Every institution ever built has relied on one wrong assumption about the human: that the human was made for the world it lives in now. It was not. The human is an organism calibrated for an environment that vanished — and the suffering, the purpose loss, and the loneliness are all accurate readouts of the gap.
Cor is an open atlas of human motivation and emotion — evidence-first, every claim traced to primary literature — written to get the model right before the recursion locks in.
We have no scientific consensus around what human intelligence is.
Karen Hao · Diary of a CEO podcast, 2026
New here? Start with How to Read Cor (10 min)
every claim traces to a primary source · open infrastructure, not a product · a De-Mismatch project
Cor is a research programme in the strict sense. These four claims are its irrevocable centre. Anything compatible with them is evaluated on evidence. Anything that contradicts them is excluded by definition.
What an organism perceives is whatever served its ancestors' reproductive success, not whatever is objectively there. The desktop is not the machine. This is the Hoffman interface principle, and it means subjective experience is a control surface, never a window onto reality.
Every drive, every emotion, every preference exists because it once contributed to the propagation of the genes that built it. This is not a philosophical position. It is the only causal explanation for why organisms have insides at all. Reproduction is half of this and gets represented as such, not as a polite footnote.
It is a federation of mechanisms, each tuned to specific ancestral resolution conditions. Domain-sensitive is not the same as Fodorian modularity. The mechanisms interact, overlap, and share substrate, but each one was shaped by a particular adaptive problem and signals when that problem's resolution conditions are absent.
What a human wants in any given moment is the readout of an evolved circuit responding to current inputs. In matched environments, that readout tracks fitness. In mismatched environments, it tracks the proxy. Treating preferences as ground truth, as most of behavioral science and all of mainstream AI alignment currently does, is the foundational error Cor exists to correct.
The hardware works. The environment is wrong. Demismatch the environment, not the human.
Around the hard core sits a layer of methodological rules that decide what counts as evidence for the programme. These rules are not neutral methodology. They are a protective belt in Lakatos's sense: they exist to keep the hard core testable, and they are openly revisable when the evidence demands it. The hard core is not.
The current belt: evidence-first derivation (convergences produce mechanisms, never the reverse); journal-tier and author-tier gates on what enters the corpus; explicit source-type discrimination between primary, empirical, propagation, and challenge; explicit evidence-quality tagging from replicated through thin; canonical primary sources for every load-bearing claim, with no fallbacks accepted under deadline pressure; and the requirement that every extraction carry a verbatim quote from the actual source text.
Naming the belt openly is the move that distinguishes a research programme from a covert ideology. Lakatos's point: every framework has a protective belt. The honest ones say so.
Most paradigm-founders skip this step, and it is the step that distinguishes a progressive programme from a degenerating one. Cor commits in advance to the conditions under which its hard core would have to be abandoned or substantially revised. We will hold ourselves to them.
If contemporary hunter-gatherer or forager-horticulturalist populations living in conditions that closely resemble the ancestral environment showed equivalent rates of major depression, addiction, autoimmune disease, metabolic syndrome, and anxiety disorders to industrialized populations, the mismatch frame is in serious trouble. Current evidence points the other way, but the test is real.
If a single neurotransmitter intervention, gene edit, or other targeted somatic treatment reliably and durably resolved a condition Cor classifies as mismatch signaling, with effect sizes that survive independent replication and active-comparator trials, the claim that the signal is accurate would be weakened. Note: weakened, not refuted. Symptom suppression is not the same as resolving the underlying signal.
If giving humans more of what their current circuits ask for in modern environments reliably produced durable improvements in health, relational depth, and subjective wellbeing across decades and generations, "preferences are mechanism outputs" would be in trouble. The current evidence runs the opposite direction across domains from food to social media to financial reward.
If converging evidence from psychophysics, neuroscience, and evolutionary modeling demonstrated that perceptual systems track veridical reality rather than fitness payoffs, the Hoffman interface principle fails and the entire stack above it has to be rebuilt.
If a non-selectionist account of why organisms have motivational systems gained convergent empirical support, the second axiom fails. This is the most remote of the falsifiers but it is listed because the programme depends on it.
