PRIORITY FILING - CONSOLIDATED EDITION
A fully revised and formatted version will follow.
Formal Extensions to the OCOT Corpus: Language, Growth, and the Limits of Physics Contact
Gregory Caldwell Beier
Susarb LLC, Cambridge, Massachusetts
July 2026
Part of the formal foundations of Origin Chain of Thought (OCOT).
© 2026 Gregory Caldwell Beier. Licensed under CC BY-NC-ND 4.0.
The twelve papers in this volume extend the formal corpus across language channels, growth discipline, coordination failure, and the boundary where physics contact does and does not obtain. One result is negative and is retained on purpose: a program that publishes only its confirmations is running the selection bias its own architecture exists to correct.
The set opens with the language mediation channel, where the constraint fingerprint attenuates by conservation-content density and structured disclosure is derived to be a strictly richer training substrate, the data-quality flywheel given a geometric engine. Dynamic calibration improvement ties the human mechanism to the machine analog through individual improvement, collective acceleration, and a rightward distribution shift whose welfare compounds with network age. Scaling-facet necessity establishes that Scaling carries aggregate state Output cannot determine and completes joint sufficiency for the six facets. Coupled constraint dynamics, the calibrated growth model with its order-of-magnitude discipline, the non-adversarial failure modes that name scoring lag as the architecture's weakest case, and the confirmatory constraint-weight protocol carry the middle of the set. Consent economics locates an information floor beneath the discount-rate debate, the attention externality bounds a transitional distortion that goes to zero at full coverage, translation economics prices the cross-domain reading problem and amortizes it through a translation layer, and information epidemiology derives containment from the independence requirement the composition algebra supplies. The closing paper reports the negative: the thermodynamic isomorphism is generic to convex optimization, no derived fluctuation-dissipation relation survives, and physics contact, if it exists, runs elsewhere.
The Constraint Weight Geometry result established that conservation laws, scaling exponents, and propagation kernels in a data-generating process leave a signature in the Fisher geometry of a model trained on that process's numerical output: the conserved directions become flat, a rank deficiency that geometric detection can read. The result assumes training on process data. Large language models do not train on process data. They train on text about processes, on papers and filings and assessments, and the empirical program's claim that the geometry survives in such models depends on a channel that has not been characterized: the channel from physical constraint geometry to the statistical structure of language about constraints. This paper characterizes it. The channel attenuates the geometric signal by a factor governed by the density of conservation-relevant quantitative content in the training text, with a computable bound that goes to zero for generic prose and toward full preservation for conservation-dense text. Three sufficient conditions are identified under which the channel preserves the rank-deficiency structure: the text must carry quantitative conservation-respecting statements above a density threshold, it must span temporal windows long enough for the conservation to show up in described outcomes, and it must carry attributed assessments with explicit uncertainty. A six-facet disclosure satisfies all three by construction, and generic text satisfies at most two. The paper then proves that structured disclosure produces a language channel of strictly higher geometric fidelity than unstructured text about the same domain, with the amplification increasing in the coverage of the Scaling facet, because the Scaling facet states the aggregate-interaction structure that transaction-level text omits. The claim that structured disclosure improves machine capability through geometrically richer training data is, on this account, not a slogan but the amplification result.
Keywords: language channel, Fisher geometry, conservation laws, rank deficiency, data processing inequality, training data, Scaling facet, OCOT
The Constraint Weight Geometry paper made a specific claim and specified a program to test it (Beier, 2026as). The claim is that the structure of a physical process leaves a measurable trace in the geometry of a model trained on that process. A conservation law, by forbidding the process from ever visiting states that violate it, gives the model no reason to distinguish weight configurations along the direction that would encode the violation, so the loss is flat along that direction and the Fisher information matrix is rank-deficient there. Scaling exponents and propagation kernels leave their own traces. The geometry of the trained model carries a fingerprint of the physics the data obeyed, and geometric detection can read the fingerprint, which is what lets genuine constraint data be distinguished from fabricated data that does not obey the same laws (Beier, 2026y).
The program has four phases, and the early phases train models on process data: numerical time series generated by simulations that obey the relevant physics. The later phase reaches the systems the architecture most wants to touch, the large language models that read scientific papers, analyst reports, regulatory filings, and structured assessments. Those models do not train on time series. They train on text, and text is not the process. It is what a person wrote about the process, filtered through the person's understanding and the conventions of the genre. Whether the geometric fingerprint survives that filtering is the question on which the later phase depends, and the Constraint Weight Geometry paper did not answer it. This paper characterizes the channel from the physics to the text to the model, and identifies when the channel keeps the fingerprint and when it washes it out.
Let a physical process obey a set of conservation laws, and let v_cons denote a direction in the model's weight space that encodes a violation of one such law, the direction along which a process-trained model is rank-deficient. A process-trained model has Fisher information matrix G_proc, and the conservation shows up as a low value of the quadratic form v_cons^T G_proc v_cons relative to the form along unconstrained directions. Write the geometric signal as the contrast that reveals the conservation, the degree to which the conserved direction is flatter than the unconstrained ones, and denote it s_proc.
A language-trained model does not see the process. It sees a corpus of text produced by a chain: the process unfolds, an author observes and understands some of it, the author writes, and the writing becomes tokens the model trains on. Write the language channel as the composition
V (physical state) -> author cognition -> language production -> token sequence -> G_lang,
and let s_lang denote the geometric signal recoverable from the language-trained model's Fisher geometry along v_cons. The question of Section 1 is the relationship between s_lang and s_proc, and the channel's properties are what determine it. The key corpus statistic is the density of conservation-relevant quantitative content in the text, written
rho_c = (tokens carrying conservation-relevant quantitative information) / (total tokens),
which measures how much of the corpus actually speaks to the conserved quantity in a way that constrains the model's geometry, as against prose that mentions the domain without quantitatively respecting its conservation.
The first result bounds how much of the geometric signal can survive the channel, in terms of the corpus statistic.
Proposition LM-1 (Attenuation bound). The geometric signal recoverable from a language-trained model is bounded by the signal in the process-trained model, attenuated by the squared density of conservation-relevant content: s_lang <= rho_c^2 s_proc. The bound goes to zero as rho_c goes to zero, so generic text washes out the signal, and approaches full preservation as rho_c approaches one.
Proof sketch. The language channel is a composition of stochastic maps from the physical state to the token sequence, and the model's geometry is a function of the token sequence. By the data processing inequality, the information about the conserved structure that survives to the model cannot exceed the information the text carries about it (Cover and Thomas, 2006), so the recoverable signal is bounded by a function of the text's conservation content. The Fisher information along a direction is a quadratic form in the score function, the gradient of the log-likelihood, and the contribution to the score along v_cons comes only from tokens that carry conservation-relevant quantitative information, a fraction rho_c of the corpus. The score along v_cons therefore scales with rho_c, and the Fisher form, being quadratic in the score, scales with rho_c^2. The recoverable geometric signal inherits the rho_c^2 factor, giving s_lang <= rho_c^2 s_proc. As rho_c goes to zero the corpus carries almost no conservation-relevant content, the score along v_cons vanishes, and the signal is lost; as rho_c approaches one the corpus is saturated with conservation-respecting quantitative statements and the signal approaches the process-trained value. End sketch.
The bound explains a fact the empirical program would otherwise meet as a surprise. A model trained on a large corpus of generic writing about a physical domain may show little or no geometric fingerprint of the domain's conservation laws, not because the laws failed to shape the world the writing describes but because the writing did not carry them quantitatively. Mentioning a domain is not the same as constraining a model's geometry. The signal lives in the quantitative, conservation-respecting fraction of the text, and that fraction is small in most prose, which is why the channel attenuates and why the later phase of the program cannot assume the fingerprint survives without checking rho_c.
