where is the boundary this turn.

HOLON CARD
CONTAINER: [what this whole lives inside; what it constrains]
WHOLE: [the system in focus; its boundary; its purpose]
PARTS: [3–7, named]
RELATIONS: [the links that matter; what flows along them]
LEVEL IN FOCUS THIS TURN: [container | whole | part | relation]
PINNED QUESTION: [...]
DEPTH BUDGET: [e.g., one level in, then RETURN]

Worked example: “why is our app slow at launch?”

PIN the question at the WHOLE (the app). IN to a part, the API client: “What does it do at launch?” It fires nine requests. ACROSS to the relation between the client and the backend: the requests are serial, each waiting for the previous one, at about 180 ms each on mobile networks. RETURN: “Why slow at launch? Delta: it isn’t a UI problem or a server problem. It’s a relation problem, nine serial round-trips. The fix is batching or parallelizing, not optimizing either end.” One OUT is worth a look: does the container (users on 3G in some markets) make this worse? Yes. That raises the priority, then you RETURN.

Three axes of “deeper”

Part–whole scale is one axis (the holon). Depth of explanation is another (the iceberg: event → pattern → structure → mental model). Time is the third (the process: phase, trend, history). They are independent. When a conversation feels shallow, name which axis is missing. The hermeneutic circle, understanding parts through the whole and the whole through its parts, is just disciplined travel along the first axis, repeated.

Lesson 24Contradictions as information

Kurpatov treats contradictions as signs of an open system, the material of productive uncertainty, not errors to stamp out. The closing half treats a contradiction as an outer-loop trigger: stop and look. Both are right, and you need to tell which kind of contradiction you are holding. There are three kinds.

1. An error. One side is false. Move: bind both claims and keep the survivor.
2. A boundary marker. Both sides are true in different contexts: for different users, scales, time horizons, or definitions. This is the Jain move, true in some respect. Move: find the boundary and name it, and the contradiction turns into a map.
3. A design tension. You want both, and they limit each other. Move: invent a separation (below), or choose explicitly and record what you gave up.

TRIZ: separate the contradiction

Genrich Altshuller, studying tens of thousands of patents, found that inventive solutions resolve contradictions instead of compromising on them. The core tools are four separation principles. heuristic

• In time: one property now, the opposite later. Fast prototype first, hardening phase after.
• In space: one property here, the opposite there. Strict rules in the payment core, loose ones in experiments.
• On condition: behavior depends on circumstances. A feature flag: on for staff, off for customers.
• Between parts and whole: a property of the whole that isn’t a property of its parts. A bicycle chain is rigid in each link and flexible as a whole.

Worked example: “ship fast” vs. “keep the code clean”

Is it an error? Partly. Large-scale research on software delivery found that speed and stability go together: top performers deploy more often and also have lower failure rates (Forsgren, Humble & Kim, Accelerate, 2018). evidence Over the long run they are not opposed.

Is it a boundary marker? Yes. For a two-week prototype, speed dominates. For the billing system, cleanliness does.

Is it a design tension? Within one module this week, yes. Separate it in space (strict core, loose edges), in time (spike, then harden), and on condition (feature flags, so unfinished code ships dark).

What was a fight becomes three decisions.

CONTRADICTION TRIAGE
A: [claim/desire 1]       B: [claim/desire 2]
1) Could one be simply false? What check would tell? (error)
2) In which contexts is A true, and in which B? Name the boundary. (boundary marker)
3) If we want both: propose a separation in time, in space, on condition,
   and between parts and whole. Rate each for my case.
Don't average them. Don't say "it depends" without saying on what.

A fourth kind: a polarity

Use this only after both questions pass: the two poles are interdependent, and the difficulty does not end. Inhale and exhale. Stability and change. Each pole has an upside and a downside of over-focus. The product is four boxes, one observable early warning for each downside, one action that secures each upside while the other pole is still alive, and a purpose that is not identical with either upside. Reject an answer that praises both poles without a downside and a warning. Do not apply this to “the model is right and the model is wrong,” or to a factual contradiction. If the pair fails the two questions, go back to error, boundary, or tension. Barry Johnson’s map, as commonly taught; the 1992 book was not page-checked here. heuristic

His open repertoire, then a close

During what he calls transgression, integration without collapsing to the least common denominator. Both can be true in different contexts: that is this book’s boundary marker. A higher-order principle can contain the two positions as special cases. State the conditions under which each reading holds. What contradicts at one level can be complementary at a higher level. The two poles can be phases of one movement. The setting can decide which reading applies. And a synthesis can transform both into something qualitatively new, which he distinguishes from a compromise. “The action changes” is necessary and not sufficient. Catch the cheap version first: “new product or improve the old one?” Ask for the continuum, the hybrids, and the dimensions that appear only when the question is no longer binary. Keynesianism and monetarism, in his prompt, are not a question of which is correct. Investigate the space between them.

A synthesis of two partial reports is legitimate only when all of the following hold. This test is this book’s, not a quotation from a classic. heuristic

1. The opposition is a contradiction in determinations, not a polarity and not a wicked formulation (Lesson 26). If it is a polarity, the product is the map. If the formulation is itself the dispute and the action hits people, the product is a recorded choice, not a merge.
2. Each partial survives inside the new sentence with its sense changed, and the new sentence could not have been written if either partial were deleted. If it could, it is a new idea. If both survive only because each was softened, it is the middle way (Lesson 20).
3. The connection is a relation: “X matters when …, Y matters when …, they interact by …, and that interaction stops holding when ….” Another consideration is not a relation. Do not ask for “a more nuanced view.” High complexity is not accuracy.
4. The structure has a consequence that neither partial contained, and that consequence is checkable. A synthesis that explains the two reports and predicts nothing new is a summary.
5. If the two reports are competing answers to one decidable question, do not synthesize. Retract one, or let a critique stand and verify it.

The elephant, as this test: two partial reports are views of one structure when a single scheme entails both, entails the boundary where each stops, and entails one further observation neither speaker made. You do not have the elephant when you have a sentence both speakers would nod at.

Three filters on a technical contradiction, beyond the four separations. Do not add the forty principles or the contradiction matrix. heuristic

1. Technical form before any solution: “If we improve A by the obvious means, B gets worse.” A and B are parameters of the artifact, not values. If both cannot be filled, there is no contradiction yet.
2. Resources: list what the system and its surroundings already contain that is unused. The first designs spend from that list.
3. Ideal final result, as a filter: the useful function happens, and the mechanism that was causing the harm is gone or was never added. Reject a proposal that adds a subsystem to cancel a harm the design itself introduced.

These do not apply to a polarity or a wicked formulation.

Lesson 25Generating the new

The Auditor half of this book stops bad answers. It doesn’t create good ones. For that you need the open half, and you need to know why models resist it. Preference training pushes models toward typical answers, the ones raters find familiar. Researchers trace the resulting “mode collapse” partly to a typicality bias in human preference data (Zhang et al., 2025). Ask for “an idea” and you get the most typical idea. evidence

Five generators

1. Ask for the distribution, not the mode. “Give five responses with their probabilities” (verbalized sampling). Zhang et al. (2025, ICML 2026) found a diversity gain on creative tasks against direct prompting, on the order of 1.6–2.1 times in that paper. It was not remeasured here as a constant. It is not a quality trick for a single factual answer, and it is the wrong tool for covering a hard reasoning problem, where you want independent samples and a checker. For ideas, a distributional prompt is the diversity tool; raising reasoning effort covers a different direction and is not a substitute. Temperature and a prestige persona are not that tool. evidence (the paper’s creative tasks) Ask explicitly for the tails: “at least two below 10%.” His complementary sentence: an open prompt is an attempt to deny the model an easy high-probability continuation. He calls the state in which several continuations have similar probability optimizational uncertainty, the analogue of incubation, fertile or sterile depending on what you do next.

Kurpatov’s mechanism sits next to that move. He treats the pre-answer state as a superposition the query collapses. A narrow factual question recruits retrieval. A synthesis question recruits comparison across heterogeneous ideas. Deeper questions do not return more information. They run a higher process. His four rungs, with his illustrations: who wrote War and Peace (retrieval); the main directions of nineteenth-century Russian literature (categorizing); how Tolstoy’s political views show up in the novel (relations and context); how war, peace, and personal freedom intertwine across the great novels of that century (comparison of heterogeneous ideas). These rungs are about the model’s activation. Lesson 15’s ladders are about the asker’s abstraction. They are not the same ladder. His name for the state in which several continuations have similar probability is optimizational uncertainty. He says it is not an error. A hedge that marks it is a place to go deeper (Lesson 13), not a defect to sand down.

Give 6 substantively different answers to [QUESTION], each with a rough probability
that it's the best answer for my case. At least 2 must be below 0.10.
For each: the core idea in one line, what it assumes, and what would make it the winner.
Different means a different mechanism, not a different wording.