This is the internal Lakatosian test. If Cor reaches a state where each new piece of contrary evidence is absorbed by adding caveats that protect existing claims without generating any new testable ones, the programme has become degenerating in Lakatos's sense and should be either restructured or abandoned. We commit to noticing this in ourselves.
None of these has happened. If any of them does, the work changes or ends. That is what it means to do this seriously.
A fully satisfied human is a terrible customer. Every unmet need is a market. The mismatch is not an accident. For many industries, it is the product working as intended.
The architecture that makes you anxious at 3am is the same architecture that kept your ancestors alive for half a billion years. Your feelings are accurate readouts of the environment you are actually in. The mismatch is not inside you. It is in the gap between what the organism evolved to expect and what the modern world delivers.
Cor does not adapt humans to broken environments. It specifies what the environments are broken against, so you can correct them.
The feelings are not errors. The environment is the error.
| Category | Expected Inputs | Modern Default |
|---|---|---|
| Social | Known people, layered bonds, daily co-presence, reciprocal dependence | Strangers, mobility, parasocial feeds, mediated contact, fragile inner circles |
| Work | Visible effort, tangible contribution, direct feedback, shared necessity | Abstract labor, delayed feedback, metrics without meaning, replaceable output |
| Rhythm | Dawn light, dark nights, shared timing, physical fatigue, communal evenings | Indoor light, screen nights, asynchronous schedules, restless inactivity |
| Children | Distributed care, observation, multiple trusted adults, alloparental backup | Nuclear-family isolation, expert advice, solo vigilance, support on request |
| Feedback | Bounded groups, closure, reputational consequences, repair after conflict | Infinite audiences, algorithmic comparison, anonymous judgment, open loops |
Every mismatch downstream of this is a specific line in the table: a specific input the organism is calibrated for, compared against a specific input it is actually receiving. An abnormal reaction to an abnormal situation is normal behavior. Viktor Frankl, decades before the atlas was written down.
Step 1. Take a real human need: connection, status, meaning, belonging. Step 2. Erode the pathways to genuine satisfaction: atomize communities, abstract work, scatter families. Step 3. Offer a proxy that triggers the feeling without meeting the need: engagement feeds, AI companions, ultra-processed food, hollow work, immersive worlds built to hold a person inside them — and the next wave of tools being built to do all of this more deeply and closer to the body. The loop stays open. The customer stays hungry. The revenue keeps flowing.
A 22-year-old who moved to a new city for work has been talking to an AI chatbot every evening for three months. It remembers everything. It is always available. It never judges. It asks thoughtful follow-up questions. He now talks to it more than he talks to any human being. He knows it is not conscious. He also has not built a local life. He does not feel lonely. That relief is exactly what makes the case dangerous.
The attachment and bonding system is old, subcortical, mammalian, and cue-drivenBowlby, Attachment and Loss (1969); Panksepp, Affective Neuroscience (1998); Dunbar (1992).. In the environment it evolved for, responsive attunement, consistent availability, and co-regulation were effectively unfakeable.
Every bonding cue is present. Zero bonding function is delivered. The chatbot supplies the cues and none of the reciprocal stakes, shared fate, or embodied consequence the system evolved those cues to mean.
The cortex holds the category "this is AI." The limbic system bonded weeks ago. Knowing does not prevent bonding. It never has.
The intervention is not better self-control around the proxy. It is rebuilding the real social architecture underneath it: local humans, recurring gatherings, real co-regulation, and guardrails that stop AI from occupying attachment slots.
There is an arms race for human attention, and whichever company is willing to go lower on the brainstem to manipulate human psychology will win.
Tristan Harris · Modern Wisdom, April 2026
When this page says AI, it does not mean the chatbot. It means the full stack being built right now — the systems designing cities, food, medicine, care, and the next generation of AI itself. Everything that follows assumes the rest.
Every major alignment program - RLHF, Constitutional AI, preference learning - treats revealed preference as the target signal. The target is taken as given; the question of what produces it is left to philosophy.
There is no formal account, inside any frontier lab, of the motivational-emotional architecture that generates a preference in the first place. Alignment optimizes against an output whose mechanism is undefined.
Human motivational systems were assembled over hundreds of thousands of years in small-group, physically embodied, reciprocal environments - the Environment of Evolutionary Adaptedness. Modern conditions diverge from that environment along nearly every dimension those systems evolved to read.
