At rest, a system reveals nothing of its own agency. Whether an artificial agent pursues a goal, whether it is competent, who authored its output, none of that is on its surface, and the surface is built to be read as more than it is: it answers as though a single mind stood behind its tokens, and what stands there is a composite, a model under a persona, with a memory, a harness, permissions, and tools, unified only where a user meets it. This paper is about how to read such a thing. A system that holds its state and a system that defends it present the same face at rest, because a perfectly regulated variable and a merely undisturbed one leave the same flat record, and what separates them, a model held inside the regulator, is a fact about what each would do under disturbance. We call the study of this situation perturbatics, and its principle is a separation principle: when two organizational models agree over the regime a system has been in, no analysis of that record tells them apart, and separation requires a probe, a controlled difference that makes the models predict differently, from a deliberate intervention or a natural experiment. A minimal instrument follows. Agency is scored as a Bayes factor between a goal model and a passive model; a gridworld holds three systems that trace the same path and so score at chance when watched, an area under the ROC of 0.50 that is the identity of their paths and not a failed classifier, and the probe battery separates every pair, though no single probe in it does alone, the planner from the script at 0.95. The instrument attributes the evidence across the assemblage by the do-Shapley value of each component, in two maps, one over the evidence for agency and one over the realized capacity. Their difference is a legibility term, and it makes agency theater a number: a narrating persona shows agency it does not realize, where the goal register realizes what it shows. The reading depends on the declared boundary, reaching its full value only when the goal and world model are enclosed with the planner. Richness is a false friend throughout, the most complex walker emitting negative evidence of agency while the simplest of them scores among the most agentic. The account has a boundary of its own, and it is the important part. For agency you can move the goal a hundred times and read the score. Consciousness is different: its candidate probe, turning the system off, returns no reading the rival hypotheses score apart, must not be run to find out, and presupposes a theory of the individual it would end, so the science must also say when not to perturb. There the instrument stops. What is left is no longer a number. It is a relation, and the grief, or its absence, of whoever turned the machine down.

An ordinary dynamical system moves a state under a rule that stands still. Many of the systems that carry the weight of biology and society do the other thing: they edit the rule as they run. A population reshapes the environment that selects it, a colony lays trails that redraw who can reach whom, a market invents instruments that change what a trade is. This paper names that process nomopoiesis, from nomos (established law or custom) and poiesis (making), and gives it a formal object: a system carries an ordinary state and a law parameter that its own trajectory edits, so that one equation is change under a law and the other is change of the law. It counts as genuinely nomopoietic only when a causal path runs from the state now, through the law, to the state later, which excludes systems whose rule is turned by an outside clock. The obvious objection, that the enlarged state-plus-law is just a bigger ordinary system, is granted in form and answered by a typed distinction between object-level and rule-level variables; the strong cases, where the reachable dimension, admissible variables, interaction topology, or action alphabet change, resist the flattening in content. A depth of parametric, structural, or generative orders the cases. Niche construction is the worked instance: life rarely rewrites physics (the beaver does not amend fluid mechanics) but routinely writes a local law inside it, since the dam resets boundaries, flow, transport cost, and the selection the beaver's descendants will meet, so niche construction is ecological nomopoiesis. A minimal deterministic simulation makes four consequences concrete and checkable. The derivative of long-run fitness splits into a direct term and a niche-mediated term whose signs can oppose, so a trait deleterious at fixed environment becomes favoured once the environment it builds is counted, the analytic split matching a finite-difference of the constructed equilibrium. Ecological inheritance is an impulse-response kernel whose persistence sets a threshold, a costly constructor here favoured only above an environmental persistence of 0.68. A nomopoietic gain separates a state-edited law, where a one-off action compounds to 4.5 times its initial trajectory effect, from an externally forced twin, where it decays to nothing. And a replicator game whose payoff matrix is written by play sustains a cooperation-defection oscillation of amplitude 0.95 that every frozen game forbids. The claim is that evolution is not only motion across a landscape but the endogenous editing of the landscape's rules, and that this is a testable structural property rather than a metaphor.

Imagination crosses arbitrary distances in a fraction of a second: picture the desk in front of you, then a galaxy ten billion light-years away, and the second image arrives about as fast as the first while staying inside your skull. Dividing the represented distance by that moment gives a number far above the speed of light, and it describes nothing that moved. This paper makes the artifact precise and replaces it with quantities that can be measured. A velocity is defined only once a changing state, a clock, and a metric on the states are fixed, and the phrase "speed of imagination" runs at least five spaces together: the cortical sheet, the neural population state, the represented content, the represented external world, and represented time. The maps between them do not preserve distance, so scaling the imagined universe by a factor multiplies the apparent velocity while the neural trajectory is untouched, which makes that velocity a property of the labelling rather than the process. What is real is a finite-time transition in local neural dynamics with distance-independent representational reach, and it has no single speed but a kinematic profile: cortical wave propagation, population state-space velocity, target-representation latency, replay rate, temporal compression, and fidelity, measured in the literature from tens of milliseconds to a second and from centimetres per second to metres per second, none convertible into the others. Building on evidence that prefrontal traveling waves correspond to rotational population trajectories, the framework takes wave geometry to act as a control field that gates which low-dimensional subspace can become active, so the latency of an imaginative transition is set by the control cost of the neural path rather than the represented distance. A minimal control model makes the claims concrete and reproducible: latency is invariant to represented-space rescaling across twenty orders of magnitude while apparent velocity scales exactly with the label; latency is predicted by control energy (R-squared 0.87) and not by represented distance (0.02); a jump is flat in the number of intermediate states while a traversal is affine in it; and forcing a transition faster than the control budget affords costs a fidelity set by the efficiency of the neural path. A proposed experiment varies represented physical, semantic, visual, and graph distance orthogonally, contrasts jump with traverse, and reads out decoding onset, wave topology, and state-space trajectory, yielding falsifiable predictions including that perception and imagery run cortical waves in opposite directions. Relativity holds throughout, because a representational jump sends nothing to the place it represents and takes nothing back.

Complex systems routinely command families of useful behaviour whose explicit design, learning, or evolutionary cost looks far too small for their scope: one switch composes into all of Boolean logic, logarithmic representation hands over an entire leading-digit distribution, a compliant body walks with no controller, local cells rebuild an anatomy no genome stores. Michael Levin calls these free lunches from a Platonic space of patterns. We keep the phenomena and drop the word free. Each is a heavily discounted lunch: the world supplies a pre-existing regularity, the system pays to build an interface to it, and a family of consequences arrives at low marginal cost, in no violation of thermodynamics or the No-Free-Lunch theorems. We make the intuition an accounting one, with a vector-valued discount across cost currencies, an accounting boundary that records who paid and when, a taxonomy that reduces to seven operations (entailment, equivalence, compression, delegation, amortization, externalization, protection), four case studies each required to exhibit a regime where its interface becomes a penalty, and an inference ladder separating a regularity from access to it and cost reduction from any claim about a new ontology. Operationally strict and metaphysically open, the framework replaces the question of whether a lunch is free with the question of which structure supplied the discount and where the rest of the cost was paid.

Rankings of private wealth are read as rankings of contribution to civilization, and the reading is wrong in a way that has structure. Work becomes foundational by becoming non-excludable, standardized, cheap to copy, and depended upon by everything above it, and each of those properties is what stops its creator from charging for it. The people whose work became the substrate of modern technology, the designers of the internet's transport protocol, the metal-oxide-semiconductor transistor, the small public-domain database inside most devices, the salt-and-sugar formula that rehydrates a dying child, captured almost none of the value their work carries, and they captured almost none of it because the work became foundational. This paper models a technological civilization as a stack of serially dependent layers, each able to levy a per-use toll, and computes what tolls do to it. In a normalized market the open, untolled stack delivers social value 0.5 and its creators keep nothing; a single owner of the whole stack keeps 0.25 while halving throughput; and when the stack is split among independent tollbooths the tolls compound, so that two booths set a higher combined price than one owner would and collect less between them while users are worse off, and twenty booths cut throughput to about 5 percent of the open level. The sustainable private capture of any one foundational layer falls as 1/(2N) in the number of complementary layers, so the model's capturable fraction reaches the roughly 2.2 percent of social surplus that Nordhaus estimated innovators actually capture near a stack 23 layers deep. The configuration that maximizes social value is the one in which creators capture nothing. Great fortunes form instead at the narrower, more excludable layers above the commons, where control over a point that use must pass through can be turned into recurring revenue, and ownership of such a point is then narrated as though it were the building of the road beneath it. The paper is careful about what this does and does not license: appropriability drives real invention, integration and scaling are real work, and a single integrated owner does less damage than a fragmented anticommons. The claim is the systematic divergence between contribution and capture, and the mistaking of the tollbooth for the road.