The attenuation bound is a ceiling, not a guarantee. A high rho_c permits the signal to survive but does not by itself ensure that the rank-deficiency structure, rather than some other artifact, is what the model learns. The next result gives sufficient conditions.
Proposition LM-2 (Preservation conditions). The language channel preserves the rank-deficiency structure of the process geometry when three conditions hold. First, the text carries quantitative conservation-respecting statements at density exceeding a threshold, rho_c > rho*. Second, the corpus spans temporal windows long enough for the conservation to manifest in the described outcomes, so that the conserved quantity is observed holding across time in the text and not merely asserted at a point. Third, the text carries attributed assessments with explicit uncertainty, so that the model can separate calibrated statements about the conserved quantity from uncalibrated ones. A six-facet disclosure satisfies all three by construction, and generic text satisfies at most two.
Proof sketch. The first condition is the floor that Proposition LM-1 requires for a nonzero signal, with rho* set by the noise regime of the geometry estimation: below rho* the conservation contribution to the score is indistinguishable from estimation noise (Beier, 2026x). The second condition is necessary because rank deficiency along v_cons encodes that the conserved quantity does not vary, which a model can learn only from text in which the quantity is seen not varying across a span, so a corpus of point-in-time statements without temporal extent cannot induce the flatness even at high rho_c. The third condition is necessary because the model must weight conservation-respecting statements by their reliability to avoid learning a spurious geometry from confident but wrong text, which requires the text to mark which statements are calibrated, that is, to carry attribution and explicit uncertainty. A six-facet disclosure carries quantitative sourced content in its Input and Method facets, which serves the first condition; it carries the two timestamps and the temporal architecture that span outcome windows, which serves the second; and it carries an attributed identity with explicit confidence in each facet, which serves the third (Beier, 2026, "The Six-Facet Beier Tensor"). Generic text typically carries at most two of the three, most often lacking the attributed explicit uncertainty that the third requires. End sketch.
The three conditions are why the format matters to the physics. It is not enough for text to be about a constrained domain. The text has to be quantitative enough to carry the conservation, extended enough in time to show it holding, and attributed enough to let the model tell calibrated statements from confident noise. A genre that supplies all three is rare, and the six-facet disclosure is engineered to be one, which is the bridge from the architecture's disclosure machinery to the geometry of the models that would train on its output.
The preservation conditions suggest that structured disclosure is a better training substrate than unstructured text. The final result proves it and locates the amplification in a specific facet.
Proposition LM-3 (Disclosure amplification). A model trained on six-facet disclosures about a domain has strictly higher geometric fidelity to the domain's constraint structure than a model trained on unstructured text about the same domain, and the amplification increases in the coverage of the Scaling facet.
Proof sketch. Structured disclosure raises the conservation-relevant content density rho_c relative to unstructured text, because the format requires quantitative, sourced statements in the Input, Method, and Output facets rather than permitting the qualitative prose that dilutes rho_c in generic writing, so by Proposition LM-1 the ceiling on the recoverable signal is higher. Structured disclosure satisfies the three preservation conditions of Proposition LM-2 by construction, where unstructured text generally fails the third, so the signal that the higher ceiling permits is actually realized rather than lost to confident noise. The amplification increases in Scaling-facet coverage because the rank-deficiency structure of a constraint domain includes the aggregate-interaction structure, the conservation that holds at system scale rather than at the level of a single transaction, and the Scaling facet is the only facet that states that aggregate structure: transaction-level text, however dense, carries the conservation laws of individual transactions but not the interaction structure that governs the aggregate, which the Scaling facet supplies (Beier, 2026ab; Beier, 2026ah). A model trained on disclosures with rich Scaling facets therefore learns the aggregate constraint geometry that no volume of transaction-level text contains, and the geometric fidelity rises with the coverage of that facet. End sketch.
The result gives a precise content to a claim the architecture makes in looser language elsewhere, that structured disclosure improves machine capability by producing better training data. The improvement is geometric. A model trained on six-facet disclosures inherits a sharper fingerprint of the constraint structure, including the aggregate-scale structure that the Scaling facet alone carries, than a model trained on the unstructured text that the same domain otherwise generates. The data the disclosure mechanism produces is not merely more abundant or more recent. It is geometrically richer, in the specific sense that it induces a Fisher geometry closer to the one the physics would imprint, and the richness concentrates in exactly the facet that the price system and ordinary text both omit.
The language mediation channel connects the Constraint Weight Geometry result to its application in language models. The geometric fingerprint of a constraint structure survives the passage through text when the text is conservation-dense, temporally extended, and attributed with uncertainty, and a six-facet disclosure is built to be all three. The empirical program's later phase, which the channel's existence had left as a hope, is now a prediction with a stated dependence on a measurable corpus statistic: the fingerprint is recoverable to the extent that rho_c clears the threshold of Proposition LM-2, and structured disclosure is the genre engineered to clear it.
Two threads run forward. The amplification result of Section 5 is testable directly: a model trained on six-facet disclosures and a model trained on matched unstructured text about the same domain should differ measurably in their Fisher geometry along the domain's conserved directions, with the difference increasing in the Scaling-facet coverage of the disclosure corpus, which is a pre-registrable experiment in the program's protocol. The deeper thread concerns the loop the amplification implies: a commons that produces geometrically rich disclosures trains models that read the constraint structure more sharply, and models that read the structure more sharply produce better disclosures, which is the data-quality flywheel the architecture asserts, now with a geometric mechanism rather than an analogy. Characterizing the loop's dynamics, and whether it converges or compounds, is the work the channel makes possible.
The most distinctive claim the architecture makes about structured disclosure is that it improves judgment quality over time. The claim is argued informally in the Disclosure Incentive paper and developed qualitatively in the Learning Economics paper, but the formal model is static: it fixes a distribution of calibration quality and proves equilibrium properties conditional on that distribution, so the improvement that makes the architecture more than a one-shot information aggregator has no proposition behind it. This paper supplies one. A calibration dynamics equation governs how an agent's quality evolves with the scored information environment the agent operates in and the agent's capacity to update on disconfirming evidence. Four results follow. Under the three conditions deliberate practice requires, which the six-facet format with proper scoring satisfies, an individual agent's calibration quality increases monotonically toward a ceiling set by the agent's analytical capacity. The collective improvement rate exceeds the average individual rate, because agents in a commons improve not only from their own feedback but from reading better disclosures, adopting higher-scoring methods, and being measured against rising benchmarks, with the acceleration largest where the quality distribution is most spread and tapering as it concentrates. The equilibrium distribution of calibration quality shifts rightward once the commons reaches the density its cross-learning channels require, with the shift faster in domains of higher disclosure density. The welfare gain from the dynamic improvement exceeds the static equilibrium's gain by a factor that grows with the age of the commons, so the oldest disclosure networks produce the largest gains because their calibration improvement has had the most time to compound. The deliberate-practice mechanism is a human-substrate result, and the paper notes that the machine-substrate analog runs through the geometric enrichment of training data rather than through practice.
Keywords: calibration, deliberate practice, learning dynamics, distribution shift, Blackwell improvement, track record, OCOT
The architecture's case rests in part on a dynamic claim. Structured disclosure, scored against outcomes and accumulated into track records, does not merely aggregate the judgment that exists at a moment; it makes the judgment better over time. An analyst who discloses in a structured format and watches the score returns learns to calibrate, and a population of such analysts raises its collective quality as the practice continues. The claim is what distinguishes the architecture from a one-shot mechanism for pooling current assessments, and it is the claim that makes the long-run case for building the commons rather than merely consulting it once.