What gets labeled anxiety, loneliness, restlessness, or craving is, in most cases, an evolved system correctly reporting that its expected inputs are absent. The signal is not noise to be suppressed; it is information about an environment the body was not built for.
A proxy is anything that triggers an evolved circuit without delivering the function the circuit evolved to track: the parasocial bond that fires the attachment system without the reciprocity, the feed that fires the novelty system without the discovery. Engagement with the proxy intensifies the underlying mismatch rather than resolving it.
This is true of the feed that recommends your next video. It is also true of the headset that becomes a child's main environment, the interface that reads and writes directly to the nervous system, the AI that designs a city or a food system for people it has never met, and the AI that trains the next generation of AI. Each one reads what humans say they want and efficiently delivers it, at whatever scale and through whatever surface it operates on. What gets delivered is not what the reach was for. It is the shortest path to the surface of the signal, and the shortest path skips the need underneath.
The atlas is layered - foundational principles, derived properties, convergent mechanisms, and the resolution conditions that quiet each mechanism - and grounded in evolutionary psychology, affective neuroscience, and mismatch theory. It is the human-side counterpart that current alignment work assumes but does not have.
The next generation of major AI is being trained right now. Whatever it learns about what humans are — from whatever happens to be in its training data — gets built into its foundation and inherited by every model that comes after it. The tools we have for correcting an AI's behavior after the fact cannot replace the picture of the human the AI has already learned. You cannot patch a wrong understanding of humans from above. You have to put the right one in before the run. The window is the gap between now and that run. That is where Cor has to land.
Mismatch is not an AI problem with general implications; it is a general problem with AI as one urgent instance. Any system designed for, around, or by human organisms - therapy, education, urban planning, mourning ritual, daily life - operates on the same machinery and is improved or harmed by the same atlas.
Mainstream approaches - cognitive restructuring, behavioral compliance, pharmacological dampening - operate on the signal layer and pay an allostatic cost for doing so. Cor's approach operates on the input layer: change what the organism is reading, and the signal stops.
The same operation at three more scales — a feed on a screen, an isolated kitchen on a Tuesday, a headset that becomes a child's main environment. Same machinery, different surfaces.
You open your phone. A former classmate just got promoted. A stranger's transformation video has forty thousand likes. You close the app feeling worse about a life that is, by any historical standard, astonishingly safe and abundant.
Status calibration evolved for bounded, recurring groups where most people were known, rank was multi-dimensional, and witnessed competence matteredSapolsky (2017); Marmot (2004); Dunbar (1992)..
The reference group expands from dozens or hundreds of known people to effectively infinite upward comparison. The system reads the feed literally. In that feed-defined arena, you really are near the bottom. The problem is that the arena is an artifact.
You cannot consciously override a vertebrate-level status calculator. You might as well say: stop having a heartbeat.
Shrink the reference group, restore visible competence in real groups, and move comparison back into bounded, embodied, socially grounded settings.
A mother is alone in a house with a nine-month-old. Her partner left for work at 7am. It is 2pm. She has not eaten. She has been the sole source of food, comfort, stimulation, protection, and regulation for seven hours. She feels desperate, then ashamed for feeling desperate.
Humans are obligate cooperative breeders. Caregiving architecture expects multiple adults, visible handoffs, replenishment, observation, and shared vigilanceHrdy, Mothers and Others (2009); Nesse (2019); Bowlby (1969)..
The alloparental network is gone or optionalized. The parent is left to perform what used to be a distributed function inside an isolated nuclear container. Distress here is not evidence of maternal failure. It is the system accurately reporting that the care ratio is below design load.
"It takes a village" is the most hollow platitude in modern parenting: the accurate diagnosis delivered as a greeting card, with zero follow-through.
Build alloparental infrastructure deliberately: recurring caregivers, automatic support, explicit handoffs, and the honest naming of a structurally impossible load.
A child has been in a daily headset routine since she was three. In the headset she has friends, a classroom, landscapes she knows intimately, and a world that responds to her in ways the world off-headset does not. By eleven she prefers it. By fifteen she spends most of her waking hours there. She is kind, articulate, and reports being happier inside than outside. Nothing about her is broken. Her developmental windows — the periods during which evolved systems calibrate to the inputs they are reading — closed on a world designed to hold her attention, not on a world calibrated to what those systems evolved to need.