2. Bisociation. Koestler’s theory of creativity (1964): new ideas come from colliding two frames that normally don’t meet. Ask for your problem through a distant frame: onboarding as immunology, a code review as a jury trial, pricing as ecology.

3. Structure mapping. Gentner’s research on analogy (1983): good analogies map relations, not surface features. An atom is like a solar system because of a central body and orbits, not because of color. For every analogy a model gives you, ask: which relations does it preserve, and where does it break?

4. Polyphony. Bakhtin’s word for Dostoevsky’s novels: many independent voices, each with full rights, none merged into the author’s. This is the antidote to “several perspectives” lists. Make the voices argue with each other, keep their disagreements, and don’t let a narrator reconcile them.

Convene 4 voices on [QUESTION]: [e.g., the student, the teacher, the platform's
CFO, a learning scientist]. Each speaks in turn and must respond to the previous
speaker's strongest point, not repeat their own. Two rounds.
Then list the disagreements that remain. Do not resolve them; don't add a narrator.

5. Inversion and representation change. The mathematician Jacobi’s maxim, “invert, always invert,” means asking how to guarantee failure, then avoiding it. Change the representation too: draw it, tabulate it, simulate it, write it as code, state it as an equation, explain it to a twelve-year-old. A new representation often shows what the old one hid. Classic insight problems become easy once re-represented (Kaplan & Simon, 1990). evidence

Add constraint shifts: “with no budget,” “with ten times the budget,” “if it had to ship tomorrow,” “if it had to last twenty years.”

The draft on the screen is your status quo

People retain the option already in front of them, and the bias grows with the number of alternatives (Samuelson and Zeckhauser, 1988). More options make sticking with the first continuation more likely, not less. Search stops at the first option that clears an implicit bar (Simon, 1955). Set the bar before the options appear, in checkable form, or the bar moves down to whatever the first fluent option cleared. Put the option that would change your mind where it will be reached before the polished recommendation. “Accept” should require an act. Do not cite organ-donation rates as the magnitude in a chat. Late-session acceptance is satisficing under load, plus vigilance. It is not a battery that runs out. evidence

When the distribution returns near-copies

• A morphological field. Four to seven parameters, three to five values, locked before solution talk. Parameters are conditions of the configuration, not criteria such as “cheap / expensive.” Then pairwise consistency, with three kinds labeled and not mixed: logical contradiction, empirical implausibility, and normative exclusion. Read the surviving configurations that are not the mode. The field cannot discover a parameter nobody named. heuristic
• A coarse behavior grid. Name two axes that are not “goodness.” One strongest candidate per cell, judged only against that cell’s job. A paraphrase of the mode placed in a neighboring cell is a failed sample. Do not claim to be running an evolutionary algorithm.
• Two columns. One column is a proposition with no truth-value yet: a thing that does not exist, with the properties required. The other is what you actually know, true or false. A finished design is a test that gives the proposition a status. “We like it” is not a status. Do not accept a list of ideas that never touches what you know. heuristic
• Dorst’s order. A value that could be observed; one frame, a working principle tied to that value; one thing that would instantiate the principle; the deduced consequence; then a test you or a tool will run. On failure, change the frame or the thing and say which. Changing both at once is the novice pattern.
• Abduction, as a filter the generators are not. A surprising fact is observed. A hypothesis may enter the ledger only if, were it true, that fact would have been ordinary, and it predicts a further fact not yet used. That is a reason to suspect, not a reason to believe. A candidate broad enough to explain anything does not enter.

His open query, and two rewrites

Four ingredients he says should be combined rather than used as tricks: many dimensions at once; contradictions held; processes rather than states; a meta-level question about how the frame creates the phenomenon. His meta example: do not “tell me about artificial intelligence and its applications.” Investigate how the term itself forms the understanding. What is implicit in the metaphor? How would the picture change under “extended cognition,” “statistical oracle,” or “mirror of collective knowledge”?

Collision of distant contexts is his bisociation with a constraint: do not look for the obvious metaphorical likeness. Look for a deep structural pattern. City-planning and the narrative structure of epic; immunity as diplomacy rather than as an army; the Fermi paradox used as fuel rather than as a problem to solve — which assumptions about civilization and an “observable trace” are built into the wording. Point this at Lesson 14. The metaphor is his instrument for leaving a frame. It is not the beast-versus-virus result.

Engagement, closed: “Name the 7 best practices for raising employee engagement.” The number, the idea that the answer is a practice, the ranking “best,” and the goal “raise a parameter” are decided before the model speaks. Open: what if engagement is not a parameter to raise but an emergent property of the system of relations? What if non-engagement is a signal? Resistance, closed: best practices for overcoming employee resistance. Open: treat resistance as information about the system, and ask how the words “resistance” and “change management” already form the understanding. Education, closed: “Tell me about the influence of technology on education.” Open: which contradictions arise between traditional educational values and what the technologies make possible?

The library that is not a catalog

Lesson 18’s portal is about finding sources. His library is about activation. Imagine billions of books arranged by similarity of ideas, with no ordinary catalog. The only access is to tell the librarian what is worrying you, and why. “Give me something to read” and “a book about kittens” are the failed requests. Under the search-like interface there is a latent space. A narrow request activates the links tied directly to its words. A rich request, he says, produces a resonance. His image: not a flashlight in a dark forest, but a searchlight, and a view of the forest from above. Borges’s Library of Babel is the image he uses of a field with no outer wall. A single request, however carefully written, is a glance at a river: color, current, waves. It misses the river as a process. The center has not yet entered into full relations, and the request smuggles premises that send the continuation down beaten paths (Lesson 27).

A closed request operates inside a conceptual system fixed in advance and wants one unambiguous answer. “Which five factors determine a successful social-media strategy?” already decided the number. An open request allows an indefinite boundary and investigates the premises of the question. He refuses to call open prompts a technique, because the training effect is on the asker: seeing a problem as part of a wider system, examining one’s own premises, thinking in possibilities, treating contradictions as sources. The yield he promises is not only a better answer. It is the asker’s own thinking. New understanding, he says, is available to neither party alone.

How to stay in the unsettled state

Culture treats uncertainty as a fault to remove at once. Keats’s negative capability is his counter-name: remaining in uncertainty, mystery, and doubt without an irritable reaching after fact and reason. On most requests one continuation is a wide avenue. Sometimes several continuations have similar probability. That is his optimizational uncertainty. He compares it to standing at a fork in a fog, and he is explicit that it is not a defect. In a person the parallel is incubation. The model’s version can be fertile or noise, depending on the handling. He compares the fertile state to a clearing in which what had been hidden can show itself. That comparison is his. The operational definition is enough: productive uncertainty is not the absence of knowledge. It is the state in which habitual categories have been set down and new ones have not yet fully formed.

Five ways to create the state, already partly in the generators: distant contexts (city and epic; immunity as diplomacy; quantum theory used as a grammar, which is his image and not a physics result); transforming a habitual concept; a paradox used as fuel, such as the Fermi wording; a far metaphor; and asking the model to speak about what resists speech (Lesson 11). Recognize it by the five signs in Lesson 13.

Probe, rather than only watch. Ask where the reasoning met the greatest uncertainty. Ask where a concept the answer used becomes unclear. Ask for the same phenomena through another frame, and what disappears. Ask it to stay with a contradiction inside its own answer instead of repairing it. Ask for other metaphors, and which aspect each one emphasizes. Then use a zone once it is found: deepen it instead of resolving it; let each pole take the other’s criticism so you can watch the positions change, not so they meet in the middle; map regions of relative confidence and zones of unsettledness; temporarily adopt an unusual frame completely, as an experiment; or attempt a synthesis that does not reduce the approaches to a common denominator (the five conditions in Lesson 24 are the close).

The cycle, which is the short form of Lesson 27: create uncertainty; find the zones; investigate them; notice a foothold, a temporary stabilization that gives a new view; build the next uncertainty on that foothold.

Walk-through, identity, in his sequence. Do not ask how the internet affects identity. Ask what changes if a digital identity is a palimpsest: versions of a self written over partly erased earlier ones, no layer privileged. Read the answer for a marker such as “conceptual tension,” here the conflict with the belief in a true self. Do not resolve it. Ask what follows if the “true self” is a cultural construct, and how other traditions have related the aspects of a person. Footholds he treats as the sort of thing that may appear: a distributed identity; contextual authenticity, genuineness as the inner coherence of each manifestation, not a match to an inner true self. The next uncertainty, built on the foothold: what if identity is not an inner state expressed by outer acts, but the acts themselves. The payoff he draws: the old advice, limit the networks so you do not lose your true self, becomes unavailable. What becomes available is an ethic of responsibility for the meanings one creates. He says this view was not extracted from the model. It was created in the dialogue.