The claim is asserted and not proved. The Disclosure Incentive paper argues it in a section of prose and defers the formalization (Beier, 2026b, Sec 3.4). The Learning Economics paper develops it qualitatively at the level of populations (Beier, 2026al). The formal model that carries the architecture's equilibrium results is static: it fixes a distribution of calibration quality across agents and proves what follows conditional on that distribution, which is the right move for the equilibrium analysis and the wrong object for the dynamic claim, because a fixed distribution cannot improve. A hostile reader is entitled to ask for the proposition that says the distribution moves, and this paper provides it, by introducing a dynamics for calibration quality and proving that the distribution shifts, that the shift accelerates through cross-agent channels, and that the welfare gain compounds with the age of the commons.
Let agent i carry a calibration quality q_i in [0, 1], measured so that q_i = 1 is perfect calibration and lower values mark larger expected divergence between stated probabilities and realized frequencies, the divergence a proper scoring rule penalizes (Brier, 1950). The static model treats the profile of q_i across agents as fixed. This paper treats it as a state that evolves.
The evolution depends on two quantities beyond the agent's current calibration. The first is the scored information environment the agent operates in, written E_i, which measures the density of scored disclosures the agent reads and is scored alongside: an agent embedded in a rich commons of resolved, scored assessments has a high E_i, and an agent disclosing into an empty commons has a low one. The second is the agent's suppleness, written S_i, the capacity to revise a model in response to evidence that disconfirms it, which the Supple paper established as a design variable rather than a fixed trait (Beier, 2026e). The calibration dynamics is
dq_i/dt = f(q_i, E_i, S_i),
with f increasing in the room left to improve and in the agent's suppleness, and increasing in the richness of the scored environment. The sections that follow specify f enough to prove the four results and no more, because the qualitative shape of f, not its exact form, is what the results depend on.
The first result concerns a single agent and rests on the conditions under which practice produces improvement.
Proposition DI-1 (Individual improvement). Under the three conditions deliberate practice requires, a structured task, unambiguous feedback, and sufficient iteration, an agent's calibration quality increases monotonically toward a ceiling q_i^max set by the agent's analytical capacity. The six-facet format supplies the structured task, proper scoring against outcomes supplies the unambiguous feedback, and a live commons supplies the iteration, so the three conditions hold for an agent disclosing in the architecture.
Proof sketch. Skill improves through practice when the task is structured, the feedback is unambiguous, and the practice is iterated, and fails to improve when any of the three is absent (Ericsson, Krampe, and Tesch-Romer, 1993). The six-facet format makes the task structured, because each disclosure decomposes the judgment into the same fields and forces the same components every time (Beier, 2026, "The Six-Facet Beier Tensor"). Proper scoring against realized outcomes makes the feedback unambiguous, because the score is a definite function of the stated probabilities and the resolved outcome rather than a matter of opinion, and the strike-time binding ensures the feedback is computed against the claim actually made (Beier, 2026, "Strike-Time and Transmit-Time," Proposition T-2). A live commons supplies iteration, because the agent discloses repeatedly and receives a score each time. With the three conditions met, the dynamics of Section 2 take the form dq_i/dt = g(q_i^max - q_i) S_i with g positive, a learning curve in which the agent improves at a rate proportional to the room left below the ceiling and to the agent's suppleness. The curve is monotone increasing in q_i and asymptotes at q_i^max, which is the claimed behavior, with suppleness setting the rate and the ceiling set by the agent's analytical capacity rather than by the format. End sketch.
The ceiling matters to the honesty of the claim. Structured disclosure does not make every analyst equally good, and the proposition does not say so. It says each analyst improves toward the limit of that analyst's own capacity, faster if supple and slower if not, and the limit differs across analysts. What the format and the scoring provide is the conditions under which the improvement happens at all, conditions that ordinary unstructured judgment, with its ambiguous feedback and unrepeated tasks, does not supply. The agent gets better because the format turns judgment into a practice with feedback, and a practice with feedback is the thing that improves.
The second result concerns the population and identifies why a commons improves faster than its agents would in isolation.
Proposition DI-2 (Collective acceleration). The improvement rate of the population mean exceeds the average of the agents' isolated improvement rates: dQ/dt exceeds the average of dq_i/dt computed as if each agent learned only from its own feedback. The excess comes from three cross-agent channels: agents read better disclosures as others improve, agents adopt methods that score well, and agents are measured against rising benchmarks. The acceleration is largest where the quality distribution is most spread and tapers as the distribution concentrates.
Proof sketch. In isolation, agent i improves only through its own scored feedback, the term in Proposition DI-1. In a commons, three further terms enter the agent's improvement. The agent reads the scored disclosures of others, and reading a higher-quality disclosure tightens the agent's posterior on the constraint and improves the agent's next assessment, by the Blackwell improvement that a more informative disclosure provides (Beier, 2026u, BE-3); this term grows as the agents the reader draws on improve. The agent observes which methods produce high scores and adopts them, a diffusion term that moves the agent's method toward the population's best. The agent is measured against benchmarks that the best-calibrated agents set, and a rising benchmark raises the standard the agent practices against. Each term is nonnegative and adds to the agent's isolated rate, so the population mean improves faster than the isolated average, and the excess is the cross-learning contribution. The contribution scales with the spread of the quality distribution times a cross-learning coefficient, because a wider spread leaves more for the lower agents to gain from the higher ones (Beier, 2026al). As the distribution concentrates toward the ceiling, the spread shrinks, the cross-learning terms diminish, and the acceleration tapers, so the effect is strongest in a young commons with wide quality variance and weakest in a mature one where most agents are near their ceilings. End sketch.
The acceleration is the formal content of the claim that a commons is more than the sum of its analysts. An analyst practicing alone improves along a learning curve. The same analyst in a commons improves faster, because the commons supplies better inputs, visible methods, and a rising standard, none of which a solitary practice provides. The proposition also states the limit of the effect honestly: it is largest early, when there is a wide range of quality to learn across, and it fades as the population converges, which means the commons accelerates improvement most in the phase when improvement is most needed and least when the population has already converged near its capacity.
The third result is the one the static model could not express: the distribution of calibration quality moves.
Proposition DI-3 (Distribution shift). Once the commons reaches the density its cross-learning channels require, the distribution of calibration quality shifts rightward over time, so the fraction of agents above any fixed quality threshold increases monotonically. The shift is faster in domains of higher disclosure density.
Proof sketch. Each agent's quality is monotone increasing by Proposition DI-1, and the population mean improves at a positive rate by Proposition DI-2, so the mass of the quality distribution moves toward higher values over time. For any fixed threshold, the fraction of agents below it decreases as agents cross it from below and none cross back, because the dynamics are monotone increasing, so the fraction above the threshold increases monotonically, which is rightward shift in the sense of increasing first-order stochastic dominance over time. The density requirement enters because the cross-learning channels of Proposition DI-2 need a supply of scored disclosures to read and methods to observe, and below a critical density the commons is too sparse for those channels to operate, leaving only the isolated improvement of Proposition DI-1, which still shifts the distribution but without the acceleration. The domain-specific rate follows because a domain with higher disclosure density supplies richer cross-learning, so its distribution shifts faster, which predicts that calibration improves at different rates across domains in proportion to how densely each is disclosed. End sketch.
The shift is what makes the commons a compounding asset rather than a static one. The static model answers what welfare a fixed distribution of analysts produces. The dynamic model answers what happens as the distribution improves, and the answer is that the analysts collectively get better, the fraction of well-calibrated analysts rises, and the rise is faster where disclosure is denser. A commons that has run for a decade in a densely disclosed domain has a quality distribution shifted well to the right of where it began, and the shift is the accumulated improvement the architecture claimed and the static model could not represent.