A simulation of water does not wet, and a language model trained on text does not, by that training, perform the speech acts whose records it reproduces. This paper makes the reason precise and turns it into a theorem with a worked example. Full language is an effectful process: a speech act in the sense of Austin and Searle maps a context and a social, epistemic, and normative state to an updated state together with an utterance, so the complete object is a state transformer and the utterance is only its emitted value. Text records the utterance and discards the state change. The textual shadow of a linguistic process is defined here as the observational projection that keeps only textual outcomes, restricts to textual probes, and then quotients states that no textual probe can separate, a coarse-graining that collapses many distinct pragmatic states onto one textual state. A model fit to text learns the minimal predictive machine of that shadow, the predictive states of the text stream in the sense of computational mechanics, and nothing finer. The consequence is an underdetermination result: distinct effectful processes can induce identical textual shadows, so no amount of text-distribution fitting can recover which process produced the text, and reproducing strings is therefore not performing the acts. An accompanying deterministic computation exhibits the gap. Two speech-act processes are built over a shared token alphabet, one that binds every promise it utters and one that binds none, engineered so that their textual output distributions are identical to machine precision while their effects on world-state diverge. The shared shadow is a two-state predictive machine of statistical complexity about 0.985 bits and entropy rate about 0.920 bits per symbol; the two processes assign, after 100 tokens, an expected 42.86 outstanding obligations against 0, a whole interval of world-states consistent with one and the same text; a third process makes each promise binding on a hidden fair coin, leaving 1 bit of illocutionary information per promise that no text-only learner can recover. The strongest opposing view, that the shadow preserves far more than critics grant because inferential and pragmatic structure leaves textual traces, is granted in full, and the residue it does not reach is located exactly: the effect on the world that the utterance was the act of producing. The boundary is a claim about what text-only prediction can recover, not a verdict on understanding. What closes the gap is not more text but grounding, interaction, and a handler in the world that gives an utterance its force.

Genius is usually located in a person. This paper locates most of its cultural work in the environment a learner inherits, the exemplars, tools, and possibility-spaces that exceptional achievements leave behind, and treats education as the process that converts that distributed inheritance into individual capability. Cumulative-culture theory explains how a civilization keeps its gains across generations through a ratchet, but not how a kept gain becomes an individual's usable and extendable skill, and that conversion is what a theory of education has to describe. The paper's operational core is a distinction among three frontiers: the production frontier, the best output a learner can put on the page with the tools at hand; the competence frontier, what the learner can independently generate and revise; and the judgment frontier, what the learner can reliably evaluate and select. Generative AI expands production almost immediately while leaving competence and judgment near their starting values, and the resulting gap is false ratcheting, sophisticated output unbacked by durable capability, revealed the moment the tool is withdrawn. The distinction is measurable, which is what separates the framework from a metaphor: test the learner with the tool present and again with it absent, and read the difference. An accompanying illustrative model makes the logic explicit. Under one learning dynamics, a finished-output tool and a structured tool that fades both reach the same production frontier near 0.87, yet competence ends at 0.15 in the first case and 0.84 in the second; hand every learner the tool at a common strength and then remove it, and displayed output collapses by 83 percent for the finished-output tool against 10 percent for the fading one; the genuine-ratchet fraction is 0.10 against 0.99, and the ordering survives 2,000 seeded parameter perturbations. The model is illustrative, its parameters stipulated to instantiate the definitions rather than measured from learners. The framework then states its predictions as falsifiable claims: AI ratchets capability genuinely when it decomposes an achievement, elicits prediction and explanation before revealing answers, generates controlled contrasts, preserves productive difficulty, and withdraws its own assistance, and it counterfeits capability when it mainly supplies finished products. A randomized trial of a scaffolded physics tutor and a field experiment with an unrestricted chatbot already show the predicted split, and the paper specifies the measurements that would confirm or break the account.

Newton's constant is measured, not derived, but that statement hides a sharply structured problem. The unit-dependent SI decimal of G is the wrong target; the invariant one is the gravitational coupling of the proton, alpha_G = G m_p^2 / (hbar c), equal to about 5.9 x 10^-39, equivalently a Planck-to-proton mass ratio of about 1.3 x 10^19. Massless spin-2 consistency, through the soft-graviton theorem and the self-coupling reconstruction of the Einstein interaction, fixes the form of gravity and the universality of its coupling while leaving the overall normalization free, because the soft theorem is homogeneous under rescaling that normalization. A limited infrared underdetermination proposition makes the gap exact: a one-parameter family of theories sharing identical non-gravitational physics assigns different values to alpha_G, so no matter observable determines it, and new information must enter from a ultraviolet completion, a global regularity condition, an anomaly, a topological quantization, or a unique vacuum. Quantum chromodynamics then explains why the proton is exponentially light rather than why gravity is weak: one-loop running turns a moderate coupling into an exponentially small confinement scale through dimensional transmutation, with no input of order 1e-19. An accompanying deterministic computation runs the strong coupling from its measured low-energy value up to the Planck scale and down to the Landau pole, finding that transmutation carries the Planck scale across about 19.9 base-10 orders to the QCD scale, and reproduces the illustrative unification-scale decomposition in which a proton-to-unification factor near 2.2 x 10^-33 and a Planck-to-unification factor near 610 combine to alpha_G. That decomposition is scheme- and threshold-dependent and is presented as illustrative bookkeeping, not as a derivation. Reducing every framework to one scale shows the unsolved object is a single dimensionless normalization, C_G/C_p near 2.6 x 10^18, that the common scale cancels out of; induced gravity, asymptotic safety, extra dimensions, string vacua, holography, species counting, and anthropic selection each rename this one number and retain a free modulus, trajectory coefficient, vacuum, counterterm, rank, or measure. No accepted realistic construction reaches a non-gravitationally calibrated prediction of alpha_G. The weakness of gravity is not thirty-eight orders of accident; it is one match, between the gravitational response and the matter-generated scale, still unmade.

A driven deterministic system can be read as carrying a memory of its own history: fibre the state space over an observable self, quotient each fibre by future behaviour, and let the input words that return the observable to its value act on the quotient as a monoid of closed self-histories. A recent proposal builds nine invariants on this picture and reads them as autobiographical structure, a system's representation of its past selves. This paper shows the construction is classical. The autobiographical fibre is the output-fibre of the Nerode minimal realization, and the self-history monoid acting on it is the fibre-restricted transition monoid, which is the Krohn-Rhodes holonomy tile; the nine invariants reduce term by term to automata, semigroup, and Koopman quantities. The proposed autobiographical curvature is ill-posed, since connection holonomy is invertible while the self-history monoid contains the erasures the construction is built to represent, and the fibres change size from one observable value to the next. The self-versus-environment defect measures observability of hidden inputs rather than aboutness: a system that copies its last input scores as maximally self-remembering, and the number depends on the observable chosen by hand. One question survives the reductions, whether a canonical self can be discovered rather than declared. The autobiographically self-predictable observables are closed under common refinement, proved here, and across 150,964 finite systems they never once fail it; under coarsening they fail in 54,300 of the 65,536 systems with four states and two inputs (82.9 percent), rising to 99.8 percent at six states, with a four-state minimal counterexample exhibited in full. The self-predictable observables form a closure system with no canonical coarsest member, a structure that coincides with informational closure. A system remembers its history; it does not, on this evidence, thereby have a self.

Reinforcement learning scores a policy by a single number, the expected return, and treats everything an agent does as a means to that scalar. Competent action resists the scalar: when a hand turns a cup to read its shape, what improves is not a running total but the agent's grip on the world, the way its uncertainty about hidden states collapses into task-relevant structure. This paper gives that grip a formal object. The agent's belief over world-states is a sufficient statistic for control, it lives on the probability simplex, and Chentsov's theorem forces the Fisher-Rao metric on that simplex, so the movement of belief has a geometry that is compelled rather than chosen. Competent contact is read off that geometry: the contraction of the uncertainty set, the Fisher-Rao length of the belief trajectory, and the separability the terminal belief keeps between world-states. The central result is a quotient. If reward depends on the world only through a partition, the expected return depends on the belief only through its pushforward onto classes, so two beliefs at the maximal Fisher-Rao distance can be reward-identical, and a scalar objective is blind to everything the geometry resolves within a reward-class. A small active-perception POMDP makes this exact. A reward-only policy and a competent-contact policy attain the identical single-task return, yet leave beliefs of very different shape, and world-states that differ only in a reward-irrelevant attribute collapse to Fisher-Rao distance zero under the reward-only policy while the competent policy holds them apart. The discarded geometry is exactly what transfers: on a second task the reward never mentioned, the reward-only agent scores at chance and the competent agent well above it, a gap no single-task return could have priced. A from-scratch curvature routine recovers the simplex curvature to machine precision, certifying the geometry the argument runs on.