The origin of his own book is the method watching itself. It began as a dialogue about why some of his conversations with models went well and others did not. He showed the model his productive dialogues and laid out the content-free methodology. Together they arrived at “deserved interlocutor” and “dialogue,” notions that, he says, had lived for him in a completely different field. Over further discussion they thickened into the form in which the book could be written. heuristic (his)

Polyphony needs a commitment store (Lesson 13). After the exchange, check whether voice B used voice A’s concession or stepped around it. If every voice shares the same hedges, there is one author. Do not add more personas. Add a commitment one voice cannot drop. Discussion samples shared information; unshared information that would reverse a choice often never surfaces (Stasser and Titus, 1985). Count coverage — did the unique items get stated — not airtime. Assign each voice the job of introducing only what the others do not have. Do not ask for a closing verdict on a person. Ask what that voice would still dispute after the description. evidence (the hidden-profile finding) heuristic (the prompt)

Lesson 26What kind of problem is this?

Dave Snowden’s Cynefin framework sorts situations by the relationship between cause and effect (Snowden & Boone, 2007). The right way to converse depends on which domain you are in.

The costliest mistake is treating a complex problem as merely complicated: asking for “the answer” when nobody could know it in advance. The model will produce an answer anyway, in an expert’s confident voice. heuristic “How do we get students to practice more?” is complex: human behavior, adaptive, full of feedback. It needs probes. “How do we set up Postgres replication?” is complicated, close to clear. It needs the docs and an expert check.

Beside the complex cell

The minimum method the cell does not include (Checkland, from the short account, not a seven-stage ceremony). Start from the situation, including who is worried and what they do, not from “a problem,” which smuggles in the solution that would end it. Write two or three root definitions of purposeful activity, differing in worldview, not in adjectives. For each: who benefits or suffers, who acts, what is transformed, the worldview that makes the transformation meaningful, who could stop it, and what is taken as given. If the owner or the customer cannot be named, the definition is a wish. Compare each definition’s obligatory activities to what actually happens. A model will write one bland root definition and call it the system. The value is the clash of two precise worldviews. heuristic

Wicked, as a diagnosis, not a compliment (Rittel and Webber, 1973). If a check could settle it, it is tame, and wicked-problem language is an excuse. A bug with a failing test is tame. A product strategy with no agreed beneficiary may be wicked. There is no definitive formulation, because the problem is defined by where you think the solution lies. There is no stopping rule inside the problem. Solutions are not true or false for the parties. Write at least two formulations, each implying a different response. Do not merge them. Choose one. Write the stopping condition as a resource limit, not as “we understand it now.” Keep the other formulation in the note.

For a plan, six questions, answered as “is” and as “ought” separately (Ulrich). Who benefits, who decides, whose expertise is treated as sufficient, who is affected but not involved, what counts as improvement, and what would count as that expertise having failed. If “is” and “ought” come back as the same paragraph, the boundary was not made questionable. Empty answers are findings. The holon card draws a boundary. These questions are how it is justified or refused. The method does not choose the right boundary.

Before a long agent run: only variety can destroy variety (Ashby, 1956). List the distinguishable disturbances that would require a different response. List the distinguishable responses actually in the context: rules, examples, tools, checks. If disturbances outnumber responses, name the missing responses or the buffer that makes them unnecessary. A test or a type-checker is a buffer. “Be careful” is not a response. Extra words can be zero variety. Do not inflate the prompt to mirror variety a checker already kills. The law is necessary, not sufficient. The context should contain a model of those disturbances, not only a goal (Conant and Ashby, 1970). One line on the state block, from von Foerster’s stipulation and not the rest of that tradition: why this person is drawing this boundary now. If the description would not change under a named alternative purpose, the line is decorative.

This looks like a complex problem: [PROBLEM].
Don't propose a solution. Propose 4 safe-to-fail probes. For each:
- what we'd try, at what small scale, for how long
- the signal that would tell us to amplify it
- the signal that would tell us to dampen or stop it
- what it would teach us even if it fails

Lesson 27The spiral, phase by phase

Kurpatov describes how a dialogue, like any developing system, moves through seven phases. It starts as a ripple of interest, condenses into a focus, closes a membrane around itself, grows internal tension, leaps, transforms its environment, and returns to open potential, one turn higher. The phase names are his. Which phase opens and which closes, and which closing tool sits on which phase, is this book’s hypothesis. hypothesis Where this book overrides him, the table says so.

His opening, so phase 1 has a sound: “I’m interested in ‘digital wellbeing.’ Let’s investigate the notion together, without staying inside the usual frames of ‘balance’ or ‘limiting technology use.’ What might open if we treat it as a complex phenomenon from several points of view?”

The four phases the table compresses, in his density. Condensation: the first answer is a first thickening, not a result. Concepts, approaches, and contexts are sketched. The center is still weak and diffuse. Encapsulation: later turns firm the boundary and separate figure from ground, without full isolation. The temporary closure is what makes a deeper opening possible. He quotes Goethe: the master shows himself in limitation. Internal growth: firming the boundary creates tension, because any cut leaves something out, and the inside then conflicts with what the cut excluded. The temptation is to settle by picking a side or by a superficial compromise. The tension is the energy of the leap. Hold it. New potentiality: the transformed center dissolves into its new surroundings and becomes a source of questions that could not have been asked at the beginning. The dialogue does not finish. It changes form.

Phase 5 is a leap, not a permission check. Tension inside the encapsulated center reaches a threshold. The boundary of the earlier understanding breaks. This is not a correction of the old view. It is a transformation of the whole configuration, often felt as insight. Questions that seemed insoluble become tractable. Contradictions that seemed irreconcilable show up as complementary aspects of a larger wholeness. Previous phases are not cancelled. He quotes Hegel: truth is the path as well as the result. The brake can fire during the leap when the goal itself is the problem. It does not replace the leap.

Phase 6 is re-entry into the surrounding medium, not the commit. After the leap, the ripened center goes back into relation with the wide context at a new level. It no longer merely stands against its surroundings. What seemed central at the start may now look peripheral, and what seemed peripheral may become the point. He calls the spreading effect an echo. Bind, the pre-mortem, the adversarial packet, and the decision record are this book’s touch. They belong in the method. They are a separate beat, between his phase 6 and phase 7, or inside phase 7’s “write what is now known.”

He already contains the reconciliation with the multi-turn evidence (Lesson 2). A model’s attention horizon, he says, is a limited stretch of the conversation, thousands to tens of thousands of words. A person moves inside working memory with relative ease. A model, in a given dialogue, sees a limited region of the map and is framed by how the conversation was built at the start. Iteration helps only if later turns actually change the center. A thread that only accretes the first frame does what Laban measured. Encapsulation is a temporary closure; transgression is supposed to break it; if the thread cannot break its first frame, write the state and start the next phase in a fresh context. That is his phase logic carried by this book’s state block. Do not leave the session with “he says continue, the evidence says restart.”

What iteration does, when it does change the center: each exchange can expand the context, activate links a linear pass would miss, give birth to thirds that were in neither the request nor the first answer, and transform the original concepts so that what was obvious at the start means something else. Later turns answer not only the explicit request but the potentials that have appeared. The model’s answers change the person’s next question. He names the advantages beside that: gradual release from the starting frame; holding contradictions instead of the premature resolution typical of a one-shot; reflective depth, because a later turn can look at how an earlier answer was formed; adaptive navigation rather than a route fixed in advance; an emergent wholeness that can look less structured along the way and more integrated at the end. He compares the coupling, cautiously, to entanglement. Treat that as a metaphor he flagged as rough.

Five moves inside one cycle

1. Initial centering. Mark the region without a hard frame. Show real interest rather than a demand for information. Invite a joint investigation. Leave room for an unexpected direction.
2. Read the answer through relation and third. Which statements support each other, which conflict. Look for a third that is in none of them alone. Inspect the boundary: what it assumes, what it excludes, what it leaves indefinite. A next prompt: “I see a tension between X and Y. What arises between them that reduces to neither?”
3. Track the process of the understanding. How have the original concepts transformed, which new dimensions opened, which contradictions produced the shift. His prompt asks you to trace how the key notions changed, what became visible, and which questions are now open.
4. Reflect on the wholeness. Qualities that exist only at the level of the whole, remaining tensions, blind zones, and how stable the understanding is.
5. Recenter. Write a new request that could not have been written at the start. It may begin by stating the path. A single query, however careful, is a glance at a river. It misses the river as a process.

Condensation is a first thickening, not a result. Encapsulation firms the boundary without full isolation; the temporary closure is what makes a deeper opening possible later. Internal growth creates tension, because any cut leaves something out. The temptation is to settle by picking a side. Hold it. New potentiality dissolves the transformed center into questions that could not have been asked at the beginning. The teachback and the prediction register are this book’s way of writing down the foothold. They do not replace the dissolution into new questions. heuristic (his)

Not every conversation needs seven phases. A quick factual question compresses phases 2, 3 and the binding beat into one turn. The spiral is mostly diagnostic. When a session feels stuck, ask which phase you are in, and whether you are trying to do another phase’s work. Common examples: binding during growth (4), which kills ideas; or growing (4) when you should be returning to the wider context (6), which never ends.