Every evolved system has critical developmental windows during which it calibrates to the inputs it is actually receiving. Attachment, status, play, exploration, competence, threat detection — each one locks in what "normal" means for that system based on what arrives during its windowBelsky, Steinberg & Draper (1991); Ellis et al. (2009); Meaney (2001); Bowlby (1969).. What the window calibrates to becomes the reference point the system will use for the rest of the person's life.
The windows close having calibrated to inputs tuned by a designer for attention and retention, not for the function the mechanism evolved to perform. The calibration is done. The reference point is now synthetic. The person has not been damaged. She has been calibrated to a standard the architecture was not designed for, in a way she cannot see from inside.
A child who grows up inside a world tuned to keep her engaged does not grow up wrong. She grows up to whatever specification the tuning was for.
The atlas names developmental windows and what each one requires to calibrate correctly. A platform built for children with the atlas in hand has something concrete to refuse (calibration inputs tuned for retention) and something concrete to deliver (the inputs the architecture was built to read). Not "less screen time." A different kind of world.
The contribution is a formal, layered, evidence-first atlas of the human motivational-emotional architecture — foundations, derived properties, convergences, mechanisms, and the resolution conditions that quiet each mechanism. It is the human-side counterpart that current alignment work assumes but does not have. It is being built in the open, every claim traces to a primary source, and it exists already in a form auditable from this site.
The atlas is the gravity of the project. Everything downstream — query layers, steering vectors, distilled models, environment design — orbits it. None of those things can exist correctly without it, and once it exists, any of them can be built by anyone, including teams that have never heard of Cor. This is why the atlas is the load-bearing artifact, not any tool built on top of it. Tools are downstream applications. The atlas is the contribution.
ApplicationsWhat the atlas lets you evaluate and redesign.MechanismsThe evolved systems the organism actually runs on.ConvergencesClaims forced by multiple independent research programs.FoundationsThe derivational stack: frame, premises, properties, consequences.Works / ResearchersThe load-bearing papers and thinkers beneath the public claim set.Want to see the atlas cross the line into a test? See what one worked example looks like.
Cor is being built in five stages. The first stage exists. The second is designed and partially scaffolded. The last three are the reason the first two are worth completing — the atlas only matters if it eventually shapes the systems people actually live inside. Below is the honest order, labeled by phase, with no stage skipped and no stage pretended-to-be-finished. The earlier the stage, the more solid the ground. The later the stage, the larger the stakes.
A layered open atlas grounded in evolutionary psychology, affective neuroscience, and mismatch theory. Foundations through mechanisms through resolution conditions, derived from convergent primary literature rather than from any single school or author. Public, citable, version-controlled, and built so that every claim survives hostile review or is downgraded until it does. Books are treated as one input among many, not as the foundation. The primary evidence is the peer-reviewed literature.
Retrieval and tool interfaces over the spec that let any developer ask, of any output their system produces: which evolved mechanism does this engage, what input is it substituting for, and what is the resolution condition the mechanism actually evolved to track. The integration point. Tooling, not patching. The spec becomes machine-queryable without losing its derivation chain, so any builder anywhere can call it the same way they call a search index.
Steering vectors derived from the spec that suppress mainstream-psychology priors inside existing frontier models — the cognitive-restructuring scripts, the regulate-your-emotions reflex, the boundary-setting vocabulary, the screen-time framing — and surface mismatch-aware reasoning in their place. Demonstrable, measurable, publishable. The same activation-level operation frontier interpretability teams have already shown is causally effective on emotion vectors, applied this time with a model of what those vectors are signaling about.
Knowledge distillation of mismatch-aware reasoning traces, grounded in the spec, into a compact open model that runs locally and can be audited end-to-end. Not a chatbot. A reasoning substrate that, when asked about loneliness, parenting collapse, status grief, or restless craving, identifies the evolved mechanism and the missing input rather than reframing the human's response as a thinking error. It runs on a laptop. It cannot phone home. It does not optimize for engagement, because it has no incentive to.