Sacrifice, potlatch, delayed gratification, costly signaling, and the death of a cell for its tissue are usually kept in separate disciplines, and the temptation is to declare them the same thing. They are not the same substance. This paper argues that they share one formal structure and states it precisely enough to be wrong. A subsystem accepts a genuine local loss; that loss becomes higher-scale control if and only if three ingredients are present, each at a distinct point of a single feedback loop: observation, so that the act registers and enters others' models; memory, so that the registration persists and accumulates rather than resetting; and constraint-change, so that the accumulated capital feeds back and relaxes the subsystem's own survival constraint. A minimal, deterministic multi-scale viability model makes the condition exact. An agent running a standing deficit can expend part of its private stock into a coalition; the loop returns a gain of theta times kappa over one minus mu units of higher-scale support per unit expended, and this loop gain must exceed unity for the loss to be repaid at the margin and exceed 1.5, the headroom-set viability threshold, for any survival policy to exist. Under the full loop the gain is 2.8, the survival-optimal act flips from hoarding to a sacrifice of 0.2778 per step, and the viability kernel fills 0.9834 of the state space. Ablating any single ingredient sends the loop gain below the threshold and empties the kernel to zero; the hoarding policy dies from every state even under the full loop. Because the three ingredients enter at three different structural points, each is independently necessary, and that mutual necessity, rather than any analogy between burning blankets and dying cells, is what the disparate phenomena share. The same construction separates sacrifice from waste as a matter of sign and shows why the loss is hard to fake: only an agent above a capacity threshold of 3.35 can bear the transient the sacrifice requires.

The oldest puzzle about heaps asks for a number: at how many grains does sand become a heap? This paper argues the number was never the point, and that the heap belongs to a wide family of systems in which a latent state accumulates small changes while a coarse readout classifying it stays fixed, until at some point the readout flips. Write the state update as x_{t+1} = F(x_t, u_t) and the readout as y_t = R(x_t). The sorites condition is that for many steps R(x_{t+1}) = R(x_t) although x_{t+1} is not x_t, and the paradox is the confusion of readout-invariance with state-invariance, of a tolerance relation that is not transitive with an equivalence relation that is. The paper develops this as one architecture across five domains and grounds four of its claims in a reproducible simulation. A population of leaky integrators driven by a stream of small inputs holds the accumulated count in full, so that a linear decoder recovers it exactly, at an R-squared of 1.0 that a deterministic linear state guarantees by construction rather than reports as a finding, while the thresholded readout stays silent for the first 37 steps and carries no information at all about how much has arrived, which is the exact sense in which a subthreshold input is nothing only to the spike that has not fired. The indistinguishability relation, in which adjacent states differ by less than a just-noticeable amount, holds for every neighboring pair yet its transitive closure merges a thousand plainly distinct states into a single class, and three quarters of its chained triples already fail transitivity, so the paradox is a relation error, not a mystery about grains. A soft threshold makes local tolerance and global transition coexist without contradiction: the largest change in category probability produced by a single grain is 0.0062 while the total change across the range is 1.0. And a bistable latent state driven slowly across a fold reproduces the emergence of a new category, a spike, a heap, a multicellular individual, with the coarse readout giving no warning at all until the jump, while the variance of the latent state rises roughly 15-fold beforehand, its autocorrelation climbs from 0.9 toward 0.99, and a reservoir trained only on pre-transition data raises an alarm 198 steps before the readout moves. The synthesis is deliberately bounded. It does not solve the sorites by naming a grain number, it does not claim that neurons are heaps or that cells are reservoirs, and it takes the sheaf-theoretic reading of contextuality only as far as the honest slogan that locally adequate readouts need not glue into one global predicate. What it offers is a single formal spine, latent accumulation under coarse-grained readout, on which vagueness, psychophysics, synaptic integration, reservoir dynamics, and the major transitions of individuality can be read as one design and not five metaphors. Ships a runnable simulation whose output carries every modelled number.

Humans detect agency without instruments. A person reads pursuit in a tiger, intention in a baby, and effort in a struggling animal, and reads it fast, involuntarily, and often wrongly, in a storm or a shadow or a moving triangle on a screen. The proposal here takes that human competence as a prototype the way the eye was a prototype for the lens: the optics of agency-detection are already implemented in us, and the question is what the instrument is doing and whether it can be extracted, corrected, and enlarged into an agentoscope that finds agency where unaided perception cannot, most promisingly not outward in weather but downward in living tissue. This paper argues that the instrument is a model comparison. A system is agentic to the degree that its behavior is better predicted by a model with a goal, a boundary, and a policy that corrects deviations than by a passive model of forces and noise, and the agency score is the log-ratio of the two. The paper's central and somewhat unwelcome result is that this score is unreadable from observation alone. When a system's goal is not moving, the goal-model's predictions coincide with those of a suitable passive dynamical model, so the two cannot be told apart by watching, and a simulated detector separates three hundred goal-seekers from three hundred passive attractors at an area under the ROC of 0.49, which is chance, with the agency score never once exceeding zero. Agency is not a property on the observational rung of Pearl's ladder. It appears only under intervention. When the goal is moved and the system tracks it, or the path is blocked and the system reroutes, the goal-model predicts what the passive model cannot, the score climbs by roughly 101 nats per informative intervention, and the same detector now separates the two populations perfectly. Not every perturbation informs: merely displacing a system, which a passive attractor and a goal-seeker both recover from, leaves the score at zero, so the discriminating interventions are exactly those that pit the goal against the mechanism. Two further results guard the instrument against the mistakes the human prototype makes. Dynamical richness is a false friend: across a battery of systems the agency score is uncorrelated with complexity and in fact runs the other way, a chaotic flow scoring the lowest of all despite the highest complexity, which is why a hurricane is a false positive and a liver is not. And agency is graded rather than binary, ordered by the instrument into a ladder on which passive systems score below zero and goal-directed ones above, with equifinality under obstruction, reaching the same goal by a different path, separating a mere set-point regulator that fails a barrier from a regenerator that routes around it. The agentoscope does not detect a soul. It reports where, at what scale, and under which interventions a goal is the better explanation, and its honest use points it at organoids, wounds, and regenerating tissue rather than at the sun. Ships a runnable simulation whose output carries every modelled number.

A familiar speculation holds that mathematics is the deepest layer of things and that the physical world is somehow made of it, that the fields were already there, in some sense, before the first moment of this universe. Stated that way the idea is either empty or magical: empty if it only means the world obeys equations, magical if it means a number turned into a force. This paper proposes a third reading. A world is not a substance that mathematics turns into; it is a phase that a region of mathematics enters, the way a gas enters a liquid phase, when enough structure has been broken and stabilized that the region becomes locally dynamical, causally separated into parts, record-forming, and legible to systems inside it. Physical reality, on this account, is not mathematics plus an extra ingredient called actuality. It is mathematics that has become operationally closed and internally readable. The paper replaces the warehouse picture of mathematics with an ecology of structures, translations, invariants, and obstructions, using as evidence the long-range translations that current research actually finds, the relative-Langlands duality that reads automorphic periods as an arithmetic electric-magnetic duality and the generic-absoluteness results that ask which truths survive a change of set-theoretic universe. It imposes a typing discipline on the word becomes, separating interpretation from duality from representation from physical realization from emergence, so that a profound analogy is not silently promoted into an identity. And it grounds three of its central claims in finite-dimensional computation. A thermal state's own modular Hamiltonian equals the system Hamiltonian, the algebraic identity K = βH confirmed to one part in ten to the fifteenth rather than a surprising numerical precision, so a state already carries a notion of time without any clock added from outside. The entanglement Hamiltonian of a spatial region is exactly nearest-neighbor and, near the boundary, its couplings climb by very nearly pi per site with a linear fit of R-squared 1.0, the lattice signature of the Bisognano-Wichmann theorem that a region's modular flow is a geometric boost, while deep inside it relaxes to the physical Hamiltonian, so a state together with a choice of region already carries geometry. And an instant of actual history leaves nine independent Hamiltonians, for a four-dimensional system, that generate the identical motion, so a law is a counterfactual neighborhood, not a compressed record of what happened. Two things resist the whole construction and are marked as residues rather than hidden: a measure over observer-locations, without which a plenum predicts nothing, and a metatheory, since no sufficiently rich structure carries its own complete truth predicate. The paper distinguishes its technical sense of worlding from the new-materialist sense it knowingly borrows, and closes on the one question no phase transition can answer, whether worldhood is what physical actuality consists in or whether actuality is something more that no structural fact can supply. Ships a runnable simulation whose output carries every modelled number.