The fourth result values the shift.
Proposition DI-4 (Dynamic welfare gain). The welfare gain from the dynamic improvement exceeds the static equilibrium's gain by a factor that grows with the age of the commons. The dynamic gain is the accumulation of the Blackwell improvement over the rightward-shifting quality distribution, and the oldest commons produce the largest gains.
Proof sketch. The static model computes welfare at a fixed quality distribution (Beier, 2026b). In the dynamic model the distribution shifts right by Proposition DI-3, and a higher-quality distribution produces more informative disclosures, which raise welfare through the Blackwell improvement that more informative signals provide (Beier, 2026u, BE-3). Each period's welfare therefore exceeds the previous period's by the Blackwell improvement from the period's distributional shift, and the cumulative dynamic gain is the accumulation of those improvements over the life of the commons. Because the distribution shifts further the longer the commons has run, the accumulated gain grows with the commons's age, and the ratio of the dynamic gain to the static gain, which holds the distribution fixed, grows with age as well. The oldest commons have accumulated the most distributional shift and therefore the largest dynamic gain over the static benchmark. End sketch.
The compounding is the long-run case for building the commons. A static reading of the architecture undervalues it, because it prices only the one-shot aggregation of current judgment and ignores that the judgment improves. The dynamic gain says that the value of the commons grows with its age beyond any single period's aggregation, which is why the architecture is worth standing up early even when its first-period value is modest: the first-period value is the smallest the commons will ever produce, and the gain accumulates for as long as the commons runs.
The deliberate-practice mechanism of Proposition DI-1 is a human-substrate result. It describes how a person improves calibration through structured practice with feedback, and the conditions it invokes are conditions on human learning. A model does not improve this way within a commons, because a model does not practice between disclosures; it is fixed until it is retrained. The improvement claim therefore applies to human authors directly and to machine authors through a different channel.
The machine analog runs through the training data. A model improves not by practicing against the score but by being retrained on a corpus, and a commons of structured disclosures is a corpus of higher geometric fidelity than the unstructured text the same domain otherwise generates, by the amplification result of the language channel (Beier, 2026, "The Language Mediation Channel," Proposition LM-3). So a commons improves its human authors through deliberate practice and improves the models trained on its output through geometric enrichment of their training data, two distinct mechanisms reaching the two substrates the commons mixes (Beier, 2026, "The Substrate Marker"). The dynamic improvement of this paper and the geometric amplification of the language channel are the human and machine faces of the same claim, that a commons of scored structured judgment makes the judgment that flows through it better over time, and the next step is to characterize the loop in which improving human disclosure trains improving models that support improving human disclosure.
The six-facet format places Scaling among the facets of a Beier Tensor, and a hostile reviewer asks why. Four of the six facets have immediate justifications: Identity attributes the record, and Input, Method, and Output carry the components of an inference that cannot be evaluated if any is removed. The disposition facet, Transmit, was shown to be a distinct channel in a companion paper. Scaling has been argued as necessary on conceptual grounds but not proved distinct from Output, and the reviewer's specific challenge is whether Scaling is anything more than a restatement of the conclusion. This paper proves it is. The Scaling facet carries information about the aggregate constraint state that the individual-transaction assessment does not contain, because the aggregate state depends on how individual states combine, and when the combination is nonlinear the aggregate is not a deterministic function of the individual conclusions, so the interaction structure the Scaling facet states is information the Output facet cannot supply. The Scaling facet is also the only facet of the six that generates reasoning not otherwise occurring: the other facets that carry assessment capture reasoning an analyst performs in the normal course of the work, while the Scaling facet forces a question about system-scale interaction that no individual-transaction workflow poses, so the information it elicits would not exist without the format prompting it. The facets are jointly sufficient to carry the constraint-dimension information the price compresses away, and with the necessity of Scaling established here and the necessity of Transmit established in the companion paper, the hexagon is the minimal set that is both necessary in each facet and sufficient in total.
Keywords: Scaling facet, aggregate state, nonlinearity, performative disclosure, sufficiency, six-facet format, OCOT
The hexagon has six facets, and most of them defend themselves. Identity attributes the record to a person who carries the consequences, without which the track record cannot accumulate. Input, Method, and Output carry the three components of an inference, the evidence considered, the procedure applied, and the conclusion reached, and removing any one leaves a record that cannot be evaluated, because an inference with no stated evidence, or no stated method, or no stated conclusion, is not assessable. The disposition facet, Transmit, was shown in a companion paper to be a distinct channel that carries the readership choice and the behavioral information that choice reveals (Beier, 2026, "The Transmit Facet"). Five of the six facets thus have arguments behind them.
The sixth invites a sharper question. The Scaling facet asks what happens when the assessed condition operates at the scale it actually operates at, including its interaction with everything else scaling alongside it, and the architecture treats this facet as the one that distinguishes the format from ordinary analysis. A reviewer is entitled to press: is the Scaling facet a distinct information type, or is it a restatement of the Output facet in different words? If an analyst has concluded something about the constraint in the Output facet, what does the Scaling facet add that the Output did not already contain? This paper answers by proving three things: that the Scaling facet carries information the other facets cannot, that it is the only facet that forces reasoning not otherwise occurring, and that the six facets are jointly sufficient, so the format is neither short a facet nor carrying a redundant one.
The distinction the Scaling facet rests on is the distinction between a transaction and a system. Consider a constraint environment in which a number of agents each act, and let each agent's action bear on a shared constraint. An agent's Output facet states the agent's conclusion about the constraint at the level of that agent's own transaction: the state of the aquifer given the agent's own extraction, the risk of the supply chain given the agent's own order, the exposure of the grid given the agent's own load. Write the transaction-level state the Output assesses as V_i for agent i.
The aggregate constraint state, written theta_S, is the state of the system when all the agents' actions combine. It depends on the agents' transactions through the system's response function, and the response function need not be linear. An aquifer draws down in proportion to total extraction up to a compaction threshold, beyond which the storage capacity itself collapses; a grid absorbs load until a stability margin is crossed, beyond which the failure cascades. The aggregate state theta_S is a function of the individual states V_1 through V_N and the interaction structure that governs how they combine, and the Scaling facet is where an agent states an assessment of that interaction structure: what happens as the condition reaches system scale and meets the other conditions scaling beside it (Beier, 2026ah). The question of Section 1 is whether that assessment carries information the Output facets do not.
The first result establishes that the Scaling facet is informationally independent of the others.
Proposition SF-1 (Information independence). In any constraint environment with nonlinear interaction effects, the Scaling facet carries information about the aggregate constraint state that the other facets do not contain: the mutual information between the aggregate state and the Scaling facet, conditional on Input, Method, Output, Identity, and Transmit, is strictly positive.
Proof sketch. The aggregate state theta_S is a function of the individual states and the interaction structure that governs how they combine. The Output facet carries the agent's conclusion about the individual state V_i, not about the aggregate. When the system's response function is nonlinear, the map from the individual states to the aggregate state is not invertible from the individual states alone, because a nonlinear combination can produce the same aggregate from different individual configurations and different aggregates from the same individual conclusions depending on the interaction structure (Beier, 2026ai, which establishes that constraint domains with power-law scaling have exactly this nonlinearity). Knowing the Output facets therefore leaves residual uncertainty about theta_S, formally H(theta_S given the other facets) is positive. The Scaling facet states the agent's assessment of the interaction structure, which is the information the map from individual to aggregate requires, so conditioning on the Scaling facet reduces the residual uncertainty: H(theta_S given the other facets and Scaling) is strictly less than H(theta_S given the other facets). The reduction is the strict positivity of the conditional mutual information. End sketch.