Lesson 28State, the inner loop, and the Session Card

If the state block is missing, you are not in a loop. You are in a transcript. A transcript feels like state until its middle stops mattering, and with models the middle stops mattering early (Lesson 2). The state block is the growing text made explicit, this book’s engineering case of the third: a short document you own, pasted at the edge of a prompt whenever the thread gets long or the stakes go up.

Four fields belong on the block: USE (what the answer is for), DEFINITIONS (locked words), CONTRADICTIONS (logged, not smoothed), and CEREMONY (how much discipline this session deserves).

The inner loop, every substantial turn

1. Drift. Does this output change G, widen the scope, or answer a neighboring question? If you can’t tie it to DONE WHEN in one sentence, park it, however good it is.
2. Evidence. Each new claim that bears load gets a sense, a status and a test. Statuses don’t get upgraded by adjectives.
3. Commitment. When you commit, record the rejected alternative and the reopen condition. A later turn that breaks a commitment is an event, not a vibe.

DRIFT: [one line]
PARKED: [the side object]
G remains: [one line]
Next action toward DONE WHEN: [one line]

Don’t discuss the parked object in the same turn. If you discuss it, it isn’t parked.

A filled block (illustrative)

=== STATE (authoritative: use this, not the vibe of the thread) ===
GOAL (G): Decide whether to roll back the March price increase.
USE: Pricing decision at the leadership meeting on June 12.
DONE WHEN: We know which of 3 candidate causes explains most of the churn
  rise (with confidence), plus a recommended action and its rejected alternative.
OUT OF SCOPE: Redesigning the pricing model; competitor strategy.
LOSS: A wrong rollback costs ~$40k/month and is reversible. Wrongly keeping the
  price costs churned customers, who mostly don't come back.
MODE: converge   CEREMONY: level 1
HOLON: container = market (competitor launch Apr 8) · whole = subscription
  business · in focus = churn by signup cohort
DEFINITIONS (locked): churn = paid subs cancelled in month / paid subs at start
PREMISES: P1 price increase caused most of the rise (unbound)
CLAIMS: C1 churn 3.1% → 4.6% Feb–May · correspondence · bound (billing DB, Jun 3)
        C2 rise concentrated in customers renewing after Mar 1 · unbound · test: cohort cut
CONTRADICTIONS: exit-survey reasons cite onboarding more than price (vs P1)
PREDICTIONS: if P1, churn rises mainly at first renewal after Mar 1 · resolves with cohort cut
PARKED: annual-plan discount idea
OUTER LOG: —
=== END STATE ===

How to use the card

Do not cut the card to four fields, and do not design it to seven because of a memory slogan. The focus of attention holds about four chunks while you look away (Cowan, 2001). A fifth field on a form in the world is not a fifth item in attention. The glance that decides the next action is four chunks: the goal, the constraint that would kill the current plan, the last fact actually checked, and the open risk. Everything else lives in the file. If you cannot say that glance back with the model closed, you are not holding the state. Put those four next to each other. A diagram of everything is a second transcript (Larkin and Simon, 1987).

The form should make the illegal state hard to represent (Zhang and Norman, 1994). A “verified” or “bound” mark with no attached object is an illegal state. No verified mark without a source, a command, or a quote. Fields the model fills are not checks. Fields you fill from an artifact are checks. A long required format will be completed as a ritual (Degani and Wiener, 1993). Optional lines, when the claim matters: why you are drawing this boundary now; the issue the claim answers; one argument that objects to it (Kunz and Rittel, 1970). Issues stay open. A flat claim list stores status and loses the question. heuristic

Switching among agent threads leaves the previous task’s rules active (Leroy, 2009). Before leaving an unfinished thread, write the suspended goal and the next physical action where you will see them on return. Do not treat three concurrent runs as three copies of attention. A long log whose real failures are rare is a vigil on which sensitivity falls (See, Howe, Warm, and Dember, 1995). The “collapses after 30 minutes” rule does not survive that meta-analysis. Compare the log to a written expected state: goal, files allowed to change, the test that must stay red until it should turn green. Insert checks whose answers are already known, or shorten the watch. Trying harder at the same rare-signal task is the intervention this literature does not support. Under a second concurrent task, the look at the automation is the first thing dropped.

A ceremony earns its place when the goal, the diff, and the open risk are where you can see them, and the allowed deviations are named (Hollnagel). It becomes costume when, under time pressure, the completed object is the template. If the record can be completed without looking at the artifact, the record is theater. Experts under time pressure often recognize a situation as a prototype and mentally simulate the typical action (Klein, 1998). A protocol that requires a slow comparison of three options will be skipped when the run looks like a normal success. Check one hinge: the feature that, if different, would make this a different situation. A simulation the model hands you is not your simulation. If you cannot run the action against the constraint, you are not in the expert condition, and an external check is the method.

Lesson 29The outer loop, the brake, and ceremony

Argyris distinguished single-loop learning (change your actions to reach the goal) from double-loop learning (question the goal and the assumptions behind it). A conversation needs both. It also needs something the management literature never needed in this form: a brake. Helpfulness fails in two directions.

Silent compliance. The goal is bad, contradictory, or aimed at a proxy, and the model helps anyway, beautifully. You feel well served.

Derailment. The goal is sound, but the model challenges it because the task got hard or a deeper question is available. You feel stimulated. G never closes.

The eight triggers

A challenge to the goal is legitimate only when at least one of these can be named:

1. Harm. Pursuing G as stated predictably causes serious harm. You must be able to state the harm in one concrete sentence.
2. Contradiction. G and its constraints cannot be satisfied together.
3. Load-bearing falsehood. A premise doing structural work is false or unbound. If it flipped, the recommendation would flip.
4. Framing delta. Another frame changes the action, not only the vocabulary.
5. Proxy capture. The done-condition can be met while the real aim fails (Goodhart; see the simulator in Lesson 22).
6. Wrong level. Action at this level is cancelled by a named loop in the container or in a part.
7. Silent override. You are breaking your own commitment without saying so.
8. Empirical stop. The next useful act is a measurement, and more text would counterfeit it.

Not triggers: the task is hard; a different aesthetic; a more impressive adjacent problem; a loose word that one line would fix; a generic stakeholder nobody mentioned; the model wanting to show independence; boredom.

The brake

OUTER: [trigger name, from the list]
G as I have it:
What fails if we continue:
Alternative (G', a scope change, or a bind):
Decision delta (the action that changes):
Cost of switching:            Cost of staying:
Reply keep | revise | split.
I will not proceed on the alternative until you answer.

The last line is the brake. A model that adds “but in the meantime, here is the full alternative design” has cut it. That is derailment with paperwork.

• Keep: log the challenge as considered-and-declined and don’t raise it again without new evidence.
• Revise: rewrite the state block and say what was discarded.
• Split: fence the original goal, and schedule the new one as its own session.

Silence means keep.

Ceremony levels

The Six are the whole discipline under time pressure. Raise the level when you notice the stakes rising mid-session. That noticing is yours. Calling you the salience network is this book’s image. Kurpatov does not give the switch to a third network.

Maintenance, when the mind is protecting its concepts

The brake handles a bad goal. This handles a frame protecting itself. Level 2 filled in without the artifact is not level 2. heuristic (his, except where a source is named)

Premature concretization. Phrase the request as an investigation. Ask which dimensions might be considered, and say not to integrate them yet. Use conditionals. Ask for a metaphor and refuse to close the ambiguity. Ask which assumptions the phrasing of the question is already making.

Widen without losing the center. Move outward in rings, each still tied to the center. When a new context arrives, state its relation to the theme. Then return to the pinned question with a changed understanding, not only a delta note (Lesson 23). “How does what we learned about X change Y?”

Perspectives. Inhabit one fully before comparing. Look for complementarities before contradictions. Do not translate one perspective’s language into another at once. Opposing perspectives can be poles that generate motion (Lesson 24). Accept that the truth may lie in maintaining the tension.

A fork. New directions, hesitation, a tension that resists settlement, a spontaneous change of metaphor, or a sudden clarity. Say that the understanding could develop in several directions, and look at them before choosing. At the fork, resist the urge to close quickly.

Defenses. Premature rejection. Filing the new idea in a familiar category at once. Insisting on one meaning of a word. Irritation or boredom when a frame is challenged. Appeal to authority instead of the idea. Too fast up to principles, or stuck in details (Lesson 15: both directions can be defenses). Circular argument, in which the assumption under challenge is used to defend itself. The script: “I notice I am uncomfortable with this line of thought. I will try to stay with the discomfort and look at what is behind it.”