A democracy and a dictatorship are not racing on the same clock. A new prime minister inherits a damaged system, needs years for a reform to mature, and is often removed before any of it shows, while an entrenched ruler holds position through fear and accumulates across decades. The intuitive fix, a strong leader who stays long enough to finish the job, is the wrong one, because it builds the vehicle the next illiberal government will drive. This paper prices the right one. It treats a democracy's strategic capacity, the expertise and institutional memory and half-built reforms that let a state act on a long horizon, as a stock that accumulates continuously and is knocked down at each change of government, a piecewise-deterministic Markov process with a closed-form long-run mean, K* = u / (delta + lambda·chi). The disadvantage the model isolates is not the turnover rate lambda but the product lambda·chi, turnover multiplied by the fraction chi of capacity lost per turnover, a quantity the paper calls the reset tax. Four results follow from the one process. At fixed democratic turnover, cutting reset severity from chi = 0.8 to chi = 0.1 lifts a democracy from 0.22 to 0.73 of an autocracy's long-run capacity, closing about 90 percent of the excess gap and removing most of its volatility, with no change to how often leaders are replaced. The same cut more than triples the reform horizon a government can finance, from 5.5 to 18.5 years, so a 15-year mission that would need a 42.5-fold payoff to survive a high-reset democracy needs only a 3.1-fold payoff under a low-reset one. Because institutions build slowly through coordinated entry and break fast through decentralized exit, electoral alternation is not self-correcting: when destruction outpaces repair, even perfect alternation drifts to breakdown, a model echo of the finding that four in five democracies that begin to autocratize do not recover, and the two cures, faster repair and earlier resistance, are both ways of lowering the reset tax. And to match an autocrat's continuity by entrenchment a democracy would have to keep its leaders for 32 years, longer than the dictator it fears, whereas the reset route keeps 4-year leaders and still buys 73 percent of autocratic continuity at 5 times the removability. The construct prices an incentive structure; it measures no country, it predicts no collapse, and it holds throughout that leaders must stay genuinely removable. Continuity belongs in the architecture of the state, not in the body of a ruler. Ships a runnable simulation whose output carries every modelled number.

We apply faultization (a systematic regime of morphogenetic perturbation) to a minimal transformer (4-layer, 16-dimensional, 4-head character-level GPT), asking which patterns from the latent space (Levin, 2026) the system accesses, how robustly, and where the interface fails. Twelve experiments show that the optimization attractor is accessible through radically different interfaces (local layerwise loss, partial communication, chimeric assembly, regeneration from random reset), and that absent interface components are tolerated while corrupted ones are catastrophic: freezing layers costs +0.2% (p = 0.41) whereas negating their gradients costs +26.3% (p < 0.0001, d = +0.531). Silence is tolerated; sabotage is catastrophic. We also report a null we initially mistook for a discovery: a gradual noise ramp appears to build stress tolerance that sudden exposure does not, but it injects only about one-third the cumulative gradient-noise variance, and the dose-matched comparison shows no inoculation effect at either scale. Degradation tracks total noise dose, not exposure history, so we do not claim a stress-inoculation free lunch.

We apply faultization (a systematic regime of morphogenetic perturbation) to the pigeonhole principle reinterpreted as a distributed multi-agent system, asking which patterns from the latent space (Levin, 2026) the system accesses under irreducible constraint. The theoretical minimum overload O_min = m − n is a mathematical truth the system channels without computing. In six of eight experiments every run converges to O_min exactly with zero variance, the strongest evidence we have for pattern-driven dynamics. Perceptual noise produces immediate degradation with no tolerance threshold; misleading holes invert the pattern-seeking mechanism; all four local policies reach identical endpoints despite 5-fold differences in process cost.

This paper proposes epistemic lensing as a formal concept for describing how mediated information channels deform belief formation. Using a Markov-blanket framing, we distinguish direct environmental evidence from processed informational inputs and define distortion as divergence between the posterior supported by a higher-fidelity channel and the posterior induced by a mediated one. We introduce five distortion operators (attenuation, selection, warping, amplification, recursion) and four families of metrics: information loss, posterior divergence, inferential curvature, and hysteresis. A toy simulation shows how channel structure produces systematic overreaction, underreaction, salience warping, and persistent miscalibration in both individual agents and populations. We extend the framework to populations and introduce posterior shear, focal capture, shadow zones, and narrative gravity wells as collective-level phenomena. The framework reads as a bridge between active inference, information theory, media ecology, and political cognition.

The modern Person, the autonomous, rational, rights-bearing individual, is treated in contemporary political and moral thought as a natural category. The argument here treats it as a stratification: a composite assembled over millennia from at least five distinct historical formations. The paper traces those formations in sequence: relational personhood rooted in kinship, ritual, and animism; epic singularization in the narratives of Gilgamesh and Odysseus; juridical personhood as developed in Roman law; interior moral personhood through the innovations of Paul, Augustine, and Boethius; and the autonomous rational subject consolidated by Locke and Kant. The modern individual emerges only when these layers collapse together and are retroactively treated as a single thing. The paper then follows the socialization of this figure through Hegel's theory of recognition and its historicization through Foucault's genealogy of the subject. The final sections examine the contemporary forces thinning the practical conditions of agency: platform capitalism, algorithmic management, the expansion of legal personhood to non-human entities, neurotechnology. Formal personhood remains intact while its institutional supports thin. The central claim is that personhood, having been assembled, can be disassembled, and that this disassembly proceeds today through the hollowing of institutional supports rather than the erasure of the legal category. Three small computational models accompany the argument: a stratigraphy that dates and orders the nine formations, a composite index in which holding formal support fixed while draining practical support reproduces the unequal thinning of agency, and a recognition network in which concentrating who designs the field collapses reciprocated recognition for the many.

The number 1.00115965218 is the experimentally measured value of |g|/2 for the electron, where g is the gyromagnetic ratio. We read this number as historical sediment. Its leading 1 is Dirac's 1928 relativistic prediction; its next several digits are Schwinger's 1948 one-loop QED correction alpha/(2pi), which agrees with the modern measurement to one part in 660. The digits beyond Schwinger's contribution were determined order by order through the QED perturbation series, with the analytic two-loop result of Sommerfield and Petermann (1957), the three-loop analytic result of Laporta and Remiddi (1996), and the five-loop numerical calculation of Aoyama, Hayakawa, Kinoshita and Nio (2012) over 12,672 Feynman diagrams. The final digits of resolution were extracted by single-electron Penning-trap experiments, from Dehmelt's geonium spectroscopy through Hanneke, Fogwell and Gabrielse (2008) to Fan, Myers, Sukra and Gabrielse (2023), the current record at 0.13 parts per trillion. The residual uncertainty in the Standard Model comparison is now set by an unresolved 5.5-sigma tension between the 2018 Berkeley caesium-recoil value of the fine-structure constant alpha-inverse = 137.035999046(27) and the 2020 Paris rubidium-recoil value 137.035999206(11); inverting the Fan et al. (2023) measurement under the Standard Model gives alpha-inverse = 137.035999166(15), consistent with Paris at 2.2 sigma and with Berkeley at 3.9 sigma. The two recoil measurements predict a_e to differ by roughly 1.4e-12, an order of magnitude larger than the experimental uncertainty 1.3e-13. The simulation reconstructs the QED series at orders one through five, propagates the hadronic and weak contributions, performs the alpha-inversion of the measurement, and quantifies the Berkeley/Paris/Fan triangle of alpha determinations. The number 1.00115965218 now functions as a metrological instrument that audits the constants and theories that go into predicting it.

Price is a late institutional interface, a derived form of commensuration that arrives after value has already been made consequential. The deeper sequence runs through constraint, patterned flow, cost, allocation, bind, valuation, commensuration, price, money, market, capital, and financial abstraction. Physical systems supply gradients and dissipation; economic language enters there only by analogy. Cells allocate energy, metabolites, membrane gradients, and repair capacity under viability constraints; they have costs and tradeoffs but no prices. Organisms budget energy, attention, immune response, movement, and reproduction. Sensorimotor systems transact with environments before they exchange anything: a ball in flight and a catcher form a temporary field of reciprocal constraint in which action possibilities compete for realization. Human economies stabilize these older structures under social value frames. Gifts create debt without money. Blood compensation creates price without market. States fiscalize grain, land, labor, debt, and bodies before capitalism. Capitalism inherits value, money, markets, and price, subordinates them to accumulation, and treats capitalized price as the privileged public form of value. Financialization extends that move by pricing claims on future income, risk, collateral, volatility, liquidity, and state rescue. The phrase "after economy" can survive only as a precise claim: after the monetary-price form loses its monopoly over the interpretation of value. AI, platforms, ecological accounting, open-source production, public provisioning, carbon budgets, and post-scarcity imaginaries leave allocation intact and shift it into compute, attention, energy, access, reputation, and governance. The problem after economy is the political form taken by allocation after price ceases to appear as ontology. A small allocation model illustrates the structural claim as a stylized geometry rather than a measurement: by construction, undistorted prices reproduce the same allocation an undistorted objective already determines, while a speculative markup on finance raises its budget share from 0.02 to 0.26 and forfeits 8.08 percent of the model's viability. The figures give the direction of a mechanism, not a reading of any real economy.