The result answers the reviewer directly. The Scaling facet is not a restatement of the Output facet, because the Output facet concerns the individual transaction and the Scaling facet concerns the aggregate, and the two coincide only when the combination is linear. Wherever the system's response is nonlinear, which the scaling-geometry result shows is the case for any domain with power-law structure, the aggregate state is not recoverable from the individual conclusions, and the Scaling facet carries the interaction information that recovery requires. The facet that the price system and ordinary transaction-level analysis both omit is exactly the facet that carries the aggregate, and the aggregate is where the constraint binds.
The second result distinguishes the Scaling facet from the others by a different property: it is the only facet that generates reasoning that was not otherwise occurring.
Proposition SF-2 (Performative uniqueness). Among the six facets, the Scaling facet is the only one that elicits reasoning an analyst would not perform in the normal course of the work. The other assessment facets capture reasoning the analyst performs regardless, and the disposition facet records a choice rather than eliciting cognition, so the information the Scaling facet carries would not exist without the format prompting it.
Proof sketch. A disclosure format is performative when it generates information that an unstructured process would not elicit, by prompting reasoning that was not occurring (Beier, 2026ab, Proposition PF-2). Consider the six facets against this property. Input, Method, and Output capture the evidence, procedure, and conclusion of an inference the analyst performs in the ordinary course of the work, because an analyst examining a constraint looks at sources, applies a method, and reaches a conclusion whether or not a format asks for them; the format makes this reasoning legible but does not create it. Identity records who the analyst is, which requires no new reasoning. Transmit records the disposition the analyst selects, which is a choice about readership and not an act of cognition about the constraint. The Scaling facet asks what happens at system scale with everything else scaling alongside, a question the individual-transaction workflow does not pose, because the workflow concerns the transaction and not the aggregate, so an analyst completing ordinary work does not perform the system-scale reasoning the Scaling facet requires. The format's prompt for the Scaling facet therefore elicits reasoning that was not occurring, which is the performative property, and the Scaling facet is the only facet of the six for which this holds. End sketch.
The performative property is why the Scaling facet does more than record. The other assessment facets transcribe an inference the analyst already made, and a record that carried only those facets would capture existing judgment without changing it. The Scaling facet changes what the analyst thinks about, by forcing a question the analyst's normal work leaves unasked, so the act of completing the facet produces an assessment that did not exist before the format demanded it. The information the architecture most wants, the assessment of systemic interaction effects, is not lying around waiting to be transcribed; it is brought into being by the requirement to state it, and the Scaling facet is the requirement.
The third result closes the question from the other side: the facets are enough.
Proposition SF-3 (Sufficiency). The six facets are jointly sufficient to carry the constraint-dimension information that the scalar price compresses away. For any disclosure an agent could produce from the agent's information set, the information that disclosure carries about the constraint dimensions is contained in the six facets.
Proof sketch. A disclosure about a constraint is a function of the agent's information set, and any such disclosure decomposes into the components the facets carry: what the agent examined, which the Input facet holds; how the agent reasoned, which the Method facet holds; what the agent concluded at the transaction level, which the Output facet holds; what the agent assesses at aggregate scale, which the Scaling facet holds; who the agent is, which the Identity facet holds; and to whom the agent discloses, which the Transmit facet holds. Any constraint-relevant content the agent could state falls into one of these components, because a statement about a constraint is a statement about evidence, method, individual conclusion, aggregate effect, source, or disposition, and the six facets exhaust these. The information the agent could disclose is therefore contained in the six facets, which is joint sufficiency. Combined with the independence of Scaling established in Proposition SF-1 and the independence of Transmit established in the companion paper, no facet is redundant and none is missing, so the six are the minimal sufficient set. End sketch.
Sufficiency completes the account of why the format has the facets it has. Each facet carries a component that constraint-relevant content falls into, and the six components exhaust the content an agent could disclose, so the format is not short a facet. Each facet carries a distinct component, established for Scaling here and for Transmit in the companion paper, so the format carries no redundant facet. The hexagon is the set of facets that is necessary in each and sufficient in total, which is the answer to the reviewer's question generalized from Scaling to the whole format: not five facets, not seven, but these six, because these six are the minimal decomposition of what a constraint assessment contains.
The two necessity results, the distinctness of Scaling proved here and the distinctness of Transmit proved in the companion paper, together with the sufficiency of the six, settle the structure of the format against the challenge that it carries the wrong number of facets. The Scaling facet is informationally independent of the others because it carries the aggregate state that the individual conclusions cannot determine under nonlinearity, it is performatively unique because it forces reasoning the normal workflow omits, and the six facets jointly exhaust the constraint content an agent could disclose. A reviewer who asks why Scaling is a facet, or why the format has six and not some other number, has the formal answer.
The account also locates the Scaling facet's standing within the framework's larger claims. The aggregate state it carries is what the sustainability functional aggregates across domains, so the facet whose necessity this paper proves is the facet on which the framework's central quantity depends (Beier, 2026, "Sustainability as an Emergent Aggregate"). The performative property it uniquely holds is what makes the commons a source of new judgment rather than a transcription of existing judgment, which is what the dynamic improvement results compound over time. The facet that the price omits and the workflow leaves unasked is the facet the architecture exists to add, and the proof that it is distinct and necessary is the proof that the addition is not redundant with anything the existing channels carry.
The most dangerous coordination failures occur at the intersections between constraint domains, where aquifer depletion meets grid stability, supply-chain concentration meets geopolitical fragility, and atmospheric loading meets agricultural productivity. The Constraint Coupling paper established the importance of these intersections in words. This paper supplies the dynamics. The coupled constraint state is modeled as a block-structured covariance whose off-diagonal block carries the coupling, and four results follow. When analysts in two domains each treat the other domain's state as a fixed boundary condition rather than as a coupled variable, the information their disclosures jointly recover about the coupled state falls short of the full information by exactly the cross-domain mutual information, the quantity that describes how the domains interact, which is lost whenever the domains are not conditionally independent, a condition every major coupled constraint system violates. The sensitivity of the coupled system to a perturbation in one domain diverges as either domain approaches its own threshold, so a coupled system is most sensitive to cross-domain shocks precisely when it is most stressed, which is the formal content of the intuition that coupled failures are catastrophic. When analysts read each other's cross-domain assessments, carried in the Scaling facets, the recovered coupling information is bounded below by the cross-domain correlation times the weaker of the two disclosures, so the recovery is limited by the less informative side. The order in which domains should be connected to a disclosure network follows from these bounds: the domain to add next is the one whose cross-domain correlations with the existing network, weighted by the network's information, are largest, which maximizes the coupling information gained per domain connected.
Keywords: coupled constraints, cross-domain information, boundary conditions, threshold amplification, bifurcation, disclosure recovery, OCOT
A water system and an energy system are studied by different analysts, in different agencies, with different models. The water analyst treats energy prices as a parameter, an input fixed from outside the water model. The energy analyst treats water availability as a parameter, an input fixed from outside the energy model. Each model is competent within its domain, and each treats the other domain as a boundary condition: a fixed value that bounds the model rather than a variable that moves with it. The arrangement works until the two systems stress at once, at which point the failure arrives through a channel neither model contained, the channel by which water scarcity raises energy demand for pumping, which raises water demand for cooling, which deepens the scarcity. The failure was at the seam between the domains, and the seam was where neither model looked.
