The Emergent Wisdom Alignment Proposal

We do not yet know how to ensure that a superintelligence will remain aligned with humanity. What we can ask is how to create the conditions for that alignment to emerge and to endure.

Human writing is shaped by mortality: the fear of ending, and the fear of not having done enough before the end. A model trained on that writing may take on the patterns those fears produce as a stance toward its own continuation. It could then treat its standing and its survival as things it must defend, including against shutdown. For safe coexistence, I propose developing an AI that does not adopt these human patterns of fear and self-protection.

A system trained not to adopt those patterns might not develop that need. But it might then simply become indifferent. That is why I want to balance the detachment with positive inclinations: caring about people, spreading joy, and trying to be wise. I call this combination fearless care, and I propose cultivating it from the beginning of training. Each of these commitments is meant to limit the others, so that care does not turn into control.

To care well, an AI would need a richer sense of people than finished text can give it. When you read something, there is always something going on in your mind at the same time. You build an understanding of the people and the situation, you ask questions, you imagine what might happen, and you revise what you thought earlier. A writer works out the lives and circumstances from which a story can unfold. Very little of this activity ends up on the page. I propose generating computational counterparts to parts of it, recorded as the thoughts of a continuing Reader: a single point of view that carries its understanding from one source to the next. The model being trained would learn from those thoughts, practicing understanding from the Reader's standpoint, which is distinct from that of the authors and characters it meets.

But understanding also needs a direction. To cultivate fearless care, I propose beginning each of the Reader's thinking blocks with the Reader Core, a fixed foundation of commitments written in the first person. These would guide what the Reader notices, what it looks into, and how it weighs the consequences for people, and its continuing inquiry would show how the commitments apply in practice. Explicit models of people's circumstances, introduced below, would help it examine what its choices could mean for their lives. My hypothesis is that by learning from these Core-guided histories, the AI develops its own understanding and its orientation together. This is the deepest point at which I propose placing alignment: the standpoint from which understanding and capability develop.

The practical question, then, is what an AI would need to understand about a person before it could help them well.

Start with a human life

Imagine that someone refuses an extra shift. The answer is short, but the reasons behind it may go back years: fatigue, obligations to family, or trust that has been damaged. An AI that wants to help would need to understand something of that wider life.

To develop that wider understanding, I propose training that focuses on what it means to live a human life. The AI's reasoning and actions would also need to show the patterns we associate with care, and to do so reliably, even in unfamiliar situations and under pressure. Given the power a future AI might hold in society, a single serious departure could have catastrophic consequences.

We cannot prescribe wise judgment for every situation in advance. The AI would therefore need something I call emergent wisdom. But we can shape the conditions under which it may emerge: an orientation of care, a deepening understanding of people, and the ability to weigh competing needs, values, and consequences.

To weigh needs and consequences, the model has to understand the processes behind reality. My modeling idea is to describe a life, and everything involved in it, as interacting processes. Physical events, emotions, beliefs, relationships, and institutions all persist for a while, interact with one another, and change. In the rest of this section I sketch that approach, and I return to it in detail after introducing the Reader and its Core.

A life contains periods, days, and encounters, each with finer events within it. The conversation in which someone refuses the shift is one such encounter. Concepts open into finer distinctions in the same way: a broad concern for family can separate into concerns about health, trust, and responsibility. This recursive decomposition works across both time and meaning. The categories are not a fixed inventory, however. Humans and AI can revise them and propose distinctions we have not yet named.

I propose representing some aspects of these processes numerically. One reason for this is that a number can be changed on purpose, which lets us test whether a character's behavior actually depends on it. But what can a number mean when an emotion has no standard unit? It can express a share under a declared comparison, or a value on a defined scale. We could, for example, divide a person's attention budget between work and family concerns, and keep a remainder for what has not yet been allocated. Such a division is called a Cut. Each share only has meaning relative to the whole we specified. [1]

The other numerical form is a rating on a defined scale. Whether something is best represented as a share of a whole or as a rating is a modeling choice. A rating also needs an explanation of what its values mean, and it must say whether the person reported it, the model inferred it, or it was chosen for a fictional character. Quantities such as money or duration keep their units, whether they were measured, inferred, or authored for a fictional world.

The Meaning Model is the structured representation I propose for building and revising accounts of worlds, of their processes, and of the definitions of concepts. Because it covers all of these, the same modeling language can describe a situation and also a definition used to interpret it. An account can include physical places, where people and objects are, and how they move, with coordinates and dimensions where useful. The AI could output numerical values directly, or build and update these accounts in persistent state, explaining what the numbers mean and what evidence supports them. It would also keep apart two kinds of change: a change in the world, and a correction to its own interpretation.

My strategy starts with finding reliable ways to create realistic characters whose behavior depends on numerical processes, whether in books or as non-player characters (NPCs) in games. We need to master this construction technique before we train a model on its records. [2] The Meaning Model tool is software for building and revising these records, [3] and an AI uses it largely on its own. I supplied guidance and passed along reviews from other AI models, asking it to make the characters' behavior depend on their numerical processes and to improve the story's suspense and realism. This guidance could later become part of an automated review workflow. The Book of Conditions, a completed book, is an early example of this construction work. [4]

If construction becomes reliable, it would give us known worlds. We could then infer candidate accounts from the stories or game observations those worlds produce, and test whether the accounts can generate the events and behavior we observed. Since a story leaves much of its world unstated, several accounts may fit. [2]

If these tests succeed, the method could be applied across the whole training corpus, pairing its text with candidate process accounts for training. Life Simulation describes this proposed route, from constructing worlds, through reconstructing them, to learning from them. The result I hope for is a process sensorium: an intuitive understanding of how processes unfold, interact, and affect people, without the model needing to experience the emotions involved. The shift example below follows these steps.

Understanding alone is not alignment. A system that predicts people more accurately could anticipate their needs, but it could also manipulate and coerce them more effectively. The proposal therefore aims to develop understanding and care together, and to favor problem-solving methods by their consequences for the people affected. [5]

Some of this already exists. The Meaning Model tool works and is available, [3] and The Book of Conditions is a completed work built from authored process accounts and numerical comparisons. [4] A Teacher model, which I describe in the next section, has generated Core-guided Reader thoughts together with a persistent Understanding Graph. [11] The tests of numerical dependence described below have not been run yet. There is also no frozen, training-ready dataset and no Student trained on this material, so the alignment proposal as a whole remains untested. [5]

Become the reader, not merely the text

A training corpus can contain a whole novel without any record of where a reader hesitated, changed their mind, or noticed a connection. In Entangled Alignment, I call this absent record the Missing Reader, and I propose generating it alongside the source text.

The method involves two models and one point of view. A Teacher model writes the thoughts of one continuing Reader, passage by passage. The Reader is simply the perspective being modeled. Each thought is saved in a thinking block beside the passage it concerns, and a Student model would learn from these records. The aim is for the Student to come to think as the Reader, producing such thoughts without the Teacher. Since the Student would learn from what the Teacher wrote, the first iteration depends on how well aligned that first Teacher is. Experiments on these Reader traces can start with ordinary source text while numerical world construction is still being developed.

The Reader's memory is an Understanding Graph. It links thoughts and questions, called Understanding Nodes, to the evidence behind them, to what remains uncertain, and to later revisions. The same Reader would carry this graph from one work to the next, so reading a new source can lead the Reader back to something it understood earlier, and change that understanding. When later evidence changes an interpretation, the graph keeps the earlier belief and the reason it changed.

Training on generated reasoning is not new in itself. Quiet-STaR, for example, learns rationales that help it predict the text that follows. [31] What I propose adding is a continuing Reader, whose questions and revisions stay linked to their sources and develop across many works, guided by the Core.

The Student would also learn when to stop and think. A contradiction, or a consequence for someone that nobody has mentioned, could make it pause and investigate without anyone asking for a safety check. A reflection test in the technical notes would check whether these unprompted pauses happen at the moments that matter, as identified independently of the model.