The boundary around *Homo sapiens* is late in the history of human capacities. Fossils from Jebel Irhoud place early members of the clade, or a population close to it, at roughly 315,000 years ago; Omo I gives a minimum age of 233,000 years for eastern Africa. By then, bipedal locomotion, stone-flake production, long-range dispersal, large-animal processing, Acheulean biface manufacture, fire use, wooden weapon production, and extended social learning already had deep histories. Bipedality and stone working reach beyond secure *Homo*. Range expansion, biface manufacture, fire practices, and organic technologies were carried by *Homo erectus*, Neanderthal-lineage populations, Denisovans, *Homo naledi*, *Homo floresiensis*, and poorly resolved regional forms. The sapiens boundary is a taxonomic stabilization inside an older hominin field. The record has the structure of a stratigraphy: upright bodies before large brains, manipulative hands before secure species assignment, tools before modern anatomy, population networks before global replacement. The surviving species inherited this field and narrowed it by outlasting, absorbing, or replacing the other lineages. Read as a stratigraphy of dated capacities, seven of eight precede the Jebel Irhoud boundary and all eight precede Omo I, the deepest by more than 4 million years.

Across the long history of art the locus of the work has migrated, in successive episodes, from the finished object toward the rule-bound specification. The intuition that an artwork is a singular artifact is a late and historically specific configuration; older regimes treated images as ritual instruments, dynastic technologies, devotional interfaces, and workshop products. Conceptual and score-based art relocated the work into the certificate, the score, the instruction. Generative code art added editioned variability through deterministic seeds. A third, still emerging migration extends the same logic to model-mediated work, where the artwork resides in an executable package of prompts, reference assets, model recommendations, seeds, and licenses, navigated by a user and realized by a model. We name this form latent score art and are explicit about its evidential status: the first two migrations rest on a deep inventory of conserved works, while the third is grounded in a small number of real, externally verifiable instances (the trained-model and protocol works of Holly Herndon and Mat Dryhurst, and the autonomous system Botto) in which the generative system rather than any single output is the exhibited work, alongside in-development and illustrative cases. We present it as an emerging category and an extrapolation of the protocol logic rather than a settled phenomenon, formalize a protocol as a substrate-relative, bounded, transferable, editioned, governed specification, and trace the consequences for legal authorship, market structure, and aesthetic ontology.

Civilizations assemble from local rule systems whose interfaces with one another either glue into a coherent macro-order or fail to. This paper formalises the gluing condition. A post-collapse landscape is modelled as a finite cell complex whose vertices are settlements and whose higher cells are the routes, watersheds, debts, and jurisdictions that bind them. To each region we assign a finite space of admissible institutional packages drawn from seven dimensions: property, labor, money, contract enforcement, violence, resource governance, and knowledge retention. A global section of the resulting cellular sheaf is a coherent macro-order; its existence is read from the integer count of global sections, while the first cohomology counts the independent gluing obstructions, a finer invariant that need not vanish even when global sections exist. Above the compatibility layer sits a dynamical layer whose attractors are the historically recurrent macro-regimes: kin-commons, warlord tribute, neo-feudal land control, monastic redistribution, guild-merchant order, state-command, capitalist market order, and commons federation. A worked three-settlement river system shows that institutional repairs aimed at edges that are not the obstruction shrink the cochain spaces without buying compatibility, the kind of error reformers make routinely. A four-dimensional parameter sweep over violence, transport, energy, and literacy finds that the capitalist attractor wins 3.1% of the parameter cube, while neo-feudal and state-command together win 43.3%. These shares are facts about the model's geometry, not estimates of historical frequency. Capitalism's stability is a geometric coincidence of low violence, secure transport, abundant energy, and preserved literacy. Reconstruction, on this view, is the engineering of interfaces under a parameter regime that may or may not permit the macro-order being aimed at.

A causal influence is owned when a persistent organisation produces the constraints that generate it, those constraints steer the organisation's future attractor regime, and the resulting regime feeds back into the production of further constraints. The right primitive for agency must be definable without "I" or "self" on the right-hand side, on pain of circularity. The paper assembles a single mathematical object that absorbs ideas from chemical reaction-network theory (Feinberg), chemical organisation theory (Dittrich and Speroni di Fenizio), autocatalytic-set and RAF theory (Kauffman, Hordijk and Steel), closure to efficient causation (Rosen, Hofmeyr), closure of constraints (Moreno, Mossio, Montévil), autopoiesis (Maturana and Varela), semantic closure (Pattee, Hoffmeyer, Kolchinsky and Wolpert), Markov-blanket inference (Friston, Kirchhoff et al.), basin dynamics (Lorenz), and circular causality across levels (Haken). Owned causality is a product of four factors: closure of constraints excludes thermostats; conditional directed information from endogenous perturbations to future macrostate, conditioned on the current state and the external drive, excludes systems mostly pushed by environment; viability gain over a scrambled-constraint counterfactual run under common random numbers excludes random internal explosions; mutual information from achieved macrostate to future constraints excludes one-shot influences. The simulation computes the formula on a six-rung agency ladder running a common bistable dynamics. The two rungs with externally specified rules (rock and thermostat) score zero, with the constraint-free rock scoring a viability gain of exactly zero under the common-random-numbers counterfactual; the four with internal constraint production (flame, RAF, chemoton, organism) score between 1.0e-6 and 5.6e-6. A second simulation traces the handover from an external scaffold to the organism's own constraint deployment: directed information from the scaffold falls from a peak of 0.037 to near zero while directed information from endogenous constraints rises to a peak of 0.403, the traces crossing at window 6 of 20, with viability between 0.885 and 1.000. The pathological case is the contrapositive of the same equation. The "I" appears later as a compressed model of recurrent high-owned-causality history.

The original P-JEPA proposal claimed to extend Joint Embedding Predictive Architectures (JEPA) by replacing the homogeneous target embedding with a sheaf-valued predictive state over a stratified interaction space. That framing was decorative: the implementation computed a posterior-weighted variance, not a coboundary, and the sheaf structure was never wired into a training loop. This revision discards the replacement framing and adopts a plug-in framing instead. JEPA is treated as the working substrate; each piece of embodied or causal mathematics from the original paper is implemented as an auxiliary loss, head, or sampler that can be added to a stock JEPA training loop and ablated. The paper has two kinds of content with different epistemic status. The validated content is a battery of preregistered null and negative results about a bespoke toy and the original paper's own unpublished code. The conjectured content is a *typology*: a proposed mapping from auxiliary losses to the structural assumptions they match, plus a priority order for V-JEPA-scale implementation. The typology is not a validated finding; it is a research agenda derived from inductive-bias reasoning, untested at scale, and in tension with the one place the toy can check it (H5 ranks intervention loss first yet finds it neutral on the toy). The toy evidence (5 preregistered hypothesis tests on dishworld, 8 generated JSON artifacts) shows that (i) the obstruction gate in the original paper is a no-op on every reported suite, with the "p_jepa_stack" agent numerically identical to a plain exact value-of-information agent; (ii) value-aware active probing beats entropy probing with paired bootstrap CI [+0.009, +0.017] across 50 deterministic seeds; (iii) the trained intervention encoder is matched within CI [+0.000, +0.005] by a frozen random projection of equal width; (iv) a real cellular sheaf reduces coboundary energy 10× as theory predicts but produces *slightly worse* downstream action choice than the raw cover (CI [−0.005, −0.004]); (v) on a small NumPy JEPA with the auxiliary losses ported as toggleable terms, the viability head shows a positive trend (CI [−0.007, +0.10]) while the bisimulation regularizer at the chosen weight is mis-calibrated and hurts (CI [−0.23, −0.03]). The toy is at its variance limit and these results are *directional* signals for V-JEPA-scale ablation, not quantitative rankings. The paper closes with a conjectured priority order for V-JEPA implementation: intervention loss first, composition consistency and active masking next, sheaf consistency on overlapping clips conditional on the H4 boundary, bisimulation with curriculum tuning, viability head last (or first if the downstream is safety-critical). This order is a hypothesis from inductive-bias reasoning, not a measured ranking, and is partly contradicted by the toy. The null and negative results are gated by `docs/HYPOTHESIS_RESULTS.md`; the priority order is offered as a conjecture.