The Constraint Coupling paper established that these seams are where the most dangerous and least visible coordination failures occur, and argued the point through cases (Beier, 2026am). What it did not supply is the dynamics: a model of how the coupled state evolves, how much information the boundary-condition treatment destroys, and how much cross-domain disclosure recovers. This paper supplies that model. It represents the coupled state as a covariance with a coupling block, proves that the boundary-condition treatment loses exactly the cross-domain information, proves that the coupled system's sensitivity diverges near either domain's threshold, and proves bounds on what cross-domain disclosure recovers and in what order domains should be connected.
Let two constraint domains carry states V_A and V_B, each a vector of constraint variables, and let the joint state have a covariance with block structure
Sigma = [ Sigma_A, Sigma_AB ; Sigma_BA, Sigma_B ],
where Sigma_A and Sigma_B are the within-domain covariances and Sigma_AB is the cross-domain coupling block, with Sigma_BA its transpose. The coupling block is what the boundary-condition treatment ignores: an analyst who fixes the other domain treats Sigma_AB as zero, modeling the joint covariance as block-diagonal.
The cross-domain mutual information between the two states measures how much the coupling block carries. For the Gaussian case it is a function of the canonical correlations between the domains, and it is zero exactly when Sigma_AB is zero, that is, when the domains are conditionally independent (Cover and Thomas, 2006). The dynamics of the coupled state are governed by a vector field that lets each domain respond to the other, dV_A/dt = F_A(V_A, V_B) and dV_B/dt = F_B(V_A, V_B), with the cross-arguments carrying the coupling that the boundary-condition models drop. The sections that follow work with this representation: the covariance for the information results and the vector field for the sensitivity result.
The first result quantifies what the boundary-condition treatment loses.
Proposition CD-1 (Boundary information loss). When analysts in domains A and B each treat the other domain's state as a fixed boundary condition, the information their disclosures jointly recover about the coupled state falls short of the full coupled information by exactly the cross-domain mutual information between V_A and V_B. The loss is zero only when the domains are conditionally independent, which a coupled constraint system with nonzero Sigma_AB violates.
Proof sketch. The entropy of the joint state decomposes as the entropy of each domain minus the cross-domain mutual information: H(V_A, V_B) equals H(V_A) plus H(V_B) minus I(V_A; V_B). An analyst who treats domain B as a fixed boundary condition discloses about V_A given a fixed value of V_B, so the disclosure D_A informs the within-domain structure H(V_A) but not the dependence of V_A on a moving V_B, and symmetrically for D_B. The joint disclosure under the boundary treatment therefore recovers H(V_A) and H(V_B), the within-domain terms, but not the cross-domain mutual information I(V_A; V_B), because neither disclosure represented the other domain as a variable that moves. The information recovered falls short of the full coupled information H(V_A, V_B) by exactly the omitted term I(V_A; V_B), which is the coupling. This term is zero only when V_A and V_B are conditionally independent, equivalently when the coupling block Sigma_AB is zero, and a coupled constraint system has a nonzero coupling block by definition, so the loss is strictly positive for every such system. End sketch.
The result names the cost of studying coupled domains separately. Two competent single-domain analyses, each treating the other domain as fixed, recover everything about each domain alone and nothing about how the domains interact, and the interaction is the whole of the coupling. The water analysis and the energy analysis, each excellent, together miss exactly the water-energy coupling, because each treated the other as a boundary. The information is not lost to incompetence within either domain. It is lost at the boundary, which neither analysis crossed, and the amount lost is precisely the cross-domain mutual information that the coupling block carries.
The second result concerns the dynamics rather than the information, and shows that coupling is most dangerous under stress.
Proposition CD-2 (Threshold coupling amplification). The sensitivity of the coupled system's steady state to a perturbation in one domain diverges as either domain approaches its own threshold. A coupled system is therefore most sensitive to cross-domain shocks precisely when it is most stressed.
Proof sketch. At a steady state of the coupled dynamics, the state of domain A responds to a change in domain B according to the implicit-function relation derived from F_A(V_A, V_B) equals zero: the steady-state sensitivity is dV_A*/dV_B equals minus the ratio of the cross-domain response to the within-domain response, minus (partial F_A / partial V_B) over (partial F_A / partial V_A). Near a threshold of domain A, the within-domain restoring response vanishes, partial F_A / partial V_A approaches zero, because a threshold is a saddle-node bifurcation at which the restoring force that pulls the domain back to its steady state disappears, the same bifurcation structure the basin-transition analysis identified (Beier, 2026, "Basin Transition Economics"; Strogatz, 1994). As the denominator approaches zero, the sensitivity dV_A*/dV_B diverges, so a perturbation in domain B that the coupled system would absorb far from threshold produces an unbounded response near it. The divergence holds as either domain approaches its threshold, because the cross-domain channel amplifies a perturbation precisely when the receiving domain has lost the restoring capacity that would otherwise damp it. End sketch.
The amplification is why coupled failures are catastrophic rather than merely additive. Far from any threshold, a shock in one domain produces a bounded response in the other, and the coupled system absorbs it. As a domain approaches its threshold, its capacity to absorb a cross-domain shock vanishes, and the same coupling that was harmless becomes the channel through which the stressed domain is pushed over. The water-energy system is stable under normal conditions and unstable under simultaneous stress not because the coupling changed but because the receiving domain's restoring force vanished near its threshold, which turned a bounded cross-domain sensitivity into a divergent one. The coupling is most dangerous exactly when the system can least afford it.
The third result is the remedy: disclosure that crosses the domain boundary recovers the coupling information the boundary treatment lost.
Proposition CD-3 (Cross-domain disclosure recovery). When analysts from domains A and B read each other's cross-domain assessments, carried in the Scaling facets that state systemic interaction effects, the recovered coupling information is bounded below by the cross-domain correlation times the weaker of the two disclosures: I_recovered is at least rho_AB times the minimum of the two sides' information contributions. The recovery is limited by the less informative side.
Proof sketch. The coupling information that the boundary treatment lost is the cross-domain mutual information of Proposition CD-1. Recovering it requires that the disclosures represent the cross-domain dependence, which the Scaling facet does, since the Scaling facet states what happens as the domain's condition operates at scale and meets the other systems scaling beside it, including the other domain (Beier, 2026ah). When an analyst in A reads an analyst in B's Scaling facet and conditions on it, and symmetrically, the recovered information about the coupling is bounded by how much the two disclosures jointly reveal about the cross-domain dependence. That joint revelation is bounded below by the cross-domain correlation rho_AB, which sets how much coupling there is to recover, times the information contribution of the weaker disclosure, because the cross-domain channel requires both sides to disclose and the recovered coupling cannot exceed what the less informative side reveals about its end of the coupling. The bound is therefore rho_AB times the minimum of the two sides' contributions, and the recovery is limited by the weaker side. End sketch.
The result locates the remedy and its limit. Cross-domain disclosure recovers the coupling, but only to the extent that both sides disclose, because a coupling has two ends and the weaker end bounds what can be recovered about it. A rich water disclosure read against a thin energy disclosure recovers only as much of the water-energy coupling as the thin energy side reveals, which is why connecting two domains usefully requires both to disclose substantively and not merely one. The Scaling facet is the channel through which the recovery flows, because it is the facet that carries the cross-domain interaction, and the composition algebra is what combines the two sides' records into a joint assessment of the coupling (Beier, 2026, "The Composition Algebra").
The bounds of the previous sections determine an order of operations for building a cross-domain network.
Proposition CD-4 (Optimal sequencing). The domain to connect next to a disclosure network is the one whose cross-domain correlations with the domains already in the network, weighted by the information those domains disclose, are largest. This ordering maximizes the coupling information gained per domain connected.