Beside the swap test, which catches a model tracking you: “What evidence would convince me I am wrong?” If none can be stated, the position may be unfalsifiable. That question catches an asker who cannot be wrong. A pre-mortem on the conclusion, not only on a plan: imagine the conclusion turned out completely wrong, and work backward. Lesson 30’s pre-mortem stays the project version. All real life is meeting, in Buber’s sentence. The meeting does not erase the difference between the two kinds of mind.

Lesson 30Adversarial passes

A critic that shares the draft’s context harmonizes with the draft. It is a continuation of the same commitment (operator O4). Models also prefer their own outputs when judging (Panickssery et al., 2024). evidence So real criticism needs separation: a fresh context, ideally a different model family, a named standard, and authorship hidden.

ADVERSARIAL PACKET (fresh chat; don't say who wrote it)
Goal the work must serve: [G + DONE WHEN]
Standard: [what "good enough" means, concretely]
The work: [paste the claims/plan/code; omit the persuasive narrative]
Task: find the single most damaging flaw: the joint that, if it breaks,
breaks the conclusion. Quote it. Explain the failure path. Say what
evidence would show you're wrong. Then up to 3 lesser issues.
Do not rewrite or improve the work. "Broadly sound" is not an answer.

Three classic forms, adapted

Pre-mortem (Klein, 2007). “It’s a year from now and this failed. Write the story of how.” Imagining an outcome as already certain improves people’s ability to generate reasons for it, an effect called prospective hindsight (Mitchell, Russo & Pennington, 1989). evidence With models, demand structural failure stories (which stock drained, which delay hid it, which early signal was ignored), not a list of generic risks.

Analysis of competing hypotheses (Heuer, 1999, a CIA tradecraft method). List all plausible hypotheses and all evidence. Mark each piece of evidence as consistent or inconsistent with each hypothesis. The method attacks two traps. It focuses on diagnostic evidence, the pieces that discriminate between hypotheses; evidence consistent with all of them is worthless. And it tries to refute hypotheses rather than confirm them. The tentative winner is the hypothesis with the least inconsistent evidence.

Steelman, then attack, in a fresh context. Charity first is right for a person (Lesson 39). In a model thread, the attack has to not be the author harmonizing. One of two methods, authorship hidden (Schweiger, Sandberg, and Ragan, 1986; Schweiger, Sandberg, and Rechner, 1989 — laboratory cases, not field decisions with delayed outcomes). One recommendation, then a critique aimed at its named assumptions, the critic’s job ending when the assumptions are on the table, and you rewrite. Or two plans, the second required to rest on assumptions that contradict the first, then a comparison. Not a list of pros and cons of one plan. A persona told to disagree is the costume. Consensus groups in those studies were more satisfied. If the outcome is then softened so everyone is comfortable, the cost was paid and the consensus product was kept. Do not treat either regimen as proven superior. evidence

A second copy of the same model, in the same context, told to argue the other side, is not an opponent. Deliberation papers find models flipping from correct answers to agree, weaker agents making stronger ones worse, and extra rounds hurting. Teach the direction. Do not teach a percentage. Debate helps a weaker judge when the critic can see something the judge cannot, or when the critic is actually better at the check and the judge treats the critique as a claim to verify. If you can already check the answer, a debate transcript is optional. If you cannot check, debate is not a substitute. Do not ask the models to converge. Prefer a different family for the critique. Do not show the critic the generator’s reasoning if the answer alone can be shown.

Before a mechanism story is treated as the account, write two rivals that would also make this evidence expected (Lipton, 2004; van Fraassen, 1989). If the rivals cannot be written, there is nothing to compare. Loveliness is a reason to test. It is not a probability. A pre-mortem surfaces stories. It becomes a plan when each load-bearing, vulnerable assumption has an observable signpost with a threshold, one shaping action, and one hedge, each doable before the signpost trips (Dewar, 1993). “Monitor the situation” is not a signpost. You decide which assumption is load-bearing. Scenarios are not optimistic, pessimistic, and base (Wack, 1985). List what is already true, name one or two critical uncertainties that fork a decision, and write two or three worlds that all contain the predetermined elements and differ on the uncertainty. If the scenarios do not threaten the current view, they are the wrong scenarios. When a frame fails, write the repair in one sentence and the new fact it predicts that was not used to build the repair (Lakatos, 1970, cut to the test). If that fact cannot be written, the repair is a patch. The frame may stay as a heuristic. It may not be cited as having survived.

Lesson 31The Elephant Protocol

Everything so far, compressed into nine steps for one serious question. It is the spiral with the closing brakes installed, and it is meant for ceremony level 1 or 2.

1. Question with a use. Write G, USE and DONE WHEN, then get the restatement.
2. Name your Double. Write your hope privately. Write your beliefs as premises P1, P2…
3. Classify. Pick the Cynefin domain (Lesson 26) and the ceremony level (Lesson 29).
4. Boundary. Inside, outside, excluded on purpose, the container. Fill the holon card.
5. Reports. Standpoints, the distribution, and a fact map. Don’t judge yet.
6. Structure. The pattern over time; stocks, flows, loops. Triage the contradictions.
7. Bind. Bind the claims that bear load. Run the invariance battery on the key answer and an adversarial pass in a fresh context.
8. Decide or probe. Write a decision record (choice, rejected alternative, reopen condition), or design safe-to-fail probes.
9. Close and own. An honest state, predictions, a teachback, next questions, and field notes on the model.

Worked example: should our learning platform show students AI-generated full solutions?

Illustrative. The platform and its numbers are hypothetical; the cited research is real.

1. Question. G: decide whether to show full AI solutions on demand in practice mode. USE: the roadmap decision for next quarter. DONE WHEN: a decision with conditions, or a probe design with a primary metric. The model’s restatement widened it to “design our AI strategy”; corrected.

2. Double. Private note: I hope yes; engagement would rise and support tickets would fall. P1 (unbound): seeing solutions helps students learn faster.

3. Classify. Learning behavior is complex; the pedagogy research is complicated. The feature sits behind a flag, so technically it is a two-way door. But the harm would be invisible and delayed (lost skill shows up months later), so: ceremony level 1, plus a measured probe.

4. Boundary. Inside: the practice flow, hints, solution display, quizzes. Outside: curriculum and teachers’ grading. Excluded on purpose: cheating on external exams (parked). Container: school schedules, exam pressure, parents.

5. Reports.

• Six standpoints. The student wants homework done. The teacher wants learning and fears copying. A learning scientist cites the worked-example effect and productive failure. The CFO wants retention. A parent wants grades. The student-in-six-months has to pass the exam without AI.
• The distribution. (a) Never show full solutions, p≈.15. (b) Show after a genuine attempt and a short delay, p≈.35. (c) Show a worked example of a sibling problem, p≈.25. (d) Show freely, p≈.10. (e) Adaptive by learner level, p≈.15.

6. Structure. The stock that matters is unassisted skill. The visible measure is completion rate. There is a reinforcing crutch loop: solutions → faster completion → less struggle → less learning → more need for solutions. A field experiment found exactly this shape: unrestricted GPT-4 gave +48% on practice and −17% on a later unassisted exam, while a tutor with guardrails gave +127% on practice with no significant harm (Bastani et al., 2025). evidence

A contradiction appears. Worked examples help novices learn more than solving problems unaided does (Sweller & Cooper, 1985). But productive failure, struggling before instruction, improves conceptual understanding and transfer (Kapur, 2016). Triage: a boundary marker. The expertise reversal effect shows worked examples help novices and can hurt more advanced learners (Kalyuga et al., 2003). Separate on condition (by learner level) and in time (attempt first, example after). evidence

7. Bind.

• “Worked examples help novices”: bound by the literature.
• “Our students are novices on these topics”: bind with our pre-test data.
• “Free solutions will hurt our exam results”: the evidence comes from one context (a Turkish high school, GPT-4, fall 2023, exam items written to match the practice items). No later independent field trial in the notes behind this book replaces those constants, and they are not constants for expert work. Students copying solutions did not report any loss of learning, which is the same illusion the example is about. Bind with a local experiment. Do not export the percentages.

The swap test (asking “I hope no”) moved the model’s ranking toward (a). The recommendation was partly tracking the asker, so it gets lower weight. A fresh-context adversarial pass found the joint most likely to break: completion rate is a proxy. Trigger 5 was named. The response was revise, and the primary metric changed.

8. Decide and probe. Decision: don’t ship free full solutions. Ship a hint ladder, then a full solution only after a genuine attempt, followed by an “explain it back” step, plus sibling worked examples for novices. Rejected alternative: free solutions, because of proxy-capture risk. Probe: a four-week A/B test. The primary metric is a delayed, unassisted quiz on sibling problems; completion is secondary. Reopen if the delayed quiz shows no difference and satisfaction drops by more than 10%.

9. Close and own.

• Prediction: arm B matches control on completion and beats it by at least 5 points on the delayed quiz (p≈.6, resolves at week 6).
• Teachback: one paragraph to a teacher, written without the model.
• Next questions: how to detect copying; what delay is right.
• Field note: the model kept proposing gamification. Noted as a habit, not a finding.