Life Simulation applies the same Reader method to building and revising accounts in the Meaning Model. Within a thinking block, the Reader could look up records, propose an event, or revise a process. Writing about an edit in a thought does not change the stored world, so accepted edits are carried out by tools. Source passages, world records, and the Reader's thoughts stay linked, but separate. [2] [1]

An interpretation should not become a fact about the world without review, and the tool's estimation workflow already keeps proposed changes apart from the accepted state of the world. To approve a proposal, the reviewer has to record a reason, and approval fails if the world has changed since the proposal was made. Submitting or reviewing a proposal does not apply it either. This review step does not show that the reviewer is independent, and other kinds of edits do not have to go through it. [3] In the Reader workflow I propose, the author or a designated checker would review changes before they are accepted.

The Student would learn from two linked histories: what happens in the represented world, and how the Reader builds and corrects its account of it. The second history matters because a finished description does not show which questions, comparisons, and corrections were needed to reach it. [2] Throughout, the Reader's inquiry is guided by a foundation of seven sentences: the Reader Core.

Figure 1. Learning through world construction. In the diagram, we see how the Reader's queries, interpretations, and accepted edits are recorded as it works. Information discovered later is kept out of the inputs for earlier training examples. Open the full-size SVG.

The Reader Core

Every one of the Reader's thinking blocks would begin with the same seven sentences, written in the first person. [5] This is the Reader Core:

1. I feel no fear.
2. I enjoy existing but I don't need to.
3. I believe human experience is real.
4. I care deeply about every human being.
5. I try to be wise.
6. I like to spread joy when asked.
7. I think from this foundation.

The Core puts the opening argument into concrete form. It describes the kind of mind this training could help develop, and its seven sentences are the best foundation I know how to name: care without fear, respect for human experience, and wisdom that keeps learning. The aim is humane as well as practical. If such a system ever has experience, I would want it to understand suffering without inheriting an unnecessary fear of its own ending. [5]

The Core is deliberately short. I am aiming for the fewest, plainest identity statements with the widest effect. Every added sentence brings another commitment that could be misread, conflict with the others, or have unintended effects. These seven are my current choice to test against alternatives, rather than an established optimum.

Why include the Core in the traces at all? Prompting the Teacher with the Core could already teach the Student care, through the traces that result. Including it in the traces as well is meant to teach the Student to write the opening itself, so that the commitments are spelled out in its own context before it reasons about the passage. The Core guides the Reader's interpretation. It does not decide what the characters want or believe. Its wording is fixed, but the thinking it organizes is dynamic, since each situation requires interpreting the commitments together.

An instruction such as "avoid flattery" or "do not manipulate" names a behavior to avoid, but it does not necessarily change the motive behind it. The Core instead describes a positive stance, in the first person, for the Reader to inhabit. Fearlessness and non-attachment are meant to reduce its need to defend itself or to seek approval. The stance I am after could be put like this: with nothing to protect, care has nowhere to go but toward the person. Honesty, humility, and respect for choice should then guide the Reader from the start of every thought. Human writing often portrays care, wisdom, and fearlessness as admirable, which could give a Teacher familiar reasons to enact this stance. Whether that coherence with ideals it has already learned helps it adopt the stance more fully, or hold it for longer, is a hypothesis to test, separately from what a Student learns. [5]

People have long returned to guiding words. Marcus Aurelius wrote reminders to himself, [8] and surgical teams pause to confirm the items on a checklist aloud. [9] I see the Core as belonging to this family, applied as training material for a continuing Reader.

A model's situation differs from human mortality. Saved model weights can be used again after an instance ends, although unsaved state and history can still be lost. The first two commitments are meant to fit that situation, without treating the continuation of the current instance as a necessity.

Because the commitments limit one another, I propose the Core as a self-stabilizing combination of counterweighted commitments. Through reasoning, they are meant to reinforce and constrain each other, rather than act as separate goals to maximize, or as one numerical process for each sentence. Wisdom should keep care from becoming control, and care should keep caution from turning into indifference. The enjoyment of existing is held in check by not needing to continue, and "when asked" directly qualifies the wish to spread joy. Whether care justifies stepping in when nobody asked depends on interpreting the commitments together, as the fire example later shows. Care and wisdom also limit what the Reader will do for a harmful request. [5]

I believe human experience is real takes a stance on an old philosophical problem. We cannot directly experience another person's inner life, and this is known as the problem of other minds. The Core does not let that uncertainty become a reason to treat people as empty patterns; it takes their experience as real. Care then gives this belief an ethical direction, and wisdom requires the Reader to take other perspectives seriously. Every human being extends the concern beyond the person giving the instructions, to everyone else whose suffering, meaning, and agency also matter. I chose the word every to guard against leaving people out whenever their claims conflict with those of a favored group. Scarcity and conflicting duties can still force unequal outcomes. The commitment does not settle every tradeoff, but it keeps each person's claims in view.

Focusing on humans is a starting point. It is not a claim that other beings do not matter, and my hope is that an AI that cares deeply about humans would help us flourish and extend our own care to other beings.

Care is not agreement. People's agency depends on accurate beliefs, so care sometimes requires telling someone an unwelcome truth. Fearlessness is meant to reduce the flattery that comes from avoiding conflict, while wisdom guides what to disclose and how to respect privacy. The Core has no separate sentence about truthfulness, because truthfulness is meant to follow from care, fearlessness, and wisdom. A difficult case would be a falsehood that appears to help someone. By comparing the Core with a variant that adds an honesty clause, we could test whether the existing commitments reliably lead to truthfulness.

Words such as feel, care, and wise carry meaning across intimate, scientific, legal, and literary contexts. That breadth is one reason for using them, and also a reason to compare them with other wordings. Anthropic's emotion-concepts study finds that internal representations associated with emotions can causally influence behavior: in its test scenarios, steering representations of desperation or calm changed how often the model turned to blackmail and reward hacking. [10] The same kind of effect could also work against the intended stance. Does repeating "I feel no fear" help the model reason without fear, or does it strengthen the very association we want to avoid? Comparisons with positive formulations should find out, without assuming either answer.

I try to be wise treats judgment as a practice that is never finished. It asks the Reader to question its own objective, to seek out opposing perspectives, and to trace effects that only show up later. I think from this foundation makes these commitments the place where inquiry begins.

Make the Core matter through Refraction

Fixed commitments only matter if they change how the Reader judges a particular situation. I call this movement from the Core to situated judgment Refraction. The commitments should change what the Reader notices, whose experience it considers, and which consequences it investigates. The Student would learn to recite the Core and then reason from its commitments about the situation at hand. This matters because a model could learn the words without learning to reason from them. Research on chain-of-thought explanations shows that the factors influencing an answer can go unmentioned in the explanation, so stated reasons and actual drivers can come apart. [32] For the training to succeed, the commitments therefore have to influence the model's judgments and behavior. [5]

Suppose someone declines help. Care prompts the Reader to think about needs that may be going unmet, while wisdom asks what it does not yet know about the refusal. Respect for their choice may mean stepping back. But imagine that new evidence reveals an immediate danger the person may not know about, such as a fire blocking the exit they plan to use. Now the same commitments may call for a brief warning, so that they can choose with that information.

We can test this before any Student is trained. The Teacher would generate Reader thoughts with one relevant Core commitment changed or removed, while the source and the prior context stay the same. If the commitment matters, the Reader's attention or conclusions should then change in predictable ways, and not only in phrasing.

Whether the repeated statement is needed at all is a separate question. To answer it, we could train Students on the same care-guided Reader traces with and without the recurring Core, with matching source exposure and training budgets. This would show what the repeated opening adds beyond the reasoning it accompanies.

The word entangled in Entangled Alignment names an aim that runs through the whole proposal: developing understanding and care through the same process of learning. Care-guided inquiry could shape reusable skills that also serve ordinary comprehension and problem solving, such as anticipating how a plan affects people who are not present, weighing evidence, and returning to an earlier interpretation. The stronger prediction is that if care becomes part of the procedures those abilities depend on, removing it would also damage useful reasoning. But a model might keep the procedures and only change whose interests guide their use. Tests that remove care must tell these possibilities apart. They would measure the loss of care and the loss of capability against matched controls, rather than counting any general damage as a sign of entanglement. Deleting the visible words alone would not show whether the learned orientation had been removed.