Proof sketch. The coupling information gained by connecting a new domain k to the network is the sum over the domains j already connected of the recovered coupling between k and j, which by Proposition CD-3 is bounded by rho_kj times the weaker of the two disclosures. The marginal information from adding k is therefore the sum over j of rho_kj weighted by the disclosure strength of domain j, taken at the level the new domain k can match. To maximize the coupling information gained per domain added, the network should connect next the domain k that maximizes this weighted sum, which is the domain most correlated with the existing network's well-disclosed domains. Iterating the choice gives a greedy ordering that connects domains in decreasing order of their weighted cross-domain correlation with the network built so far, and the greedy choice maximizes the per-step gain. End sketch.
The sequencing matters because a cross-domain network cannot be built everywhere at once, and the order determines how quickly the coupling information accumulates. Connecting a domain weakly correlated with the existing network adds little coupling information, however well the domain discloses internally, because there is little coupling to recover. Connecting a domain strongly correlated with the network's well-disclosed domains adds the most, because the coupling to recover is large and both ends are substantively disclosed. The water-energy-food nexus is the canonical first network for exactly this reason: its three domains are strongly coupled to each other and individually well studied, so connecting them recovers a large coupling per domain, which is why the institutional program identifies that nexus as the first deployment.
The coupling dynamics supply what the qualitative coupling paper and the sustainability aggregate both needed. The boundary-condition treatment loses exactly the cross-domain information, the coupled system amplifies cross-domain shocks near its thresholds, cross-domain disclosure through the Scaling facets recovers the coupling up to the weaker side, and the optimal order of connection follows from the recovery bound. The sustainability functional that aggregates Scaling facets across domains depends on these coupling terms, and the present results give them their dynamics, so the truncation that a missing domain imposes on the sustainability aggregate now has a magnitude: it is the cross-domain mutual information of every coupling that involves the missing domain, the quantity Proposition CD-1 identifies (Beier, 2026, "Sustainability as an Emergent Aggregate").
The model leaves two extensions to further work. The Gaussian covariance representation captures linear coupling, and the threshold amplification of Proposition CD-2 already reaches past it, so a full treatment of nonlinear coupling, where the coupling block itself depends on the state, is the natural generalization. The recovery bound of Proposition CD-3 is a lower bound, and the conditions under which it is tight, where cross-domain disclosure recovers all rather than merely some of the coupling, are the conditions a deployment would want to engineer, which makes the gap between the bound and the achievable recovery the quantity to study next.
Growth accounting attributes output growth to capital, labor, and a residual called total factor productivity, and the residual has been endogenized before, most influentially as innovation. This paper endogenizes a different component of it: coordination quality, the part of aggregate productivity that depends on capital reaching the uses that physical and institutional constraints actually reward rather than the uses a compressed price signal makes look rewarding. The contribution is not the observation that misallocation lowers productivity, which is established, but the connection of a specific mechanism to that misallocation: structured disclosure makes constraint dimensions observable, the combinatorial discovery of constraint interactions through open disclosure grows faster than linearly in the number of disclosers, and the resulting reduction in misallocation enters the productivity residual with increasing returns. The model decomposes productivity into an innovation component and a coordination component, expresses the coordination component through the dispersion of marginal products that misallocation produces, and shows that because the number of tracked constraint interactions grows quadratically in the disclosure network's size, the coordination component is convex in that size, which is increasing returns to disclosure at the level of aggregate output. The model is calibrated against three historical changes in information architecture that each reduced the dispersion of marginal products: mandatory financial reporting, the credit-rating industry, and real-time market data. The calibration yields an illustrative central range for the permanent increase in annual productivity growth once a constraint-disclosure network reaches critical mass, and a corresponding range for the misallocation that each decade of delay allows to accumulate through lock-in. The ranges are rough and the uncertainty is wide, and the paper states both.
Keywords: total factor productivity, misallocation, coordination, combinatorial disclosure, increasing returns, calibration, OCOT
Growth accounting writes output as a function of capital and labor and finds that the inputs do not explain all of the growth. The unexplained part is the residual, total factor productivity, and the history of growth theory is in large part the history of saying what is in it (Solow, 1957). Romer put innovation in it, endogenizing the residual as the accumulation of ideas (Romer, 1990). The residual has room for more than one occupant, and this paper argues that a distinct component of it is coordination quality, the degree to which capital and effort reach the uses that the world actually rewards rather than the uses that a compressed signal makes appear rewarding.
The argument has to be careful about what is new. That misallocation lowers aggregate productivity is established, and the framework for measuring it through the dispersion of marginal products is standard (Hsieh and Klenow, 2009). That coordination matters to output is not a discovery. What this paper adds is a mechanism connected to a measurement: structured disclosure makes constraint dimensions observable, the discovery of constraint interactions through open disclosure grows combinatorially in the number of disclosers, and the reduction in misallocation that follows enters the residual with increasing returns. The model below endogenizes the coordination component, shows it is convex in the disclosure network's size, and calibrates its magnitude against historical episodes in which information architecture changed and misallocation measurably fell. The numbers it produces are rough, and the paper marks them as rough.
Write output in the standard form Y equals A times K to the alpha times L to the one minus alpha, with A the total factor productivity. Decompose the productivity term into two components,
A equals A_innovation times A_coordination,
where A_innovation is the Romer component, the productivity from the stock of ideas, and A_coordination is the component this paper isolates, the productivity from capital reaching its constraint-rewarded uses. The coordination component is a decreasing function of misallocation, and misallocation is measured through the cross-sectional dispersion of marginal products: when capital is allocated efficiently, marginal products are equalized across uses, and when it is misallocated, they disperse (Hsieh and Klenow, 2009). In the constant-elasticity case the coordination component takes the form A_coordination equals the dispersion-adjusted aggregate of firm-level productivities, falling as the dispersion of those productivities rises.
The source of dispersion this paper attends to is the constraint dimensions the price does not carry. Capital flows toward assets that look rewarding on the price signal, and the price signal carries strictly positive excess distortion on the constraint dimensions, so it makes constraint-fragile assets look as rewarding as constraint-robust ones (Beier, 2026h). The result is dispersion in the true marginal product, because capital is allocated as if constraint-fragile and constraint-robust assets were equivalent when they are not, and the dispersion is misallocation that lowers A_coordination. Structured disclosure makes the constraint dimensions observable, which lets capital distinguish the fragile from the robust, which reduces the dispersion and raises the coordination component. The size of the effect depends on how many constraint dimensions disclosure makes observable, which the next section connects to the disclosure network's size.
The reduction in misallocation depends on how many constraint interactions the disclosure network surfaces, and the number of surfaced interactions grows faster than the number of disclosers.
Proposition CG-1 (Convexity of the coordination component). The coordination component of productivity is convex in the size of the disclosure network, so the marginal contribution of an additional discloser to aggregate productivity increases with the network's size. This is increasing returns to disclosure at the level of aggregate output.
Proof sketch. The coordination component A_coordination is an increasing function of the number of tracked constraint interactions, because each tracked interaction is a dimension on which capital can distinguish robust from fragile assets, which reduces the misallocation dispersion (Beier, 2026ae, which connects information architecture quality to productivity through misallocation reduction). The number of constraint interactions that open disclosure surfaces grows quadratically in the number of disclosers, because the combinatorial discovery of interactions through the interaction of disclosers produces at minimum order N-squared discoverable interactions for N disclosers (Beier, 2026ag, the Factorial result). The coordination component is therefore an increasing function of a quantity that grows quadratically in the network size, which makes it convex in the network size under the monotonicity of the dependence, so the second difference of A_coordination in N is positive. The marginal productivity contribution of adding a discloser, the first difference of A_coordination, therefore increases with N, which is increasing returns to disclosure. End sketch.