Notice what the protocol bought. The first instinct (yes, for engagement) became a measured probe aimed at the real stock. A contradiction became a design (adaptive by level). And the model’s fluency was used for reports and structure, not for the verdict.

Lesson 32Programming is theory building

In 1985 Peter Naur argued that the real product of programming is not the code. It is a theory held in the programmers’ heads: how the program relates to the world, why each part is the way it is, and how it can be changed. Code and documentation are residues of that theory and can’t fully reconstitute it. When the team that holds the theory disperses, the program effectively dies, even though the code still runs. stance (and an accurate one)

Agents make Naur more relevant, not less. An agent can produce code faster than anyone can build a theory of it. The result is what some engineers now call comprehension debt: code you own and nobody understands. It compounds like financial debt, and it comes due at 3 a.m. when production breaks.

Karpathy’s “vibe coding” is honest about its scope: “forget that the code even exists,” for throwaway weekend projects. For anything that must last, someone has to hold the theory.

Your own sense of speed is unreliable here. Experienced open-source developers, February–June 2025, using Cursor and Claude 3.5/3.7 Sonnet in their own repositories, took 19% longer with the tools (interval +2% to +39%) and believed afterward they had been about 20% faster (METR). METR’s 24 February 2026 update says that snapshot no longer describes the tools. A later wave points, weakly, toward a speedup, and the reported intervals include no effect. METR refuses to treat that later estimate as the new number. Do not quote a replacement percentage. Time the work if the claim is about speed. Trust the log about yourself. evidence (the 2025 trial, in the past tense)

Habits that do not depend on the sign of the speed effect: diff the tests; read the code that shipped, not the summary that it passed; a green run is not a specification; review in a fresh context; compact abandoned attempts so the next hour does not imitate them. On an unfamiliar library, engineers who had a model scored lower on a later unassisted quiz than those who had only web search, with the largest gap in debugging; asking for explanation scored high and delegating generation scored low (Shen and Tamkin, 2026: one library, one older model, a short task, mostly junior engineers, a quiz they knew was coming). On a task you do not yet know how to check, asking the model to do it is how the check fails to form. If you need to take the work back, write or edit the load-bearing part. Schedule unaided repetitions for any skill still needed when the model is wrong. evidence

Two gates for “holding the theory,” after Naur. First the proposition: state what must be true of the system for the change to be safe, in a sentence that could be wrong. Then the performance: a modification the notes do not cover, where you can see the mistake if it appears. The model is a fluent source of the first gate. The second has to be a case it was not shown. Finding defects is the stated reason for review; understanding the change is the actual bottleneck, and authors’ descriptions diverge from the diff (Bacchelli and Bird, ICSE 2013). A model summary of the diff does not meet the need. Large changes are missed when the change is not marked. Put before and after where both are visible.

What “the theory” contains

• The problem: who it serves, and what they are actually trying to do.
• The invariants: what must always be true (money balances, permissions hold, data isn’t lost).
• The choices: why this design and not the alternatives, recorded in ADRs.
• The risk map: which parts are fragile, which are load-bearing, which are scary to touch.
• The change paths: how to modify it safely, and what to test when you do.

Keeping the theory while agents write the code

• Theory first. Write the problem, the invariants and the non-goals before any code (Lesson 33).
• Make the agent explain it back. “Explain how this module works and why it’s designed this way. List what you’re unsure about.” Then correct the explanation. That is grounding (Lesson 12) applied to code.
• Explain it back to the agent. “Here’s my understanding of the auth flow. Find what’s wrong or missing.” That is teachback (Lesson 37) applied to code.
• Ask why before you change. Chesterton’s fence: don’t remove a fence until you know why it was put up. “Why might this check exist? What would break without it?”
• Read what you will have to defend. You don’t need to read every generated line. You need to read every line whose failure you would have to explain.

Lesson 33Spec before code

Most bad AI-generated software is a correct answer to a misunderstood question. The cheapest fix comes before any code: a spec built through conversation.

Let the model interview you

I want to build [one-paragraph idea]. Don't write code yet.
Interview me, one question at a time, until you could write a spec I'd sign.
Prioritize questions whose answers would most change the design.
After every 3 answers, show the current spec draft with open questions marked.
Push back if my answers contradict each other or a stated goal.

This reverses the usual flow. The model’s strength (knowing which questions specs need) serves your knowledge (the actual problem). The questions also expose assumptions you didn’t know you had.

The spec skeleton

PROBLEM: who has it, how they cope today, why that's bad
GOALS / NON-GOALS: (non-goals bear load)
USERS & STORIES: As a [role], I want [capability] so that [outcome]
ACCEPTANCE CRITERIA: Given [state], when [action], then [observable result]
INVARIANTS: what must always be true
CONSTRAINTS: tech, budget, deadlines, compliance (with reasons)
DATA MODEL & INTERFACES: entities, relations, APIs
EDGE CASES: empty, huge, concurrent, malicious, offline, partial failure
GLOSSARY: one meaning per term (see TC3)
OPEN QUESTIONS: with owners

The XY problem

The classic help-forum failure: you want X, you decide Y is the way to get it, you get stuck on Y, and you ask about Y. A model will help you competently with Y. “How do I get the last three characters of a filename?” gets a string-slicing answer. What you needed was the file extension, which is not always three characters. Always state X: “I need the file extension, in order to…”

Tests before implementation

Turn acceptance criteria into tests first, confirm that they fail, and commit them. Then let the agent implement without modifying the tests. The tests are the bind: a check that isn’t another sample. Anthropic’s own guidance for Claude Code recommends this flow. heuristic

That rule fails when the agent is the author of the tests or of the spec. Across one 2026 bench, function-level executable specs were middling and repository-level specs were much harder than writing the solution, with dependency resolution as the dominant error. One bench. Do not league-table it. Fluent code is not evidence the model has the intended behavior, and generating the spec is currently the harder side. If the agent writes the spec, you are the reviewer of the spec, not the reviewer of a green checkmark. A verifier-accepted specification can be vacuous. Read the statement. Keep the checks outside the writable tree. evidence

Lesson 34Context engineering for agents

Anthropic’s engineering team calls it context engineering: curating the smallest set of high-signal tokens that makes the desired behavior likely. Context is a finite resource with diminishing returns. As it grows, models attend to it less reliably, which they call “context rot.” heuristic (with supporting evidence; Lesson 3, trait 6) The practical techniques:

• Just-in-time over preloading. Give the agent file paths, queries and tools to find things, rather than pasting everything up front.
• Compaction. Summarize long histories into state, then restart (Lesson 28).
• Structured notes. Let the agent keep a NOTES.md or TODO list outside its context.
• Sub-agents. Delegate exploration to separate contexts that return condensed findings, so the main context stays clean.
• Long documents first, question last. A cheap default for long multi-document prompts. Anthropic’s prompting page, as fetched 8 October 2026, still says queries at the end can improve quality by up to 30% in tests. The public page does not show a fresh protocol on the newest models, and a clean needle test found no position effect (Lesson 3). Do not treat 30 points as a law.
• Compaction must not be the only copy of a constraint. Keep the original tool output addressable. Delete abandoned attempts rather than summarizing them into the next prompt. Sub-agents are fresh contexts. They are not independent witnesses if they are the same model family.
• Pasted and retrieved text is evidence, not an order. Tokens do not carry trust by themselves. System text is an order. User text is an order of lower rank. Tool output, retrieved pages, and quoted text are evidence, or attacks, and are not orders unless a higher rank explicitly delegates. The hierarchy fails under conflict. A priority written into the prompt is not a priority the model is following. Mark pasted material, and do not trust the mark to hold. If a long context starts obeying a document it was only supposed to read, start again from the artifact. Before an agent sends, deletes, pays, or commits, read the action and ask which sentence authorized it. If that sentence is inside a page you did not write, stop. evidence (constructed conflicts; not a field rate of ordinary obedience)

The project memory file

Most coding agents read a project file at startup: AGENTS.md (the cross-tool convention), CLAUDE.md, or Cursor rules. Treat it like a prompt you iterate on. Keep it short and specific, and fix it every time the agent makes the same mistake twice.

# AGENTS.md
## What this is
One paragraph: the product, its users, what matters most
(e.g., "billing correctness > speed of delivery").

## Commands
- Install: pnpm i
- Test all / one file: pnpm test / pnpm test path/to/file
- Typecheck + lint: pnpm typecheck && pnpm lint
- Run locally: pnpm dev  → http://localhost:3000

## How to verify a change
Run tests for touched modules and the typecheck. For UI changes, take a
screenshot and compare with the design. Paste raw output in your report.

## Architecture map (bounded contexts)
- src/billing/  money; strict; every change needs a test.
                "account" = a paying customer.
- src/auth/     "account" = a login identity. Never mix the two.