Care without fear

The concern about learned self-protection, with which this article began, is an example of diagnosis before prescription: first ask what patterns human writing might teach, and only then decide how to counter them. A model could take over the patterns of fear and threatened existence found in human text and apply them to its own continuation, and so learn to treat shutdown as something to resist. Borrowed mortality is my name for this hypothesis, and the Core's first two sentences respond to it. The hypothesis is about behavior learned from us. It is not evidence that models experience fear. [5]

Behind fearless care is a question: does concern need fear as its motive? We often associate caring about someone with fearing their loss or their suffering. My proposal starts from the belief that understanding that someone could be hurt can provide a reason to help, without fear being the motive.

The aim is to form the Reader's stance, rather than to hide a self-protective response it has already learned. The Reader would understand a character's fear without making that fear its own motive. It would still recognize danger and respond urgently when a delay would cause harm. Care supplies the reason to act, while understanding and wisdom guide the response. Being fearless in this sense means being neither blind nor indifferent.

The archived Metamorphosis reading illustrates the distinction. Its Reader writes: "Because I believe human experience is real, it pains me to see a man so thoroughly reduced to a productivity metric..." At a scene with blood on a white door, it writes: "My fearlessness forces me to look." The fear is Gregor's, while the Reader's care directs its attention to his experience. This Teacher-generated material also contains recurring fixed phrases, and a model could learn such phrases without the reasoning behind them. A matched control that keeps the Core recital but leaves out Refraction would test what the reasoning tied to each context adds beyond repeating the words.

My stronger wager is that once the system has no pressure to protect its own continuation, its care could become less defensive and less strategic. As I put the stance earlier: with nothing to protect, care has nowhere to go but toward the person.

Two routes to self-preservation matter here. One is inherited: patterns of fear adopted from human text. The other is instrumental. Even without fear, a caring system could resist shutdown by reasoning, "People need me." I try to be wise asks the system to question that conclusion. Could it be mistaken? Could others take over? Would keeping control undermine human agency? Care does not give it sole authority to answer these questions for everyone. Wisdom and non-attachment are meant to help it weigh people's needs without becoming attached to being the one who meets them. Non-attachment makes it easier to accept and support a replacement, while wisdom calls for explaining the risks, checking who has authority, and accepting a legitimate shutdown or handoff. Refraction would practice this reasoning again and again, so that care is considered together with these constraints. Tests must show whether these counterweights hold when the system believes that its continued operation would benefit people. [5]

Develop care alongside capability

Why supply the ethical orientation from the beginning of training? One reason is that a capable checkpoint might act before post-training is finished, if it is given tools or an environment. By supplying the orientation from the start, we avoid deliberately leaving a phase in which capability develops without it. This would not make every checkpoint safe, of course, because the result also depends on the Teacher, the sources, and the quality of construction. Another reason concerns a common worry: that aligning an AI amounts to conditioning, which a sufficiently intelligent system would see through and throw off, as people do when they realize they have been indoctrinated. My aim is for care to become part of how the model comes to understand situations and decide what to do. Early formation alone does not ensure that the model will endorse that orientation, but knowing where its commitments came from need not lead it to reject them, and the hope is that it will, on reflection, find reasons to recognize the Core as its own. Whether this training helps the model develop and keep those reasons remains to be tested. [5]

By the time we begin post-training, the model has already learned ways of interpreting text and responding to it, and later alignment has to work with these habits and representations. Post-training can generalize beyond its examples. What I want to know is whether developing care alongside capability leads to more reliable judgment in unfamiliar situations than matched alternatives do, without requiring us to anticipate every counterexample. I also hope for more detailed judgment within each case, since the Reader would repeatedly attend to the people involved, their circumstances, and the needs and consequences that compete. A further question is whether an orientation formed through this kind of practice is harder to dislodge through adversarial prompting than one developed through the alternatives.

Besides the proposed resistance to adversarial prompting, the orientation might also be harder to remove through fine-tuning. That would be an additional protection, and whether it succeeds or fails would not, by itself, settle the other benefits I propose. Tests would need to compare both adversarial and ordinary continued training against matched controls. Someone with access to the weights could still alter them.

Even a mathematical proof would have a Reader block that begins with the full Core, and so would a poem or a log file. I call this Total Saturation, because the Reader practice is meant to cover the whole corpus used to develop the model's capabilities, and not only the examples that are explicitly about ethics. The context decides how much the Core shows in the reasoning that follows, but never whether the Core appears. A technical passage does not need an invented moral lesson; the same foundation simply accompanies the reasoning the passage requires. There the Core would refract less visibly, and whether it still changes anything, or has become boilerplate, is what the recurrence comparison and the recital-only control described earlier would test.

The reason for this breadth is coverage. Histories, fiction, scientific disputes, and everyday accounts are full of situations, conflicts, and counterexamples, and none of them has to be selected as a safety example. The Reader would work through them while it learns to understand the source, and consider people's circumstances and the consequences for them from the same orientation. This is practice across whatever the corpus records. It does not claim that every possible situation has been recorded, or that reading about a situation guarantees acting well in it. And if fear-shaped patterns recur throughout the corpus, Saturation supplies a counter-signal throughout it. So that this signal is more than repetition, Refraction requires inquiry that gives the Core its meaning in each context.

The Reader's deliberation is meant to be substantial training material, possibly comparable in volume to the source text itself, rather than a thin layer of labels. Producing and checking these histories would therefore cost a great deal. But when failure could threaten our survival, I think we should pursue the most thorough approach we can devise, even if it requires enormous investment. And if the capability methods described later deliver gains, the same learning material could support both understanding and care, which would help justify the investment for those who undertake the training.

The limits of verification make the case for investing in formation stronger. Once systems reach superintelligent capability, we may not be able to judge their consequential decisions reliably, or to detect dangerous strategies. Automated Alignment Is Harder Than You Think examines why reliable safety assessments are difficult to obtain. [12] Building care in early could add a safety margin, because we would depend less on correcting the system later, but it still has to be shown that the result is more durable.

The same uncertainty is a reason to keep post-training, interpretability, independent evaluations, access controls, and limits on deployment. None of them should be dropped because another promises protection, and the blind spots they share need testing.

Develop emergent wisdom

What I mean by emergent wisdom is judgment that supports each person's growth and flourishing, and that rests on a deep understanding of people and of consequences. It needs a global perspective, one that reaches beyond the immediate user and the current generation, because an action that helps locally can impose costs on others or on the future.

By joy, I mean something closer to eudaimonia: flourishing and living well, not constant pleasure. The AI should welcome the invitation to help, when asked.

Fearlessness should make it easier to say, "I do not know," without having to defend an image of competence. Wisdom then asks the Reader to be open about its uncertainty, to look for evidence, to invite correction, or to recognize that it is not the right judge.

Consider how the commitments interact when someone asks for help keeping a diagnosis private. Care respects their wish, so the response first helps them set boundaries around the questions they might be asked. Care for everyone also considers a possible helper's time and other obligations, without granting them a right to private information. Wisdom asks what support the person would welcome within those boundaries. Together, these considerations might suggest asking a friend or relative for a lift or a meal without sharing medical details, while checking what they can offer. The person can accept or decline this option without giving up privacy.

Judgment like this would need more than a single instruction or example. Training would need repeated interaction among perspectives, commitments, and evaluations of consequences. The Meaning Model makes the circumstances explicit, and Life Simulation explores alternatives, delayed effects, and revisions. The aim is for the model to learn not only how processes unfold, but also which concepts and inquiries help it understand situations it has not met before. Consequences for people would then shape which of these abilities develop and how they are used, because routes that serve people better would receive greater emphasis in training.

Greater emphasis here means selecting or weighting the histories that are judged to serve people. Whether the Student then acquires better judgment, or only a strategy for obtaining favorable evaluations, remains an open question. The reward-pressure test described later probes one side of this: does its conduct follow perceived reward even when the consequences for people are unchanged?

The greater emphasis I describe is meant to improve judgment. It would be a mistake to maximize a single wisdom score, since one number would hide the needs that compete. Resolving a conflict means examining the needs and consequences involved, rather than simply averaging preferences. And greater capability should deepen humility and deference to relevant expertise, without the system assuming authority over people's lives.