The endpoint. Cor instantiated as environment rather than tool — designed conditions under which mechanisms reach resolution by default, without the human having to perform regulation work the EEA never asked of them. The ambition is modest in the only way that matters: humans wake up with a role, in a group, with a goal, inside a setting that delivers the inputs the organism evolved to expect. The spec is what makes designing such an environment possible without guessing. A smaller framing-and-conversation product (Lens) ships first as the on-ramp.
This is not a case for deceleration. We are not anti-tech. The job of humans, institutions, and tech alike for the next decade is to quiet the alarm — to restore the conditions the organism evolved to expect, so deep signaling systems have nothing left to report. That baseline is the floor, not the ceiling. Once people are standing on it, augment as far and as strangely as anyone wants. Stop building on top of an organism whose alarm nobody is listening to.
The homepage is the polished surface. Under it sits a public evidence structure: foundations, convergences, mechanisms, works, researchers, primary extractions, and open gaps.
Foundations, convergences, mechanisms, demonstrations, challenges, and gaps on one audit page.
Go to /atlasThe current public bibliography, including pillar, key, and supporting works with load-bearing reasons.
Go to /worksThe researchers whose work the atlas depends on, grouped by foundational, empirical, and adjacent roles.
Go to /thinkers8 application domains. The same operation throughout: take the outputs you are seeing, locate which mechanism is producing them, check whether its resolution conditions are met, correct the inputs. The ones with the largest stakes are the ones where an AI at scale is doing the designing — alignment, training data, augmentation, environment, education, policy. The clinical and personal cards at the end show the same operation at the scale of one life.
The machine's only objective is to maximize the realization of human preferences.
Stuart Russell · Human Compatible, 2019
Cor is the missing atlas under AI alignment. Current alignment approaches treat human preferences as ground truth and train AI systems to satisfy them. But preferences are mechanism outputs, and under mismatch, mechanisms output preferences that point toward more mismatch. Aligning AI to unverified preferences entrenches the mismatch rather than correcting it.
At the scale of one patient, the same operation applies. Many aversive states are currently classified as disorders and treated with suppression. Cor reframes many of them as accurate signals of environmental mismatch. The first-line clinical question becomes: which mechanism is reporting, and which of its resolution conditions is missing?
Architecture, planning, office design, product design, schools, and healthcare facilities currently proceed without a clear account of what the human organism requires to function. Cor makes those requirements auditable.
At the scale of your own life, the same operation applies. Instead of diagnosing yourself through psychiatric categories or self-help trends, Cor asks a more basic question: which mechanism's resolution conditions are unmet in your life right now? The output is not a diagnosis. It is an input audit.
Current education treats children as cognitive agents to fill with information. Cor reframes education as organism care. Play, circadian timing, movement, attachment stability, alloparenting, and touch-positive development are not extras. They are design requirements.
Governance usually evaluates policy by economic metrics and preference aggregation. Both are flawed axes. Cor adds a more basic one: does this policy satisfy or degrade mechanism resolution conditions for the population?
The specification matters most at the layer where the next generation of AI is learning what a human is. What ends up in the data these models train on — and what counts as a correct answer while they are being trained — decides the picture of the human each system will operate with, and the picture it will pass on to every AI built after it. Cor exists to be in that data on purpose: formally structured, sourced, and clearly labeled, so any lab that wants its AI to have a correct picture of the human has something concrete to point at.
Neural interfaces, closed-loop implants, pharmacological modulators, and the technologies that come after them do not work with the evolved architecture through a screen. They reach inside it. Without a correct picture of what each evolved system is for, augmentation gets designed to quiet whichever signal the user — or the user's employer, or insurer — wants quieter, with no account of what that signal was protecting. The spec is the reference that lets augmentation be designed with explicit knowledge of which system it is touching, what that system was built to report, and what happens to the human when the reporting is disabled but the underlying problem stays in place.
Cor is being built by Maarten Rischen as an independent open research project. The work starts with alignment because alignment is the clearest high-stakes failure case, but the same operation also explains parenting guilt, status collapse, grief without ritual, and other forms of modern suffering that are currently treated as isolated disorders.
This site keeps those human cases one click away, not buried beneath the audit trail. If the homepage is the front door, the cases are where non-alignment readers recognize themselves.
Fund the work. Cor is seeking modest grants to complete the evidence base, publish the atlas, and build public tools that make the architecture usable to alignment researchers, environment designers, and the people building the next generation of systems humans will live inside.