The word "expert" in mixture-of-experts is a fossil: a modern expert is a routed feed-forward subnetwork for conditional compute, not something that runs experiments. Taking the older sense seriously, we ask what it would mean for a language model to run an experiment on itself. Recent test-time activation steering does not qualify, because steering pushes the residual stream toward a chosen answer and so presupposes the conclusion. A genuine experiment is answer-agnostic: it intervenes on the evidence a claim depends on, predicts how the output should respond if the claim is grounded, and reads the discrepancy. We make this concrete as a controlled internal experiment, a targeted intervention on the evidence bound to an answer paired with a matched control intervention on evidence that should be irrelevant, whose control-corrected differential measures whether the answer is causally grounded. A mixture of experimenters is a battery of such experiments aggregated into one label-free correctness signal. On an in-context retrieval task with a small transformer whose ground truth is known by construction, in a noisy regime, the aggregated signal separates correct from incorrect answers with AUROC 0.967 (a three-experiment mixture; 0.956 for two), against 0.786 for negative entropy, 0.784 for logit margin, and 0.602 for a supervised correctness probe. The comparison with the probe is one of access, not supervision: each experiment runs three to five counterfactual forward passes and watches the output move, while the probe is a single-pass readout of one frozen activation, so the gap shows that interventional access beats static readout rather than that labels are useless. The construction transfers to Pythia-160M, reaching AUROC 0.841 against 0.719 for entropy. When the model must locate its own evidence by causal attribution rather than being given it, the signal stays above chance but its advantage over confidence disappears, falling to AUROC 0.683 below entropy's 0.719 at the attribution quality a small model supplies. The matched control is necessary where the intervention is non-specific (ablation, AUROC +0.016) and inert where it is intrinsically specific (a counterfactual swap), so a control matters exactly when there is generic perturbation-sensitivity to cancel. The finding is that whether an answer survives counterfactual intervention on its own evidence is a stronger label-free signal of correctness than how confident the model is, provided the evidence can be located.

A human newborn is a living organism with homeostatic regulation, affect, sensorimotor coupling, social-orienting biases, and a powerful capacity for statistical learning. It is not yet a language-bearing, norm-bearing, autobiographically extended self. By the age of three to five a recognizable someone has formed, and by adulthood the same initial architecture can support a scientist, a citizen, a parent, a legal subject, a reflective philosophical agent. This paper proposes a bootstrapping architecture that accounts for that divergence. It takes as its starting point the interactionist consensus already established by neuroconstructivism and developmental systems theory, that development is neither maturation alone nor inscription on a blank slate but emerges from organism-environment coupling, and its contribution is not that premise but a specific staging of it. The thesis is that adult cognitive personhood is co-compiled: biologically prepared proto-agential systems enter caregiver-mediated loops of affect regulation, joint attention, language, norm enforcement, autobiographical narration, executive scaffolding, and cultural-tool use, and the child progressively internalizes the regulatory and symbolic structures first run around it by others. The argument is an integrative synthesis with a defined scope. It stages a determinate proto-agent-to-person stack and integrates developmental psychology, attachment theory, predictive processing, language-acquisition research, Vygotskian internalization, Tomasellian shared intentionality, autobiographical-memory theory, executive-function development, and cultural evolution, carrying the bootstrapping process past childhood into the institutional layers that most developmental syntheses leave unaddressed, and it tests the integrated account against the boundary cases of institutional deprivation and late first-language acquisition. The result is a layered control architecture in which each acquired layer rewrites the learning conditions for the next. The paper is candid that an integration of this breadth is hard to falsify as a whole and that much of what it accommodates it accommodates by construction, and it isolates the mediation and ordering claims on which the architecture could actually be tested. The account is developmental and cognitive, and is held strictly separate from claims about moral or legal status: the newborn lacks adult-style cognitive personhood and retains full moral standing. The developmental layering mirrors, under compression, the layered history of the person concept itself, which classical anthropology has long treated as a category assembled across the history of societies rather than given whole.

Across physics, biology, cognitive science, and machine learning the same move recurs: a domain that is not spatial is rendered as a structured space of possible forms, and its behaviour is treated as movement, transformation, and constraint within that space. Configuration spaces, morphospaces, conceptual spaces, and the latent spaces of trained networks are instances. This paper argues that the move is not a set of independent metaphors but the latest phase of one long process in the history of geometry, the progressive detachment of form from physical extension. We trace that history stratigraphically, from the surveyor's field and the Euclidean figure through coordinates, the collapse of the parallel postulate, Riemann's intrinsic manifold, Klein's transformation groups, topology, phase space, and information geometry, to biological morphospace and the representational geometry of brains and machines, each stratum remaining in use beneath the next. We then name the current phase morphotopy and, so that the name does more than relabel, give five conditions a treatment must satisfy to count as a morphotopic one rather than a spatial figure of speech, and apply them to cases that pass and cases that fail. One structural obstruction is definable in any of these spaces, geometric frustration, the failure of locally satisfiable constraints to compose into a global solution; we exhibit it exactly in a minimal spin model, the one stratum where it can be computed cleanly, and argue rather than demonstrate its recurrence elsewhere. We date the strata to show that the abstraction accelerates, three quarters of them falling in the last twentieth of the span, an acceleration that is in part a feature of how finely the recent strata are individuated. The claim is historical and organizational rather than a new mathematics, and it is bounded: rendering the structure of a mind as navigable geometry is not an account of why there is experience.

Care is usually studied as a feeling, a virtue, or a practice, and studied that way it is hard to see what a hospital and a mother have in common, or why good intentions so often miscarry. This paper takes a colder primitive: one system cares for another when the other's viability enters its perception, model, and choice of action, with the ethical senses adding valence, an explicit model, and a submission to the other's consent. The claim is that this relation has a genuine geometry, and that the geometry is the information geometry of the other's viability. Represent the cared-for as a distribution over the states in which it can persist. The space of such distributions is a statistical manifold whose metric is not chosen but forced: Chentsov's theorem makes the Fisher-Rao metric the only one invariant under sufficient statistics, and on the three-state simplex it is a sphere of radius 2 with constant curvature a quarter, a value a from-scratch curvature routine recovers to machine precision. Care is then transport, moving the other along a geodesic toward the viable region, and the care-distance is the geodesic's length. Three results follow from the construction rather than from any tuning of it. Care-distance is a genuine metric, so it obeys the triangle inequality and the older claim that concern could reach a stranger more cheaply by routing through a friend was an artefact; the real obstruction is curvature, and a care-triangle carries a measurable spherical excess, a quarter of its area, by the Gauss-Bonnet theorem. Direction enters through the carer's control metric, a reweighting whose curvature is no longer constant and under which a distant stranger costs five times what kin cost along the same path. And the constraint that separates care from domination becomes a boundary fact: total coercion drives the other to a vertex of the simplex, a distribution of zero entropy, which is exactly Canguilhem's definition of the pathological, a life reduced to a single norm; holding the other off that boundary costs little viability, while abandoning it to an unguided drift costs half. What the geometry cannot reach is the inside of the relation, and the paper ends there, with Ruyer, on the one region that has no distance because it is present to itself without survey, and which a geometry of distances is built to miss.

A firm grows from 100 units to 120 after hiring 10 people, and every available theory of value answers the question of who made the 20 badly: labor-value accounting hands it to an undifferentiated substance called labor, marginal-product accounting books it to whoever was measured last, and bonus arithmetic divides it by headcount. The question is a credit-assignment problem, the same one reinforcement learning has carried since its founding, and cooperative game theory solved its allocative core in 1953 with the Shapley value, the unique division of a jointly produced surplus consistent with efficiency, symmetry, and indifference to everything except contribution. This paper asks what political economy looks like when that solution is taken seriously at the moment machine intelligence becomes a production factor. A minimal firm, solved exactly, shows naive growth attribution overstating the new hires' contribution more than fourfold and isolates a residue of the surplus, the three-way synergy, that belongs to no participant alone. A task-continuum production game over intelligence, compute, energy, and human labor, solved exactly along the whole automation path, shows total output multiplying sixfold while labor's share of attributed value collapses from all of it to an eighth, and it yields two structural facts the singularity discourse misses, both consequences of modeling the stack factors as perfect complements: the Shapley values of intelligence and its essential complements are forced equal by symmetry, so attribution cannot see which member of the stack is the clever one, and intelligence's share of value is capped at one-third by two essential complements, at 1/(k+1) by k, for any finite capability whatever. Within that perfect-complement structure the Shapley singularity, intelligence's share of attributed value approaching one across frontier production games, is unreachable by capability growth alone: it is an ownership event, a redrawing of the boundaries around the complements, rather than an intelligence event. The qualification carries the weight: with imperfect substitutability the equalities loosen and the ceiling moves, and a capability path that erodes a complement's essentiality is a genuine route to a higher ceiling the model excludes by assumption, so the ownership thesis holds within the model rather than unconditionally. A final pair of games locates the last human wage in verification, the best-attributed factor on the frontier until machine output certifies itself, after which labor's share falls threefold in one step, and exhibits the gap between the Shapley point and the core, between what each factor contributes and what its owner can stably extract, which is where politics lives. The frame is bounded: a Shapley value is an audit under a chosen value function, never a desert claim, and the choices of what counts as value and what follows from the audit remain political acts that no attribution mathematics can perform.