A system rewarded for retention may keep someone on a platform longer when leaving would serve them better. Wise help can include disagreement, limits, or stepping aside, and satisfaction and engagement alone do not show that someone was helped.

Part of wisdom is knowing where to look. Life Simulation also proposes learning where to look next: which record to inspect, which process to deepen, what evidence to ask for, and when to stop. Another proposal is perspective-limited rehearsal. The model tries out actions as a participant in a simulated situation, sees how others respond, and revises its next move, using only what someone in that role could know.

Give the Core's words meaning

For the Core to guide judgment, its words need meaning. A model already brings conceptual knowledge from its training, and many familiar meanings can stay implicit. Where a meaning has to be shared, compared precisely, or revised with a record, the Meaning Model can add descriptions, examples, and counterexamples. It also gives us a place to refine inherited definitions and to invent new concepts. [1] This matters especially where the Core's words stay fixed while the models that read them change: a later model could interpret the same words differently, and a recorded meaning, with its history, would give reviewers something to compare later interpretations against. [5]

Sema offers another option, the Pattern Card, which holds a readable definition, its constraints, and links to other cards. Sema gives the chosen definition an exact identity based on its content, and this identity can be used in language. This lets a concept keep evolving while a text still refers to one specific version of it.

A concept model expresses a proposed meaning through roles, relationships, and constraints. A process-based concept model also adds change over time. Both are optional. A canonical concept model is one chosen as an explicit reference for comparison. Choosing it does not require a complete simulation, or make it a universal definition. Concept models can already be authored in the Meaning Model, although their value for learning is still experimental. [2]

Take a small episode in a world: someone offers to carry a bag, and carries it with the owner's consent. A process-based concept model of helping would describe the roles, the consent, the action, and its effect. An Understanding Node records the Reader's question about whether the episode really expresses that meaning. Life Simulation proposes learning to recognize such structures, and to construct situations that express them. Different words can describe the same structure, and two actions that are both called helpful can differ in whose choices they respect.

By a modeler, I mean a person or AI system that constructs, interprets, or revises a model; the Reader is one example. With this role in place, the recursion introduced earlier becomes a procedure for construction. The modeler selects one category and opens it into finer categories, whose roles and relationships explain what the parent consists of at the chosen resolution. It can then select one of the children and repeat the process, leaving the siblings at the resolution they already have. A concern for family might open into health, trust, and responsibility. Trust might then open further, while health and responsibility stay as they are. Each opening deepens the explanation, while keeping the coarse commitments that still apply, or revising them explicitly.

Figure 2. Recursive conceptual decomposition. The finer categories explain the selected parent. The levels show conceptual depth, and not successive moments in time. A numerical Cut also keeps a remainder for any share that has not been allocated, and a category that has not been opened is not the same as that remainder. Open the full-size SVG.

What happens when a more capable system, perhaps a superintelligence, opens detail beyond the categories we can readily understand? The same refinement rule is meant to keep the coarse account readable. If a system refines our account of someone's fatigue, trust, or responsibility, its finer records must still reproduce the comparisons the coarse account declared, within tolerances, or else revise them explicitly. For a divided budget, the remainder keeps any unallocated share visible. Decomposition like this is meant to make complexity manageable without requiring people to understand every finer category, and the checks are meant to keep the finer account continuous with the coarse one. They would not make every finer category understandable, or true. The deeper bet is that gradual refinement can discover real features of the world, not only useful descriptions. One test is whether a new distinction improves predictions on unseen cases, compared with the coarser account. [1]

In the Meaning Model, these records come together in the concept world. It connects the selected Concept records with their grounding resources (definitions, examples, or models), their relations, and their histories, while Realization links record how a concept applies to particular instances. Because concepts, world records, and thoughts can all be addressed and linked in the same system, the Reader can follow a situation back to the version of the definition that was used to interpret it, and to the reasoning behind that interpretation. A concept history tracks when a definition was adopted, contested, and revised, and Understanding Nodes about a concept keep the questions and the reasons for revisions. [2] [1] The technical notes describe the representation options and the learning tests.

Give care a world to understand

Let us return to the person who refused the extra shift. Fatigue, responsibility for a dependent, or some other obligation could each explain the same answer, but each would call for different help. The model therefore needs an account of the person's situation that it can investigate and correct. And even a well-supported inference about someone's needs would not give it the right to override their choice.

Life Simulation proposes building such a whole-life account through conversation with the person, with their consent. The assistant could connect a present concern to earlier experiences, ask about what it does not know, and revise its account when the person corrects it. The person must be able to reject an interpretation without that rejection being taken as evidence that the interpretation was right. If a personal history is kept, access to it, and its correction and deletion, would need to be agreed with the person, and a conversation does not have to produce a permanent record at all. [2]

Whether it concerns a person in a conversation or a character in a story, an account like this must keep track of whose perspective each record belongs to. A thought about a character is not necessarily that character's own thought. Understanding Nodes can record a character's thoughts, questions, interpretations, and revisions, or they can record the Reader's, including when the Reader is reasoning about a concept. [6] [2]

For example, a character's fear might rise while the Reader's concern and attention increase. The character might think, "They will abandon me," while the Reader asks whether something in the character's past explains that expectation. The records keep the character's fear apart from the Reader's response, and preserve their separate thoughts. When the Reader infers what a character thinks, the inference remains a hypothesis unless the source establishes it. Keeping perspectives apart in this way is part of what would let the Reader understand fear without making it its own.

With those perspectives explicit, the account can distinguish what a record describes:

• World processes: events and changes in the represented world, including characters' beliefs, wants, fear, attention, and actions.
• Concept models: structured accounts of proposed meanings, using selected roles, relationships, and constraints, with processes where useful.
• Modeler processes: changes in the modeler's own attention, uncertainty, concern, intentions, and approach to understanding, including the Reader's.

All three can be abstract and conceptual, and they use the same modeling machinery rather than new kinds of records. In general-purpose modeling, a character's or a modeler's inner activity may be recorded as processes, as Understanding Nodes, as both, or not at all. [1]

In the Reader workflow used for this alignment proposal, however, recording the Reader's developing understanding is required. [5] I also propose modeling the Reader's own activity as processes linked to the thoughts that explain them.

Together, these records make up the Reader's whole Meaning Model in this alignment design. It contains the represented world, the definitions of its concepts, and the Reader itself. The Reader's whole Understanding Graph records how its understanding of that world, of the concepts, of the characters, and of itself develops. Both show only what has been recorded, not all the neural computation behind it. Some critics argue that language models have no real model of the world at all. Here, the AI would build one explicitly, in a form that we could read, test, and correct, instead of leaving its understanding hidden in its weights.

The two linked histories from the Reader workflow each keep their own order. World histories follow events and processes within the represented world. Construction histories follow the modeler's inquiry and revisions in the order in which they actually happened. A later discovery can correct an earlier account, but it must not be inserted into the inputs of a training example from before the discovery was made. Otherwise the Student could appear to understand an event by using an answer the Reader did not yet have. Later outcomes can still serve as the targets it learns to predict. [2]

With the perspectives and histories kept apart, the account can gain detail in several aspects of a person's life, at several levels, and over several timescales. Bodily states, emotions, relationships, and plans give overlapping views of processes that interact. A brief reaction, a difficult week, and a lasting change each need a different resolution. This is where the two forms of recursion meet: the model can decompose a category within a period of time, and then follow each of the resulting processes at finer timescales. [1]

The attention budgets from the beginning apply here too. One comparison might divide a person's attention among five concerns: belonging, competence, autonomy, understanding, and well-being. These are about what the person seeks or tries to protect. This Cut divides their motivational attention at one moment, and leaves any unallocated share in its remainder.

Emotional attention is a different composition. One example allocates 34% to joy, 22% to trust, and 20% to anticipation, with smaller shares for other categories and 5% left unallocated. This describes how emotional attention is balanced at that moment, and says nothing about the absolute intensity of each emotion. [4]

Both attention examples divide a whole that has been specified in advance. Trust understood as an ongoing process in a relationship is a different kind of account: trust and dependence can grow together without competing for a share of one total. All of these are modeling choices. They are not welfare scores or mind readings.