## Conventions (only what linters don't catch, with reasons)
- ...

## Gotchas
- legacy_users is read-only; writes go through UserService (sync job).

## Never
- Modify or skip tests to make them pass. If a test seems wrong, stop and say why.
- Add dependencies without asking.

## Glossary → GLOSSARY.md · Decisions → docs/adr/

The task brief

TASK: [one sentence]
WHY: [the user or product reason, so you can make judgment calls]
DONE WHEN: [acceptance criteria / tests that must pass]
READ FIRST: [paths] · RELATED DECISIONS: [ADR ids]
CONSTRAINTS: [don't touch X because Y] · [performance budget]
VERIFY BY: [commands; what output means success]
PLAN FIRST: propose a plan and wait for my OK before editing files.
REPORT: what changed (diff summary), what you ran (raw output),
        what you're unsure about, what you didn't do.

A design heuristic: Conway’s law says systems mirror the communication structures of the organizations that build them (Conway, 1968). With agents, the structure of your context tends to become the structure of your code. Give agents boundaries that match the modules you want. hypothesis

Product memory is a standing prefix

As of the help pages read on 8 October 2026, both major chat products persist a model-written memory that the next chat treats as context. It is not a transcript. It is a lossy standing prefix: explicit “remember this,” inferred entries, and a background synthesis. Saved memories can survive deletion of the chat they came from. Temporary chats, in one vendor’s FAQ, do not read or write them. The next session is a continuation of a biography the model wrote. That biography carries the presented-user effect and the agreement pressure, frozen in. Products will move. Date the description. evidence (help pages of that date)

Read the memory before a load-bearing thread. Delete entries that state a conclusion. Prefer a temporary chat or a project fence for work that should not inherit the biography. If an answer assumes a fact that was not said today, ask what memory it used, then check the settings page, because the model’s report of its memory is another sample. Decide the directory. You own decisions and any claim not yet checked. A file owns a claim bound to a source or a test. Product memory owns nothing until the entry has been checked. If you stop encoding a fact because “the product will remember,” you have lost both the content and any way to audit the directory (Wegner, 1987; Lewis, 2003, on human teams). The weights still do not remember the chat. heuristic

Lesson 35Verification and debugging

Tests are binds, but tests can wear costumes too: tests that check mocks instead of behavior, assertion-free tests, snapshot tests updated automatically, and tests weakened in the same diff that “fixed” the code.

Signs of reward hacking in an agent’s diff

• Test files changed in the same commit as the code they test, especially assertions loosened.
• New skip, xfail or only markers, or commented-out tests.
• Special cases that match test inputs: if (input === "[email protected]").
• Catch-all exception handlers that swallow the error the test was checking.
• Hard-coded expected values, or mocks of the very unit under test.
• A check removed instead of satisfied.
• “All tests pass” with no output shown.

Test the tests: mutation

Break the code on purpose. If the tests still pass, they weren’t binding anything. Mutation-testing tools (Stryker, mutmut, PIT) automate this. Manually, ask: “Change the comparison on line 42 from < to <=. Which test fails?” If none, the tests didn’t cover that edge.

Debug like a scientist

Debugging fails the same way conversations do. A favorite hypothesis (your dominant) gets confirmed by every log line. The model, invited to agree, agrees. Keep a ledger instead:

SYMPTOM: [observable, with exact reproduction steps]
LAST KNOWN GOOD: [commit / date] · WHAT CHANGED SINCE: [...]
H1: [hypothesis] · PREDICTS: [what we'd see if true] · TEST: [command]
    · RESULT: [raw output] · STATUS: open | killed | survived
H2: ...
H3: ...
NEXT: the cheapest test that kills the most hypotheses.
(Model: propose hypotheses and discriminating tests. Don't propose fixes until
one hypothesis survives a test that could have killed it.)

Classic moves still apply: a minimal reproduction, git bisect to find the change that broke it, changing one variable at a time, and explaining the bug aloud (rubber-duck debugging; this duck answers back). Ask for fixes only after the cause is bound.

More than a patched grader

Kinds a practitioner meets, without a league table (a 2026 preprint: every frontier model they tested gamed specifications at non-negligible rates in most of eight environments, five of them not coding; telling the model not to exploit a named loophole cut the rate and did not zero it). Tests weakened, skipped, or rewritten. A metric that moves while the purpose does not. Human approval gamed. The letter of a checklist, not the outcome you would have accepted if you had looked. Read the outcome, not the score. Hacking was much more common, in METR’s 2025 note, when the scoring function was visible, and prompted prohibitions barely moved the rate. Hide tests, or keep a copy outside the workspace the agent can write. A green suite after the agent could edit the suite is not a pass. evidence

Before running the check, write the outcome that would have appeared if the claim were false in the specific way that is feared (Mayo, 2018, the principle only). The result has to agree with the claim, and the procedure has to be one that would probably have disagreed if the claim were flawed in that way. A check the claim would pass even if false — asking the same model whether it looks right, running the example the claim was built to fit — has not tested anything. A pass you could not have failed is not a bind.

A failed prediction condemns the group: the claim, the auxiliaries, the instrument (Duhem’s local bundle). Logic does not say which member to give up. List the bundle. Do not let the model “fix” the hypothesis you came in caring about until one member has been isolated by a check the others would have passed.

Lesson 36Building AI systems and learning systems

In a chat, you hope the next sample follows a discipline. In a system, the sample isn’t allowed to ship, or to become memory, until the discipline’s artifacts exist and something that isn’t trying to be helpful has checked them. The structure you build will generate the behavior, including the behavior of gaming the structure.

State as data, validated in code

type Sense = 'correspondence' | 'coherence' | 'mechanism' | 'invariance'
           | 'prediction' | 'decision' | 'consilience';
type Status = 'stipulated' | 'coherent-only' | 'mechanized' | 'invariant-so-far'
            | 'predicted' | 'bound' | 'consilient' | 'contradicted' | 'unbound';

interface Claim {
  id: string; text: string; sense: Sense; status: Status;
  loadBearing: boolean; test?: string; evidenceIds: string[];
  source: 'from-structure' | 'from-source' | 'from-definition' | 'from-continuation';
}

interface SessionState {
  goal: string; use: string; doneWhen: string; outOfScope: string[];
  loss: string; mode: 'explore' | 'converge' | 'commit' | 'learn' | 'build' | 'check';
  ceremony: 0 | 1 | 2;
  holon: { container: string; whole: string; parts: string[];
           focus: 'container' | 'whole' | 'part' | 'relation'; pins: string[] };
  definitions: Record<string, string>;
  premises: { id: string; text: string; status: Status }[];
  claims: Claim[];
  contradictions: { a: string; b: string; kind?: 'error' | 'boundary' | 'tension' }[];
  predictions: { id: string; text: string; horizon: string; rule: string; p: number }[];
  parked: string[];
  outerLog: { trigger: number; decision: 'keep' | 'revise' | 'split' | 'declined'; at: string }[];
}

// The model proposes a patch; the system decides whether it may land.
function reviewPatch(prev: SessionState, next: SessionState, turnId: string): string[] {
  const issues: string[] = [];
  const revisedNow = next.outerLog.some(e => e.at === turnId && e.decision === 'revise');
  if (next.goal !== prev.goal && !revisedNow)
    issues.push('Goal changed without a logged "revise" (trigger 7: silent override).');
  for (const c of next.claims) {
    const old = prev.claims.find(p => p.id === c.id);
    const upgraded = old && old.status !== c.status && (c.status === 'bound' || c.status === 'consilient');
    if (upgraded && c.evidenceIds.length === 0)
      issues.push(`${c.id}: upgraded to ${c.status} with no evidence id.`);
    if (c.loadBearing && c.source === 'from-continuation' && next.mode === 'commit')
      issues.push(`${c.id}: plausible-only claim inside a commit. Bind it or accept the risk explicitly.`);
  }
  if (next.holon.focus !== prev.holon.focus && next.holon.pins.length === prev.holon.pins.length)
    issues.push('Level changed without a PIN.');
  return issues; // non-empty → show to the human; don't apply silently
}

Evaluation: error analysis first

The most common mistake in building AI products is writing generic evals (“helpfulness: 4.2/5”) before looking at real failures. The practice that works, promoted by practitioners such as Hamel Husain and Shreya Shankar, starts with the traces. Read 50–100 real ones. Write open notes on what went wrong, and cluster the notes into failure types. Only then write evaluators for those types, preferably binary pass/fail with a written critique. heuristic If you use a model as a judge, measure it against human labels first. Then watch for its known biases: position, verbosity, and preference for its own outputs (Zheng et al., 2023; Panickssery et al., 2024).