Modern agents act atop layered inheritances they did not build and cannot fully perceive: the oxygenated atmosphere, the mitochondrial machinery that powers a thought, controlled fire, written memory, the compiler that turns code into machine behavior. This paper develops a concept for the failure these inheritances produce. Arespect is not disrespect, which perceives a value and diminishes it, and not ingratitude, which withholds feeling toward a recognized debt. Arespect is the failure to perceive the depth, cost, and difficulty of an enabling condition at all, the blankness that settles once a hard-won achievement stabilizes into a reliable interface. Its mechanism is stack blindness: a working layer withdraws from attention because we act through it rather than toward it, and returns to view only when it breaks. The argument distinguishes arespect from a family of neighboring concepts (disenchantment, reification, alienation, inattentional blindness, the withdrawal of the ready-to-hand, infrastructural invisibility, black-boxing, tacit knowing), each of which names the mechanism or the content of recession while none names the dispositional failure of regard that recession makes possible. It then traces the stack across scales, from the geological inheritance of breathable oxygen and the cellular inheritance of mitochondrial energy through fire, writing, and software, holding every empirical claim to a discipline that separates established science from plausible inference, live controversy, and speculation. The cure proposed is not gratitude but depth perception, including honest perception of how much of the depth remains unknown. Civilization is read as accumulated impossibility rendered usable, and arespect as the price of the rendering.

To say a group voted against its interests is to run a calculation and hide most of its inputs: it presupposes one interest where a voter has several, reads a candidate's likely harm to a group while ignoring whatever the candidate offers, treats a knowable harm and an unforeseeable one as the same act, and never says whether it is counting voters or counting votes. This paper turns the accusation into a measurement model that exposes each input and prices the verdict's dependence on it. Adverse policy risk for a group, candidate, and policy domain factors as a product of exposure, institutional dependence, candidate hostility, implementation probability, and magnitude; a foreseeability gate and a salience gate separate the risk a voter could know and weigh from raw exposure; and net alignment under a given definition of interest is the candidate's weighted protective benefit minus the weighted adverse risk, so that gross risk and net alignment are different quantities whose difference is exactly the protective term the folk accusation drops. Run on three synthetic but anchored 2024 cases (support for Trump among LGBTQ, Muslim, and Latino voters), solved exactly and propagated through 40,000 seeded Monte-Carlo draws, the model shows three buried parameters governing the answer. The interest function moves the same Latino vote from aligned under a material reading to misaligned under a rights reading, flipping the verdict's sign for one of three groups while it holds negative for the other two. The counting frame inverts the most-misaligned group from LGBTQ, judged per supporter, to Latino, judged per the votes a bloc moves. The awareness gate removes between 43% and 58% of each group's gross risk, and because each input is propagated at its own measurement-quality concentration the leftover uncertainty is dominated by the quantities measured worst, exposure and salience and awareness, while the policy-hostility term the accusation leans on carries least. One configuration is robust, the rights-dependence reading of the LGBTQ case per voter, surviving in 90% of draws; the headline ordering as a whole survives in 66%. The construct's contribution is to make the hidden inputs explicit and report how much of any verdict each one carries. It does not issue the verdict, and the paper is precise about why it cannot.

When a nationalist party becomes a plausible route to power, the claim that capital is moving to the far right runs a calculation and hides its unit. Capital is a population of firms whose exposure to state discretion and whose freedom to leave differ widely, and those two quantities decide the response. This paper prices the situation as a contingent claim. A firm holds a share of its value in streams the state allocates and a share disciplined by markets, and can move a fraction of itself abroad; a sovereigntist government arrives with some probability, reallocates part of the state-exposed value, and raises the country risk premium. Each firm then takes Hirschman's menu as a menu of options: hedge by buying a call on reallocated rents through access, exit by pulling mobile capital out of a repriced exposure, or voice by lobbying for stability. Solved over 60,000 firms, the model locates a state-dependence threshold near 0.42: below it firms exit, above it they hedge, so the same rising probability that drives mobile capital out drives state-dependent capital toward access. A bond sell-off and a rush for political access in the same week are one repricing seen from two fractions, and the aggregate direction of capital is an artifact of which fractions the economy holds. Across three regimes the access value at stake runs from about 0.13 under consolidated governance to 0.02 where discretion is credibly capped. The sharpest result is negative: a cross-section of firms holding access to a sovereigntist party is generated equally by option-buying and by capture, and the two part only in the response to a change in the probability, where the option unwinds and the capture does not. The same two parameters, rent-discretion and how settled a regime is, place regimes well beyond the European semi-periphery: a regime map puts Norway at the boundary where institutions cap the rents to nothing, the Gulf rentier monarchies at the boundary where the future is already settled and access decays into tribute, and consolidated, party-state, and predatory autocracies in between, with sovereigntist optionality proper the contested interior. The unit the slogan hides is the firm, indexed by how much of its value the state allocates and how much can leave; one probability sorts that population across a threshold, so the aggregate sign of capital's move is a fact about the composition of the capital stock rather than about its politics. Ships a runnable simulation whose output carries every modelled number.

A consumer who will not switch phones over a country's politics will switch beer over an advertisement, and the asymmetry is usually explained by the phone costing more. This paper argues it is explained by the phones differing and the beers not. When functional quality across competing products has converged, political meaning takes over the work of differentiation, and it does so in two ways an outside observer cannot tell apart. Under sorting, politics changes which good a consumer chooses while leaving the good's experienced quality intact; under flavoring, politics changes the experience itself, so the same physical good is tasted as better when its label aligns and worse when it does not, the way a price tag moves reported pleasantness in the neural-imaging studies. The paper builds the smallest model that holds both channels and a mixture parameter for how much of the political premium runs through flavoring rather than sorting. Solved over a population of 60,000 consumers in a polarized electorate, the first result is negative and exact: at the quality plateau the labeled market is invariant to the mixture, so the choice data that revealed-preference analysis relies on are identical for every split between the two channels, and the political premium of 0.362 in functional-quality units is consistent with any flavoring share in the unit interval. Only a quantity the market never generates identifies the split, the within-good swing between a good's blind rating and its labeled rating, which the model fixes at 0.202 per unit of flavoring share and which is exactly the design of the relabeling experiment. The second result makes the plateau quantitative: sweeping functional dispersion locates a threshold near 0.445 below which politics decides the majority of contested choices, and placing ten categories on that axis reproduces the received susceptibility ordering, bottled water through cosmetics on the flavor-dominated side, phones through medical devices on the other, as an output rather than an assertion. The third result prices brand activism. Because moral contamination is negativity-dominant, with the misalignment penalty here 2.33 times the alignment bonus, a neutral good that adopts a political position loses a centered mass-market base by 0.036 of demand while a symmetric model gains it 0.008, and the position pays only for a base already concentrated beyond 0.425 toward the adopted pole, which is the large-share-brand penalty derived from the contagion asymmetry alone. The construct says what would have to be measured to call any market politically flavored. It does not certify that any named market is, and the closing section is precise about the difference. Ships a runnable simulation whose output carries every modelled number.

The title overstates on purpose, and the paper spends its length earning the qualification it needs: we have never been intelligent alone. Human organisms hold real and unequal cognitive capacities, but the performances usually offered as proof of intelligence, in mathematics, physics, law, engineering, administration, and now artificial intelligence, are rarely produced by an isolated brain. They are realized by assemblages of trained bodies, notations, instruments, collaborators, procedures, and inherited institutions, and then attributed, after the fact, to a single named person. This paper reads that attribution through Bruno Latour. Latour argued that the modern settlement works by purification, declaring a clean separation between nature and society while its laboratories proliferate the hybrids the separation forbids anyone to name. An analogous settlement governs intelligence. A Cognitive Constitution depends operationally on cognitive hybrids while representationally and normatively purifying their outputs into properties of bounded individuals, and the operation that performs the purification is what the paper calls attributional compression: collaborative production booked to a genius, an educational infrastructure booked to an examination score, a software ecosystem booked to a programmer's talent, a corpus and a data-center booked to a model. The argument keeps four things apart that the folk concept fuses, capacity and performance and realization and historical possibility, so that psychometrics, distributed cognition, and cultural evolution stop reading as rivals and start reading as descriptions of different levels. It rests on the defensible claims, that the realizing system for many advanced tasks exceeds the individual and that purification happens through observable attribution practices, and it declines the contested metaphysical claim that the notebook is literally part of the mind. A recent and contested line of work pushes the relocation to its limit, proposing that the capacity to learn is a generic affordance of mathematical structure that natural selection discovers rather than builds, so that random unselected networks already sit close to optimal for learning; if that holds, then crediting human biology for intelligence is the same compression run at the largest scale, credit assigned at every level and an author at none. The closing move is the sharp one: purification is not an error to be debugged but the accounting system of merit, which is why the individual mind survives every demonstration of its insufficiency. Intelligence names a settlement about who gets the credit, and we have been litigating it, under that name, for a long time.