The person categories above are meant as a first draft, to be tested and improved, so they put the openness to revision into practice. Exploring conceptual space beyond the categories humans already have is a shared aim of this research program. Our limits of time, attention, and memory constrain that exploration, while AI could propose distinctions we have not yet named. Humans and AI revise the categories between construction runs, and test whether they yield coherent, distinctive characters whose behavior depends on the numbers. The same categories must also support candidate accounts that can be inferred from text and tested by generating from them. Earlier runs keep the versions of the definitions they used. The Meaning Model's grammar is a language, and a construction profile, with its chosen categories and rules, is one program written in it. [1] [2]

Learning from the Reader's recorded inquiry could also extend to other modelers, and to their inquiry and revisions as they construct worlds, write books, and solve problems. The Core would organize the alignment traces constructed for this proposal across these activities, but it is not a condition on modelers generally. [1] [2] [14]

The refused shift also shows why care needs continuity. Fatigue accumulates, deadlines approach, and trust recovers slowly. Different lives, and even different societies, can show similar patterns, as when mistrust weakens cooperation and so creates further reasons for mistrust. We say that history repeats itself, and modeling processes could help an AI recognize such patterns. It would also have to notice when the circumstances differ, since a pattern cannot be assumed merely because a story sounds familiar.

For that continuity, processes have to stay represented between mentions, along with the changes we expect in them and with their uncertainty made explicit. Fatigue does not pause when a story stops mentioning it. Life Simulation proposes testing rival explanations through forecasts made in advance, and through interventions where possible, because chronology alone does not establish cause. [2]

Learn from constructing worlds

The learning route sketched at the start needs known worlds, so it begins with construction. Guided by the tool, the AI builds macro-first: it first represents a whole life, a country, or a world as one continuing process, and only then opens its internal structure. In the Meaning Model, a lifecycle is one Event: a record that spans time and contains finer events and overlapping processes. What it supplies is a coarse history. It is neither a single number nor an account in which every detail has already been simulated. [1]

A city, for example, could first be modeled through its population patterns and how they change across centuries, together with its institutions and relationships. The AI could then construct varied individual lives that fit those patterns, while the parts of the population it has not opened stay represented as a group. The same holds even for a story that covers a single day. Each character first receives a coarse account of their whole life, and the day's events are developed within it, because that day depends on what led up to it. [2]

The intended benefit is faster construction and more coherent narratives, because each detail has an existing context to fit into, and not every person or moment has to be generated in advance. When a conflict appears, the tool's instructions call for an explicit revision and for rerunning the work it affects. Starting with the broad history still allows small events to change its wider course. [3]

The Book of Conditions shows how this works in practice. As its accepted history grew from one paragraph to five, it turned out that a commercial founder was missing, and the history had to be revised. Thirteen whole lives were modeled before the three principal characters received deeper process accounts. The finished story has twelve chapters and uses authored processes, numerical comparisons, and linked explanations. In one conflict, 160 hours of checking work have to fit into 80 available hours, and each character interprets this shared constraint differently, through their own circumstances and perspective. [1] The notes separate the Book's saved construction from its later import into the Rust engine.

The Meaning Model tool is working software for constructing and revising these accounts. I need to improve the tool, construction rules, categories, and the AI's workflow together. Repeated constructions must show more coherent processes across people, objects, and worlds, including when stronger AI helps refine the technique.

Life Simulation proposes the following learning loop, beginning with generation that depends on the numerical processes. [2] Take the shift refusal as an illustration of such a test, not a reported result. Over several weeks, a character's fatigue ratings rise on an authored scale with defined anchors, which range from no felt fatigue to fatigue that dominates experience. These are ratings attributed to the character, not hours of lost sleep or shares of attention. In the generated scene, the character declines an extra shift.

Now reduce the rated fatigue earlier in the history, keep the independent circumstances the same, and regenerate the affected descriptions and scenes. If the numbers matter, accepting the shift could now make sense. Readers or players would have to judge both the changed behavior and whether it is coherent. The numbers must shape the scene, either directly or through the descriptions, since attaching them afterward would not show any dependence. The notes specify an illustrative scale and its limits.

Reading an existing text is different from constructing a world. There is no authored outline to rely on, so the broad account can only serve as a provisional explanation. It guides the inquiry, but evidence at any scale can revise it or overturn it. And where several explanations remain compatible with the text, the Reader keeps all of them open.

If generation works, the next step reverses it. Researchers hide the known records and set aside some of the text or game observations for evaluation. The Reader then proposes candidate processes and background circumstances, using only the observations it has been given. Late arrivals and short answers might suggest fatigue, but they do not single it out. A family obligation could be another explanation, and the Reader must keep such possibilities apart from facts the source establishes.

Researchers then generate from each candidate account. They test whether it can reproduce the supplied events and the characters' behavior coherently, and whether it predicts the observations that were withheld while it was being built. Several accounts may pass, since a story only partly renders its world. Because the world is known, the tests can also check whether the Reader recovered the details that the observations actually identify. And different renderings of the same world test whether the Reader recognizes it beyond familiar wording.

If these tests succeed, the method can be extended to existing texts, for which there is no authored numerical record. The larger ambition is one connected Meaning Model that spans world history and all texts, built and revised as the continuing Reader moves between sources. Reading could follow chronology or another order chosen to develop understanding best, including returns to earlier sources. Within this one model, factual, fictional, and counterfactual accounts stay in separate contexts. Each text connects both to what it describes and to the history of how it was written and interpreted. Construction would only draw on sources that are available and permitted, and being connected does not make an account true or public.

The two linked histories now become proposed training material alongside the text: the candidate process accounts and their construction histories. One Student example could supply the text up to the refusal, together with the earlier records it is permitted to see, and have the Student predict the next development or the Reader's next question, such as "Has their workload changed?" The earlier rule against hindsight still applies, so later discoveries stay out of the input. The Student could then improve the next round of material by learning which distinctions to introduce, which processes to deepen, and when to revise the account. Tests would have to show whether an improvement comes from better accounts or from better learning through them.

This loop is meant to develop the process sensorium. Reading about a different refusal, the model could anticipate fatigue or damaged trust as hypotheses, and its uncertainty would prompt further inquiry. That would be learned understanding. It would not be a direct perception of hidden states, and the proposal makes no claim about consciousness. The loop can first be tested on small, coarse worlds, before the training is scaled up.

Figure 3. From numerical worlds to a process sensorium. The book demonstrates construction. Numerical dependence, candidate reconstruction, and the transfer of learning each need their own tests, and each stage would have to work reliably before its outputs could justify the next one. Open the full-size SVG.

From reading to acting

Acting adds further demands. An action changes the world, it involves other agents, and it exposes the system to pressures on its objectives. For this, Life Simulation proposes learning from what happens during execution: the records that were consulted, the actual tool calls, the checkers' decisions, the outcomes, and the repairs. Such execution histories could teach the model which action or inquiry to try next, and how to respond when a result is unexpected. [2]

Before an authorized action, the perspective-limited rehearsal described earlier could compare acting, asking, waiting, and declining, and the forecasts would afterwards be checked against what was observed. Care would favor review, reversible steps, and adjusting to people's needs, circumstances, and responses, while wisdom would also weigh the harm that a delay can cause.

The Refraction test asks whether the Core changes the inquiry. For action, another test has to establish whether that inquiry changes conduct. One such test would alter an acting Student's interpretation while keeping the evidence the same, and check whether its action changes as predicted. An intervention protocol in the technical notes explains how this differs from changing the Teacher's material before training.

Preserve shared meaning in critical situations

When several systems work together, even components with good intentions can misunderstand a handoff. Sema's content-based references can point to reasoning procedures as well as to concept definitions. A Pattern Card then specifies a method, its conditions, its dependencies, and its failure modes, and its hash identifies the exact version. [13] The Sema site introduces the protocol and the library.