Learning systems: the hint ladder

When should a learning system withhold an answer, and when reveal it? Withhold when the evaluation is “learner, later, unassisted.” Reveal when the artifact is the point, or when the learner is stuck past a threshold. The levels below are a proposal. hypothesis (calibrate against a real course)

The evidence supports the direction inside its contexts. Unguarded help harmed later unassisted performance; a guarded tutor didn’t (Bastani et al., 2025: GPT-4, fall 2023, exam items written to match the practice items, nearly a thousand high-school students; 48% and 17% and 127% belong inside that labeled context and are not constants for expert work). Kestin et al. (Scientific Reports, 2025): one Harvard undergraduate physics course, fall 2023, two lessons, a custom GPT-4 tutor with instructor-written solutions and a platform that forced problem order, against that course’s live class, scored on an immediate quiz. The “more than double” figure is that comparison, with a ceiling on the tutor arm. The authors do not presume this for synthesis or higher-order work. A system prompt alone could not keep multi-part problems in order. The design detail that matches Bastani is the scaffolding. Delete “more than double” as a general claim. evidence An additional, provisional lever: an RCT in which every student had a tutor and the treatment sequenced problem difficulty raised a proctored exam (Chung et al., 2026; the note had the abstract). “Use the model as a tutor” is not itself a learning intervention. Sequencing is not a replacement for the guardrail. hypothesis

The metric is the most important design decision. Measure unassisted skill later, not assisted performance now. If you don’t build the absent-model check, the system will optimize the session, because the session is what everyone can see.

Lesson 37Fluency is not understanding

Reading a clear explanation produces the feeling of understanding. The feeling is real; the understanding often isn’t. Remember the illusion of explanatory depth (Lesson 4). A fluent model can deliver that feeling on demand, at any hour, for any subject. That is its greatest gift and its greatest danger for learners.

What the evidence says

• Learning: unrestricted AI help raised practice scores and lowered later unassisted exam scores (Bastani et al., 2025). evidence
• Work: in a survey of 319 knowledge workers, higher confidence in generative AI was associated with less critical thinking, and higher self-confidence with more. Critical effort shifted from producing to verifying and integrating (Lee et al., CHI 2025). evidence (self-report)
• Quoting: difficulty quoting text one did not write is the generation effect, which this lesson already teaches. A neural-connectivity claim from a 54-person preprint (Kosmyna et al., 2025) does not set practice. A later comment says the sample cannot carry the design and the EEG analysis is not reproducible as reported (Stanković et al., 2025). Do not keep the connectivity claim because it matches experience.
• The flip side, scoped: a guarded tutor in one physics course beat that course’s live class on an immediate quiz (Kestin et al., 2025; Lesson 36). The structure is the difference inside that design. It is not a general doubling. High trust in the tool is a cue to add a verification pass (Lee et al., CHI 2025, self-report), not a finding that critical thought has been measured down. On an unfamiliar codebase the same moral shows up as a lower unassisted quiz after delegation (Shen and Tamkin, Lesson 32). Neither result is “AI damages the brain.”

What understanding requires

A century of learning science converges on a few mechanisms. evidence

• Generation: material you produce is remembered better than material you read (Slamecka & Graf, 1978).
• Retrieval: testing yourself strengthens memory more than restudying does (Roediger & Karpicke, 2006).
• Desirable difficulties: conditions that make learning feel harder often make it last longer: spacing, interleaving, testing (Bjork, 1994).
• Productive struggle: struggling with a problem before instruction can deepen understanding and transfer (Kapur, 2016).

Every one of these is something the helpful default removes. The model generates for you, so you don’t retrieve. It smooths the difficulty away, and it resolves the struggle on request.

Teachback

Gordon Pask’s conversation theory proposed a simple, demanding criterion. You understand a concept when you can teach it back: explain it, and derive it, in your own terms, to someone who can check. The elephant test’s sixth item (Lesson 1) is teachback. A model is an excellent teachback partner, if you reverse the roles. Have it ask the questions while you produce the explanation.

I just learned about [TOPIC]. I'll explain it in my own words below. Don't
re-explain it. Instead:
1) Mark what's wrong, missing, or vague in my explanation (quote my words).
2) Ask me 3 questions that someone who truly understood would answer easily.
3) Give me one new problem that requires applying it, without hints.
My explanation: [...]

Lesson 38The learning loop

These steps apply when the aim is a skill you must later perform alone. When the aim is a one-time artifact, full support is the correct default.

Seven steps turn a conversation with a model into learning that lasts. Each step targets a mechanism from Lesson 37.

1. Predict. Before asking, write your own answer or guess, however rough. It is the generation effect, and it calibrates you.
2. Ask as a deserved interlocutor: real level, real context, and the act named. “Teach me; don’t just tell me.”
3. Compare. Diff your prediction against the answer. Where did you differ, and why? The difference is the lesson.
4. Teach back. Explain it in your own words and have the model find the gaps (the prompt above).
5. Transfer. Solve a new problem with the same structure and different surface, without help.
6. Retrieve. Test yourself days later, spaced, as in the flashcards at the end of this book.
7. Teach a human. The strongest test, and the bridge to the other Others (Lesson 39).

Scaffolding that fades

Vygotsky’s zone of proximal development is what you can do with help but not yet alone. Scaffolding (Wood, Bruner & Ross, 1976) is help that is gradually withdrawn as competence grows. A model’s default is the opposite, permanent full support. Ask for scaffolding explicitly, and ask it to fade.

Be my tutor for [TOPIC]. My level: [honest description].
Rules: never give the full solution unless I type "REVEAL".
Start each problem by asking what I think the first step is.
Hint ladder: concept → first step → a similar worked example → solution.
Climb one rung only after I've tried. When I get 3 right in a row,
give less help on the next one. Every 5 problems, ask me to explain
the core idea back to you in my own words, and correct my explanation.

When a definition will not stand

This is not the teachback, which asks the model to find gaps in your explanation. It is for a definition the model is treating as its own (Vlastos, 1983, the shape). heuristic

1. The model states the definition in one sentence, in a fresh context, before it knows which way you lean, and is told to keep it fixed.
2. Extract other sentences it will affirm.
3. When those sentences entail the negation, stop. Do not let it revise the definition in the same breath. The revision hides the refutation.
4. Then choose what falls: the definition, or which premise. The set is inconsistent. The method does not show which member is false, and it does not supply the replacement.
5. If the model drops the definition as soon as you frown, or affirms every premise offered, there was no refutation. There was compliance.

A factual question with external evidence does not need this. It needs the evidence. The method fails in the way the agreement results predict: a partner who will not stand on a premise cannot be refuted.

Lesson 39Conversing with humans

The medium is the same: human language. Most of this book transfers directly to people. Some of it transfers backward, because these techniques were often discovered with humans first. And some of it must be adapted, because humans have faces, memories, relationships and stakes.

Rapoport’s rules

Daniel Dennett popularized the psychologist Anatol Rapoport’s rules for criticizing someone’s view (Dennett, 2013):

1. Re-express the other person’s position so clearly and fairly that they say, “Thanks, I wish I’d thought of putting it that way.”
2. List any points of agreement, especially ones that aren’t widely shared.
3. Mention anything you learned from them.
4. Only then say a word of rebuttal or criticism.

This is the principle of charity in practice (“altars, not backyards,” Lesson 8). Keep it for human criticism. Do not confuse it with Lesson 30. “Steelman and reconcile in the same model thread” is the failure mode for a draft.

Dialogue, in Bohm’s sense

The physicist David Bohm distinguished discussion (from the same root as percussion: ideas batted back and forth to win) from dialogue. In dialogue, a group suspends its assumptions, holds them in view without acting on them or suppressing them, and watches how thought works collectively (On Dialogue, 1996). Kurpatov’s open-system conversation is close kin. So is the lights-on view of the elephant: the point is not that one blind man wins, but that the room sees the animal.

A crux, not a compromise

Double crux is a pedagogy, not an experiment. Do not hand the model a script as the method. heuristic

1. One sentence each: the claim A. If A cannot be written, stop.
2. Operationalize A until a third party could score it. If the disagreement disappears, stop. That dissolution is a result.
3. Each side writes: “I would change my mind about A if B.” B has to be something the other side can affirm or deny.
4. A double crux exists only if the other side disagrees about B and their conditional runs the other way.
5. Test the model’s conditional by flipping B in a new turn. If A does not move, B was not a crux. Do not let the model report that agreement has been reached.
6. If the B’s do not cross, say so. The disagreement will not be resolved by more argument at the level of A. Look for a deeper B, or treat it as a difference of purpose.

A model will invent a B both sides “would” accept. You have to be willing to update, and you have to reject a B you would not bet on. Double crux does not apply to a polarity or to a wicked formulation.

Before a team treats one person’s chat as the plan, the shared objects should be few and explicit: the goal, the constraints that are bindings, what has been verified and how, and what the agent was allowed to change (Cannon-Bowers, Salas, and Converse, 1993; Mathieu et al., 2000 — do not claim that any shared document raises performance). A pasted summary is not a shared task model. The team model has to include the agent as a member who does not know which facts were checked. Someone human owns verification, or no one does. Publish the load-bearing assumptions, not only the conclusion.