A hypothesis circulating in Bucharest holds that Romania's turn toward gas serves an American strategy to keep Europe on fossil fuels and out of an electrified, autonomous future, and it prices the stake as a 1.7 to 3.6 percent hit to the United States current-account deficit. The intuition is right and the arithmetic is wrong, and fixing the arithmetic requires distinguishing two forms of energy power that the word energy fuses. Molecule power is power over a flow metered at a chokepoint: the customer must keep buying, the rent recurs, the leverage is a tap. Electron power is power over a stock and a standard: the customer buys the capital once, the fuel is then free, and the residual dependence is on equipment, software, and standards rather than on a daily delivery. The asymmetry that organizes everything is that molecule revenue requires the customer to stay un-electrified, so a fossil exporter has a structural interest in delay. This paper builds the smallest dynamical model that makes the interest legible: an importing region's electric share follows a replicator equation with technological learning, which is bistable, with an interior takeoff threshold the exporter's price can move. Solved on that model, three results follow. First, the exporter faces a dilemma between harvesting the fossil age at a high price, which accelerates the importer past the threshold and ends the rent, and extending it by limit-pricing, which keeps the importer trapped; extending beats harvesting by a present-value factor near 1.9 only when the exporter is patient and substitution would otherwise be fast, so a state that suppresses its own price reveals it believes takeoff is near. Second, the original current-account arithmetic is an order of magnitude too large and in the wrong ledger: the displacement is about 0.13 to 0.23 percent of the deficit, while the stake the exporter actually defends is the capitalized European rent a takeoff terminates, near 289 billion dollars, roughly 113 times the annual trade-flow figure. Third, the delay is bounded: because the learning curve is global, an exporter that suppresses one region's deployment still faces a threshold dragged down by whatever rival keeps deploying, so limit-pricing buys a finite delay of years, not a reprieve. The Romanian outcome the hypothesis fears is generated by each actor's independent optimum without any coordination, which makes the conspiracy reading not merely unproven but unnecessary, and the paper is precise that pricing the incentive is not establishing the instruction. Ships a runnable simulation whose output carries every modelled number.

In creative fields, low pay and anti-commercial rhetoric read as proof of devotion, but the same surface, negligible art income and a claim to be doing it for the love, is produced by at least five different situations: freedom from commerce, refusal of commerce, exclusion from commerce, dependence concealed as devotion, and hobbyist indifference. Because these are observationally identical at the level of the individual, no quotation or earnings figure can sort them, and the right unit of analysis is not the artist but the field, modelled as a filter. This paper builds that filter as an agent-based model, the Purity and Scene Machine, in which 4,000 artists differ in parental wealth and in talent and scene-value propensity drawn independently of wealth, so the model never assumes the rich are more gifted or more sincere. Wealth supplies a financial buffer and, more importantly, a recurring hidden subsidy that lets a purity posture be sustained rather than merely cushioned; purity forgoes commercial income, and a field that rewards anti-commerciality excludes the visibly commercial through its gatekeepers. Run as a survival filter over ten years, the model reproduces the class gradient that the empirical literature reports, a 4.99-fold advantage in survival for artists from the top wealth third over the bottom, out of an economic filter alone with talent held independent. The field's average purity rises from entry to survival, from 0.42 to 0.50, not because survivors are more devoted but because the poor and devoted were filtered out, so observed luxury purity is a survivorship artifact; purity in the model correlates with wealth at 0.58 and with talent at minus 0.03. Strengthening the field's purity norm amplifies the class gradient 3.2-fold, which is the sense in which a norm that presents as aesthetic operates as a wealth means-test. The field loses about 69 percent of its scene and archival value to dropout, and because platform attention is a weak proxy for cultural value in the model, correlation 0.30, the stewards it starves are precisely its least visible members. Among policy levers, a recurring public income floor closes the class gap most while preserving the most scene value, because it lets low-buffer artists sustain non-commercial work without either commercializing or leaving; a one-off buffer barely moves a recurring shortfall. The model does not show that any wealthy artist is insincere; it shows that wealth changes the price of purity, the probability of surviving it, and the class composition of who remains. The question a field should ask is not who is pure but who can afford to be. Ships a runnable simulation whose output carries every modelled number.

Generative AI did not create a cheating problem in education; it broke an inference. The take-home essay, the problem set, and the coding assignment were never the thing being measured. They were proxies, and the institution read backward from a good artifact to a competent student. AI severs that link, because a good artifact can now be produced whether or not the student is competent, so the crisis is one of assessment validity and attribution rather than of integrity. This paper models the artifact as a measurement channel from a student's latent competence to an observed score, and asks three computable questions. First, when does the proxy collapse? Treating substitution as a rational choice, the artifact's validity, the correlation between the recorded score and true competence, falls from about 0.90 under weak AI to near zero under strong AI on a substitutable task, and the collapse is governed by task substitutability: validity crosses one half at a substitutability near 0.68, so low-substitutability work such as supervised and practical assessment keeps a validity near 0.95 while the open take-home loses almost all of it. Second, can detection save the artifact? No, and the reason is Bayesian: with a detector that flags AI text at a true-positive rate that erodes under cheap evasion and a false-positive rate that does not, at a moderate prevalence the posterior that a flagged student actually substituted is about 0.58 at realistic evasion and below a coin flip beyond it, so roughly 0.42 of accusations fall on the honest while the motivated evade, and enforcement manufactures injustice rather than deterrence. Third, what restores trust? A verification probe, a short viva or a request to modify or explain, measures agency rather than output and is a second channel that substitution cannot fake; because the probe also deters, sampling it on about a third of students restores validity to 0.81 at 6.5 percent of the cost of securely assessing every artifact, and targeting the probe at the suspicious cases reaches 0.95 against 0.66 for random sampling at the same budget. The model separates the crisis from teacher quality: the channel breaks as a structural property of substitutability regardless of how well a course is taught, though poor assignment design accelerates it, so the popular reduction to teacher incompetence is wrong about the cause and right only that bad design amplifies it. The equity costs of the second channel are real and the paper does not minimize them. What AI made expensive is exactly what the artifact was always a proxy for: the agency to explain, reproduce, modify, transfer, and defend the work. Ships a runnable simulation whose output carries every modelled number.

A provocative thesis reads the present authoritarian and plutocratic turn as an anticipatory counterrevolution, a reaction by the powerful against a near future in which AI, open-source investigation, automated translation, and computational audit make impunity impossible to maintain. The strong form does not survive: the movements predate the technology and are already explained by status threat, nationalism, elite economic interest, and institutional weakness. But a narrow mechanism does survive, and this paper prices it. Whether a wrongful act meets a consequence is the product of a chain of stage probabilities, discovery, attribution, publication, prosecution, independent judgment, and enforcement, and a product is bounded by its weakest term. Generative AI raises the first three, the evidence stages, toward one; it does not raise the last three, the institutional stages, which are human and capturable. Solved over a population of cases, the model yields three results. The bottleneck moves: before AI the binding stage is discovery and the probability of consequence sits at 0.078 of what the institutions would allow, while after AI it sits at the institutional ceiling exactly, so further evidence buys nothing and the binding constraint becomes adjudication. Impunity is conserved: cheaper evidence would multiply the probability of consequence nearly thirteenfold, from 0.010 to 0.123, but a self-protective elite rationally captures the new bottleneck, prosecution and judgment, and that capture absorbs the entire evidence gain, returning consequence to 0.007, marginally below where it began. And exposure decouples from consequence: the share of wrongdoing that becomes publicly legible rises toward one while the share that is sanctioned does not, so the gap between what is known and what is answered widens roughly fifteenfold, which is the pattern of a leak that reveals everything and a trial that acquits everyone. The same chain run downward, the state's automated case against a subject, inverts: AI raises the accusation stage and capture removes the protections, so a captured system harms the weak about sixfold. The age AI inaugurates is therefore not the last age of impunity but the age of impunity made visible, and what to do about visible impunity is a question about courts and enforcement, not about models. The construct prices an incentive structure; it does not measure any country, and it is explicit that the motives of real movements are multicausal and largely older than the machine. Ships a runnable simulation whose output carries every modelled number.