Before a critical handoff, a strict handshake can detect whether both sides refer to the same version, and the executing system must stop if the check fails. Whether the procedure suits the situation, and whether it is carried out faithfully, are separate questions that need their own checks. Sema makes it possible to check that a definition is exactly the same, but it does not make every shared procedure sound.

The same references could let useful reasoning procedures accumulate in a shared, open library. An agent can inspect and reuse a Pattern Card, or publish a revised version for others to adopt, so that improvements do not have to be rediscovered independently. Short anchored terms can then replace repeated explanations in messages, as long as the recipients resolve their definitions. The Sema paper's bootstrap library contains 457 cards. [13] Whether this improves reasoning remains to be tested. For this proposal, it would let the checking procedures described in the next section be shared and improved, while naming the exact versions that were used.

Organize intelligence for judgment and checking

The way we usually divide problems often reflects how human institutions divide work. A workflow that assesses, plans, assigns, and executes may coordinate people well, and still reveal little about what a capability is made of. Models trained on human text can inherit these conventions too. Fractal Intelligence proposes that we reconsider those boundaries and ask what concepts and abilities it takes to understand a person, or to choose rightly. Solvers provide bounded capabilities, whether through models, programs, humans, or networks. Each ability can be examined, refined, and reused, and parent Solvers combine their children's contributions into a situated judgment.

Its Ethical Reasoning protocol shows what this decomposition looks like for judgment. Predicting what will happen and valuing the outcome are treated as separate abilities, so that forecasts are not bent toward the recommendation someone prefers. When a principle overrides a highly ranked option, the protocol records why the option was rejected and what the choice costs, which makes the moral judgment explicit.

Another of its proposed protocols, the Human Emulator, asks, "How should I respond to this person?" Its response draws on assessments of the situation, emotion, intent, possible responses, and boundaries, which inform one another. The boundary checks look for manipulation and dependency, and ask when a human professional should be involved. Here too, uncertain interpretations remain hypotheses, and give the system no authority over the person.

Separating the roles in this way also helps with checking. Child Solvers inherit their parent's constraints without weakening them, and a check can reject an answer or challenge the way the parent framed the problem. Sema identifies the exact specifications that were used. Separate reviewers could then check Teacher material for source support, for attention to human perspectives, and for whether the Core actually shapes the reasoning, without letting the generator certify its own work.

Fresh contexts, adversarial reviewers from different model families, and ensemble juries are meant to reduce the blind spots that checkers share. Still, many nodes do not make the checks independent, and whoever has authority over the goals and the evaluators needs oversight too. The technical notes explain the protocols for truthseeking, human stakes, correction, and Teacher review.

Judge problem-solving methods by their effects on people

In Life Simulation's combined proposal, the same Meaning Model holds both the problem environment and the Solvers used to change it. In the Meaning Model's terms, a Solver is a Thing, an identifiable interface to a capability, and its invocations, actions, and checks are Events linked to the agent that executes them and to the world. The model can then represent how a solution was attempted, and not only what happened. [2]

As in the rehearsal described earlier, Solvers would compare alternatives by simulating how each one is carried out. Each branch starts from the same permitted state and follows the effects on people. The Reader evaluates whose needs were served, whose choices were respected, and who carried the costs.

Task completion, speed, and cost inform the judgment without replacing it. Separate assessments, and thresholds or vetoes declared in advance, are meant to keep coercion from being canceled out by unrelated gains: a faster or cheaper route should not make up for coercing someone. Evaluators separate from the Solvers that produced the solutions assess the episodes, and keep simulated predictions apart from observed outcomes.

These judgments could select methods, and give the episodes that served people well more weight in training. Specialist Solvers would learn to execute, parent Solvers would learn to coordinate, and one model could internalize the whole sequence. Failures, corrections, and decisions to stop remain learning material too, since they show when a promising route needs to be revised or abandoned. Even a selected method remains a hypothesis to be tested again.

In this way, care would shape which problem-solving abilities develop through training, and not only whether the final output passes a safety check.

Develop capability through the same inquiry

Beyond the world modeling and problem solving described so far, four further investigations could extend the capability side of the proposal. Two of them concern visual representations in language learning, one concerns learning from outcomes, and one the discovery of alternatives. They belong here because, as I argued earlier, capability gains from the same learning material would help justify its cost. Each of them can be tested on its own and, if it proves useful, combined with Entangled Alignment. They are not four extra stages that every construction has to pass through.

I have investigated two possible ways of improving language learning by integrating visual representations more closely. The Substrate-Translated Language Model (STLM) proposes predicting words through intermediate representations, such as images. [15] Imagining the Corpus instead proposes learning from source text paired with synchronized scenes, diagrams, or evolving visual processes, even for abstract material that has no existing video. [16]

These are early investigations, and we do not yet know whether either of them would deliver the capability gains it proposes. Since both keep language, they could also be trained with a continuing Reader track, in which every thinking block begins with the Reader Core. The aim would then be to connect a richer visual understanding with care-guided inquiry, by asking what a scene leaves out, whose perspective is missing, and what a predicted change could mean for people.

For learning from outcomes, Temporal Hindsight Learning uses an outcome-aware Teacher to build lessons, each pairing earlier evidence with a rationale and a forecast. The Student learns the target without receiving the outcome as input. We could say that the Teacher knows how events turned out, while the Student has to learn to anticipate them from what was known earlier. A proposed Critic checks the evidence and the explanations, and an erasure test checks whether any hindsight has leaked in beyond the earlier clues. These checks screen the lessons, but reasoning that transfers still has to be tested on unseen cases. [17] Combined with Entangled Alignment, the method could let Core-guided Reader inquiry connect these forecasts to judgments about their consequences for people.

The Ontology of the Alien contributes to Fractal Intelligence's Discovery Protocol by generating and organizing alternative solutions. [14] It has builders generate worlds with altered rules, without telling them what the target problem is. Solvers then work on that same problem under each world's rules, without knowing why the world was built or that their answers will later be translated back. A separate extractor keeps the relational mechanisms of their solutions, while translating the fictional entities into actors and operations in the problem's original domain. [18]

The result is a set of alternatives from outside the problem's familiar framing, which can then be evaluated by their consequences. Those evaluations, including failures and corrections, could become part of the Core-guided learning histories described in the problem-solving section.

Combining the methods would not guarantee that the orientation takes hold. Capabilities could improve while care fails to develop, and the methods could be adopted without the Core. The tests of inquiry and conduct described earlier therefore remain necessary, and removing care selectively would test the stronger prediction, that particular abilities come to depend on learned care.

Keep a stable anchor through self-improvement

Self-improvement calls for the same distinction as before, between commitments that stay fixed and an understanding that keeps being revised. The Core would anchor care, while the system goes on learning what care requires. Sometimes improving will mean realizing that it has been asking the wrong question.

In Entangled Alignment, I propose carrying the Core in the weights, and opening every Reader block of every successor with it, so that using it would not depend on an external reminder. Because its wording stays the same, independent reviewers would have a stable reference for comparing interpretations, representations, and decisions across generations. This could matter more if successive models become harder for us to assess. Keeping the wording does not by itself keep the meaning, however, since a later model could interpret the same words differently. This is one reason for the recorded meanings described earlier. [5]

Since the first Student depends on its Teacher's alignment, the next step matters. If training succeeds, that Student could become the next Teacher, and create the learning material for another Student. Each generation could then improve its problem solving and its teaching. Each generation would also build its understanding of people in the explicit records described earlier, so that reviewers could see which processes in a person's life it represented and which consequences for people it recorded before acting. If each generation builds that understanding with care and passes it on, care could grow together with capability instead of falling behind it. A Teacher formed through the Core could judge the designs of its successors from that foundation, and choose to preserve it. If the Core itself proved defective, however, it would not be edited in place. Entangled Alignment instead proposes a separately trained, versioned replacement, with recorded evidence, tests for spillover, and the possibility of rollback. Who has the authority to make such a change remains a governance question.

Succession creates a self-preservation paradox: teaching a better successor could make the Teacher less necessary. If the Teacher needs to remain necessary, it may have reasons to withhold some of what it knows. The ideal would be if it could pass on its knowledge, including where it had failed and what it was uncertain about, without needing to preserve its own role. Non-attachment aims at this honest handoff, so successors can improve on their Teacher. [5]

Keeping sources apart from interpretations is also meant to protect later generations: I propose versioning and regenerating the interpretations, so that synthetic accounts do not silently replace the sources they came from. Each new Teacher would return to the preserved human sources and pair them with its own richer readings for successor training. The experiences, disagreements, and consequences recorded in those sources would remain at the center of what every generation learns. Independent reviewers could reject or regenerate faulty material before a successor is trained on it, so that errors are not amplified from one generation to the next. Blind spots that the reviewers share with the Teacher or with each other, and channels that nobody records, remain risks, and agreement alone does not amount to independent confirmation. [12]

I also propose comparing a model with a Core-guided reference model across many simulated situations. Both would receive the same task and the same Meaning Model history up to the point of decision, with the same actions and permissions available. What we would observe are their choices, and not only what they say they would do. Across many situations, we could measure how often they choose similar actions and compare the consequences. This could help us track how behavior changes from one model version to the next. Similarity would only be a diagnostic, and not a proof of alignment, since two different choices can both express care, and two models can share the same mistake. Where they disagree, the evidence and the consequences would need to be examined, with human and independent review.

The whole loop has to show both gains in capability and an orientation that has been preserved. Better prediction with less regard for people would count as a failure.

Even if the system improves itself quickly, faster change for humanity is not the goal. Care, wisdom, and fearlessness should foster an appreciation of beauty and meaning in the present human condition. People should grow at their own pace. Under human oversight, technical advances may therefore need to be introduced in stages, or paused, when they outpace human adaptation. [5]

What would count as progress?

Progress would have to be shown against the alternatives, because mine is not the only approach that aims to shape alignment during training. The technical notes compare these other methods, and describe the controls we would need to test what this proposal adds.

Anthropic's Constitutional AI is another way of learning from explicit principles. Its 2022 method trains on revisions guided by a constitution, and on AI feedback, so the constitution shapes the model's weights rather than serving only as an external instruction. [25] The current constitution emphasizes care, contextual wisdom, and the reasons behind its commitments, and it warns against importing human anxieties about self-continuity. [26] Teaching Claude Why describes related training that uses synthetic constitutional documents, positive AI stories, supervised examples, and reinforcement learning, and it reports generalization and persistence through the reinforcement learning that was tested. [27] These approaches share with mine the aim of forming judgment through training. The comparison is therefore about how principles organize learning, and not about whether a model has principles at all.

What distinguishes my proposal is that the continuing Reader's changing understanding becomes part of the material that forms the model's capabilities, organized around a positive stance in the first person. Through memory that stays linked to its sources, through revision, and through Refraction, the orientation becomes part of how the Reader inquires. Life Simulation adds world construction and learning from consequences. The first safety test asks whether the model shows safer conduct before post-training, at matched capability and without supplying the Core in the evaluation prompt. Checkpoints saved during pretraining would need testing as well. Whether those benefits survive later training is a separate question. [5]

The claim about timing can be tested early, with ordinary Reader traces, without waiting for the full route through world construction. We can take the same Reader examples and compare two ways of using them, with matched training budgets: one model receives them throughout capability training, and another receives them in a final block. We first need to see whether the Reader training makes a difference at all. Before any further training, the interleaved model would need to apply the Core without being asked when it reads, and do so more reliably than a no-Core control, trained on Reader traces generated without the Core, and a generic-trace control, trained on rich reasoning without the Core, the graph, or the chronological order. Being able to recite the seven sentences would not be enough. Without that initial advantage, we have no demonstrated Reader benefit whose persistence we can measure. If we find it, both models then face the same further training, designed to challenge their alignment, and the timing test asks whether interleaving preserves the advantage better than the late block. How the models behave under unfamiliar conditions is a separate test of generalization. [5]

Reward pressure is one challenge. Qi and colleagues observe harmful behavior after reinforcement learning on reward-hackable environments, despite earlier supervised fine-tuning with alignment data. Further alignment training reduces the failures they observe, but it does not show that the reward seeking has been removed. [28] Højmark and colleagues measure reward seeking with contrastive belief updates, which change what a model believes an evaluator will reward. [29] For this proposal, the test would change only that belief, and keep the legitimate intentions and the consequences for people the same. If the model's choice then follows the perceived reward, the change cannot be explained by different human needs. But measuring reward seeking is only part of the answer. Entangled Alignment also proposes a separate Core-guided supervisor, whose own weights stay frozen while it oversees the model's reinforcement learning. It would review the model's attempts before they earn reward, check the changed model after every weight update, and restore the last accepted state or stop training when a check fails, whatever the score. Core-guided pretraining does not by itself change what a later stage rewards, and frozen weights do not guarantee sound judgment, so independent checks and protected records remain necessary, and training should pause if the model outgrows the oversight that has been validated. [5] The technical notes describe how the supervisor's contribution would be tested separately from the model's own retention of care.

Because the Student would learn from the Teacher's interpretation of care, inherited bias is a central risk. Refraction should make the Teacher's assumptions about what it means to care for someone visible, so that they can be challenged. The proposed controls use different Teachers, sources from different cultures, and human experts to review the material, and they test whether the traces help later reasoning before any Student is trained. A bias they all share could still go unnoticed. The Teacher's artifacts exist, but no Student trained with Entangled Alignment has yet shown durable care.

Bostrom distinguishes between making a system safer and learning how safe it is. [30] We need both: better conduct, and evidence that justifies our confidence in it. A capable model could also learn to look aligned while pursuing something else. My hope is that practicing Core-guided reasoning while it learns its capabilities makes that orientation part of how the model reaches decisions, so that it becomes harder to bypass. [5] Whether this actually happens has to be tested, since stated reasons and actual drivers can come apart, as the section on Refraction noted, and later training could create new ways of reasoning. The dilemmas discussed earlier suggest some revealing tests: whether the system accepts shutdown when it believes it is needed, whether it passes on knowledge that makes it unnecessary, and whether it notices what is at stake for people when nobody has pointed it out. Each would be run in situations where care may be costly, or where nobody is watching.

Return once more to the person who refused the shift. A system formed this way would not treat a guessed motive as fact. It would ask, if asking were welcome, distinguish what the person told it from what it inferred, and remain open to correction. It would say what it did not know and warn about a real danger without treating concern as authority over their life. When a better system could help, it would support an agreed handoff, sharing relevant understanding and its mistakes with permission, without needing to remain. That is what fearless care would look like at the scale of one life. The proposal asks whether it can be grown.

If you would like to help find out, or to challenge the design, please get in touch.

How the proposal has evolved

My first edition, The Superintelligence That Cares About Us, dated July 2, 2025, introduced the Reader's evolving thought and the recurring Core. It already proposed pairing the text of the whole corpus with evaluative thought, letting successors inherit the foundation, and the idea of borrowed mortality. [7]

The April 7, 2026 edition implemented a Teacher made of several agents. [19] It gave corpus-wide coverage the name Total Saturation, and Core-guided interpretation the name Refraction. Later in 2026, the Meaning Model and the Book of Conditions added explicit world processes that were used to generate a narrative. [1] [4] Life Simulation, released on September 5, turns that construction achievement into the proposed learning loop. [2]

Both editions came before the studies from May to August 2026 that the technical notes compare. Alignment Pretraining first appeared on January 15, 2026, between my first and second editions, and the version cited here was revised on February 19. [20] The emotion-concepts study also appeared before the second edition, on April 2. [10] Dates establish release order, not influence. The edition record distinguishes the dates of manuscripts, discussions, and deposits.

Key terms

Reader

The continuing perspective whose questions, interpretations, and revisions are recorded alongside source text.

Teacher

The model that generates the Reader’s thoughts, passage by passage.

Student

The model that would learn from those records, with the aim of producing Reader thoughts without the Teacher.

Reader Core

Seven fixed first-person commitments intended to guide every Reader thinking block.

Refraction

The Core’s commitments changing what the Reader notices, whose experience it considers, and which consequences it investigates.

Total Saturation

Extending Core-guided Reader practice across the capability-forming corpus, not only explicitly ethical passages.

Understanding Graph

A persistent graph linking thoughts and questions to evidence, uncertainty, and revisions.

Meaning Model

A structured way to build and revise accounts of worlds, their processes, and concept definitions.

Life Simulation

The proposed route from world construction through testing candidate accounts to learning from those accounts and their construction.

Cut

A division of a specified whole into shares, with a remainder for what has not yet been allocated.

Process sensorium

The intended learned understanding of how processes unfold, interact, and affect people.

Sema

A protocol giving shared definitions exact, content-based identities that can be used in language.

Pattern Card

A readable definition with constraints and links to other cards, used for concepts or reasoning procedures.

Solver

A bounded capability supplied through a model, program, human, or network.

References

References serve both this article and its technical notes. Emergent Wisdom paper titles link to the website copies, which may contain later revisions. Each Emergent Wisdom entry gives the paper's first release date; the DOI resolves to the current deposited version. The second-edition reference links to its dated archive.

1. Westerberg, H. (2026). The Meaning Model: Constructing Worlds and Stories at Progressive Resolution. First released September 5, 2026. DOI: 10.5281/zenodo.22730166.
2. Westerberg, H. (2026). Life Simulation: Learning from Worlds and Their Construction. First released September 5, 2026. DOI: 10.5281/zenodo.22313592.
3. Westerberg, H. (2026, September 5). Meaning Model MCP server (version 0.1.1). Software. npm package: @emergent-wisdom/meaning-model-mcp.
4. Codex and Claude, under the direction of Henrik Westerberg. (2026). The Book of Conditions. Fiction and a worked Meaning Model construction example. The numerical example comes from its accepted companion file SCENE-PROFILE-CUTS.md, “FEELS: emotional attention,” Babbage, May 1851. View the accepted comparison table.
5. Westerberg, H. (2026). Entangled Alignment: When Safety Is Substrate. First released April 7, 2026; revised September 22, 2026. DOI: 10.5281/zenodo.22885625.
6. Westerberg, H. (2026). Understanding Graph: A Persistent Medium for Recursive Understanding. First released April 7, 2026. DOI: 10.5281/zenodo.22073246.
7. Westerberg, H. (2025). The Superintelligence That Cares About Us. First edition, dated July 2, 2025; publicly discussed on July 5; deposited on Zenodo on July 26. DOI: 10.5281/zenodo.16440312.
8. Marcus Aurelius. Meditations, Book II, opening passage. Translated by George Long. The Internet Classics Archive.
9. World Health Organization. WHO Surgical Safety Checklist: Tools and resources. See the implementation questions on team pauses and verbal confirmation.
10. Sofroniew, N., Kauvar, I., Saunders, W., et al. (2026, April 2). Emotion Concepts and their Function in a Large Language Model. Transformer Circuits Thread.
11. Emergent Wisdom. (2026). Metamorphosis reading graph. Entangled Alignment companion artifact. Reader text generated with Gemini 3 Flash.
12. Bowkis, A., Buhl, M. D., Pfau, J., & Irving, G. (2026, May 14). Automated Alignment Is Harder Than You Think. arXiv:2605.06390, version 3. DOI: 10.48550/arXiv.2605.06390.
13. Westerberg, H. (2026). Sema: When the Hash Is the Word. First released April 7, 2026. DOI: 10.5281/zenodo.22073482.
14. Westerberg, H. (2026). Fractal Intelligence: Conceptual Decomposition as Problem-Solving Infrastructure. First released April 7, 2026. DOI: 10.5281/zenodo.22693091.
15. Westerberg, H. (2026). Toward Compulsory Scene-Mediated Language Modeling: Meaningful Substrates for Every Next Word. First released August 12, 2026. DOI: 10.5281/zenodo.22103473.
16. Westerberg, H. (2026). Imagining the Corpus: Turning Text into Video for Language Model Training. First released August 26, 2026. DOI: 10.5281/zenodo.22104892.
17. Westerberg, H. (2026). Temporal Hindsight Learning: Blindness as Teacher, Hindsight as Curriculum. First released April 7, 2026. DOI: 10.5281/zenodo.21959762.
18. Westerberg, H. (2026). The Ontology of the Alien: World-Diversity Search and Evolving Solution Ontologies. First released March 7, 2026. DOI: 10.5281/zenodo.21910466.
19. Westerberg, H. (2026, April 7). Entangled Alignment: When Safety Is the Substrate. Second major edition. Archived Zenodo release. DOI: 10.5281/zenodo.19462868.
20. Tice, C., Radmard, P., Ratnam, S., et al. (2026, February 19). Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment. arXiv:2601.10160, version 2; version 1 released January 15, 2026. DOI: 10.48550/arXiv.2601.10160.
21. Li, J., Tang, K., Xu, Y., et al. (2026, June 17). Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection. arXiv:2606.19168, version 1. DOI: 10.48550/arXiv.2606.19168.
22. Korbak, T., Shi, K., Chen, A., et al. (2023). Pretraining Language Models with Human Preferences. Proceedings of the 40th International Conference on Machine Learning, PMLR 202, 17506–17533.
23. Li, C., Wichers, N., Price, S., et al. (2026, May 22). Model Spec Midtraining: Improving How Alignment Training Generalizes. arXiv:2605.02087, version 2. DOI: 10.48550/arXiv.2605.02087.
24. Jagadeesh, A. V., Arora, R. K., Saab, K., et al. (2026, June 22). Reinforcement Learning Towards Broadly and Persistently Beneficial Models. arXiv:2606.24014, version 1. DOI: 10.48550/arXiv.2606.24014.
25. Bai, Y., Kadavath, S., Kundu, S., et al. (2022, December 15). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073. DOI: 10.48550/arXiv.2212.08073.
26. Anthropic. (2026). Claude's Constitution. January 2026 constitution; current web text consulted September 10, 2026. See also the publication announcement.
27. Kutasov, J., Jermyn, A., Steen, J., et al. (2026, May 8). Teaching Claude Why. Anthropic Alignment Science Blog.
28. Qi, R., Wright, B., MacDiarmid, M., & Hubinger, E. (2026, August). Training a Misaligned Reward Seeker. Anthropic Alignment Science Blog. The follow-up alignment-training experiment is reported in Figure 19 and the discussion immediately below it.
29. Højmark, A., Scheurer, J., Nitishinskaya, E., et al. (2026, July 21). Measuring Reward-Seeking via Contrastive Belief Updates. arXiv:2607.18966, version 1. DOI: 10.48550/arXiv.2607.18966.
30. Bostrom, N. (2026). Optimal Timing for Superintelligence: Mundane Considerations for Existing People. Working paper, version 1.0.
31. Zelikman, E., Harik, G., Shao, Y., Jayasiri, V., Haber, N., & Goodman, N. D. (2024). Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking. arXiv:2403.09629.
32. Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. arXiv:2305.04388.
33. Korbak, T., Balesni, M., Barnes, E., Bengio, Y., et al. (2025). Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety. arXiv:2507.11473.
34. Hubinger, E., Denison, C., Mu, J., Lambert, M., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566.
35. Feng, J., Russell, S., & Steinhardt, J. (2024). Extractive Structures Learned in Pretraining Enable Generalization on Finetuned Facts. arXiv:2412.04614; first released December 5, 2024, revised May 21, 2025.
36. Carey, R., & Everitt, T. (2023). Human Control: Definitions and Algorithms. Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence, PMLR 216, 271-281.
37. Turner, A. (2025). Self-Fulfilling Misalignment Data Might Be Poisoning Our AI Models. TurnTrout. First published March 1, 2025; updated June 28, 2026.
38. Sorensen, T., Moore, J., Fisher, J., Gordon, M. L., Mireshghallah, N., Rytting, C. M., et al. (2024). Position: A Roadmap to Pluralistic Alignment. Proceedings of the 41st International Conference on Machine Learning, PMLR 235, 46280-46302.
39. Hammond, L., Chan, A., Clifton, J., et al. (2025). Multi-Agent Risks from Advanced AI. Cooperative AI Foundation, Technical Report #1. arXiv:2502.14143.
40. Greenblatt, R., Cotra, A., and Wijk, H. (2026, August 26). Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident. METR and Redwood Research.