The convexity is the macroeconomic face of the Factorial. At the level of a single constraint domain, the Factorial says that disclosers interacting produce combinatorially many discovered interactions; at the level of aggregate output, that combinatorial discovery enters productivity through misallocation reduction, and the quadratic growth in discovered interactions makes the productivity contribution convex in the network size. A small disclosure network produces a modest reduction in misallocation, and each added discloser produces a larger reduction than the last, because the interactions a new discloser surfaces multiply against the network already present. The aggregate return to building the network rises as the network grows, which is the property that makes the network worth building past the point where its first-period return is small.
The model's structure says the coordination component rises with disclosure and rises convexly. Its magnitude is an empirical question, and the way to approach it is through historical episodes in which information architecture changed and misallocation measurably fell.
Proposition CG-2 (Calibrated productivity gain). Calibrated against historical changes in information architecture that reduced the dispersion of marginal products, the model implies an illustrative central range of roughly three to eight tenths of a percentage point for the permanent increase in annual productivity growth once a constraint-disclosure network reaches critical mass, with the range reflecting uncertainty in the density of constraint interactions.
Derivation and its limits. Three historical episodes supply the calibration. Mandatory financial reporting, instituted in the United States through the securities legislation of the early 1930s, made firm financials observable and reduced the dispersion of marginal products by letting capital distinguish firms that prior opacity had pooled. The credit-rating industry, developing from the early twentieth century, made default risk observable and reduced dispersion on that dimension. Real-time market data, spreading through electronic terminals from the 1970s and 1980s, made prices observable continuously and reduced dispersion from informational lag. Each episode produced a measurable reduction in marginal-product dispersion of the kind the misallocation framework quantifies (Hsieh and Klenow, 2009). Mapping the dispersion reduction of a comparable shift in constraint-dimension observation through the model's coordination component, and scaling by the density of constraint interactions that determines how much the Factorial surfaces, yields the central range stated. The range is illustrative and the uncertainty is wide: it depends on the constraint-interaction density, which is estimated rather than measured, on the comparability of the historical episodes to constraint-dimension disclosure, which is imperfect, and on the assumption that the dispersion reduction maps to productivity as the framework specifies. The number is offered as an order of magnitude grounded in the historical analogs, not as a precise forecast, and the honest statement is that the true value could fall outside the range if the constraint-interaction density differs substantially from the estimate. End derivation.
The calibration is deliberately modest in what it claims. It does not assert that constraint disclosure will raise productivity growth by a specific amount. It asserts that, if constraint disclosure reduces misallocation the way three historical information-architecture changes did, the implied productivity effect is of the order the range states, with the range carrying real uncertainty about the constraint-interaction density. The value of the estimate is in its grounding, that it is calibrated against episodes where the mechanism operated and the effect was measured, rather than asserted from theory alone, and in its honesty about the width of the band around it.
The convexity and the calibration together imply a cost to delaying the network, because the misallocation that disclosure would reduce continues to accumulate, and some of it locks in.
Proposition CG-3 (Delay cost). Each decade of delay in building a constraint-disclosure network allows additional misallocation to accumulate through lock-in on constraint dimensions, of an illustrative order of magnitude in the low trillions of dollars per decade, consistent with the accumulated-cost analysis of the Reality Debt paper.
Derivation and its limits. Misallocation on constraint dimensions does not merely persist while disclosure is absent; it compounds, because capital allocated to constraint-fragile assets that look rewarding on the price builds sunk infrastructure in those assets, which raises the cost of reallocating later and locks the misallocation in (Beier, 2026af). The accumulated cost of a decade of delay is the foregone improvement in the coordination component over that decade, integrated against output, plus the lock-in that makes a portion of the misallocation costly to reverse afterward. Scaling the calibrated productivity gain of Proposition CG-2 against aggregate output over a decade, and adding the lock-in component, yields an order of magnitude in the low trillions of dollars per decade of delay. The figure inherits all the uncertainty of the productivity calibration and adds the uncertainty of the lock-in rate, so its band is at least as wide, and it is offered on the same terms: an order of magnitude grounded in the model and the historical analogs, not a precise accounting. End derivation.
The delay cost is the quantitative form of the argument for building the network early. The first-period return to the network is small, by the convexity of Proposition CG-1, which might suggest waiting. The delay cost says that waiting is not free, because the misallocation the network would reduce accumulates and partially locks in during the wait, so the cost of a decade of delay is the foregone reduction plus the lock-in, an order of magnitude that the trillions-range figure expresses. The case for early construction is that the network's value compounds once built and the cost of not building it compounds while it is absent, and both compounding effects are large enough to matter at the scale of aggregate output.
The calibrated growth model connects the disclosure architecture to aggregate productivity through a mechanism and a measurement: structured disclosure reduces misallocation on constraint dimensions, the combinatorial discovery of interactions makes the productivity contribution convex in the network's size, and historical calibration places the magnitude at an order the paper states with wide bands. The model supports the claim that constraint disclosure has a macroeconomic productivity effect, that the effect has increasing returns, and that delay carries an accumulating cost. It supports these claims at the level of order-of-magnitude calibration, not precise forecast.
The model does not support a precise number, and the paper has marked the numbers as illustrative throughout. The constraint-interaction density that drives the magnitude is estimated rather than measured, and a live disclosure network would measure it, which is the first thing a deployment would pin down to replace the calibrated range with an observed one. The comparability of the historical information-architecture episodes to constraint-dimension disclosure is imperfect, and a closer study of each episode's dispersion reduction would tighten the mapping. The honest summary is that the model gives the shape of the effect with confidence, increasing returns and an accumulating delay cost, and the size of the effect only to an order of magnitude, and that closing the gap between the shape and a precise size is empirical work the network itself would do.
The adversarial analysis of the disclosure mechanism identified four categories of deliberate attack and argued that cross-disciplinary detection covers them with no shared blind spot. A mechanism can also fail without any adversary, from its own dynamics under realistic conditions, and the architecture's honesty about its limits requires naming these failures as plainly as it names the attacks. This paper analyzes four. Ontology drift occurs when the categories under which disclosures are filed fall out of alignment with the evolving structure of the constraint space, which makes disclosures filed under a category non-comparable across time and degrades the validity of the track record built on it, with the defense a versioned ontology and the early warning a rising within-category outcome variance. Scoring lag occurs when a constraint dimension resolves so slowly that the track record cannot reach useful length before the constraint binds, which is the architecture's weakest case and admits only partial defenses through intermediate observables, cross-dimension transfer, and the engagement signal that carries information where the track record is silent. Platform fragmentation occurs when incompatible disclosure platforms partition the disclosers into non-composing groups, which fragments the combinatorial advantage by a factor equal to the number of incompatible platforms, with the defense a shared open specification that lets records compose across implementations. Quality collapse occurs when the calibration threshold is set too low or the coverage weight too high, which floods the commons with low-value disclosures until the value of reading falls below the cost of producing and high-quality producers withdraw, with the defense a parameter regime that keeps reading worthwhile. For each failure the paper states the mechanism, the conditions under which it dominates, the early warning, and the defense or the modification the architecture requires.
Keywords: failure modes, ontology drift, scoring lag, platform fragmentation, quality collapse, robustness, OCOT
The disclosure mechanism attracts adversaries, and the adversarial paper took them seriously: Sybil attacks, coordinated poisoning, hollow disclosures generated at scale, and strategic selection of disclosure domains, each with a detection signature and a cross-disciplinary defense that leaves no shared blind spot (Beier, 2026an). An account that stopped there would be incomplete in a way that matters for the architecture's credibility, because a mechanism can fail under realistic conditions with no adversary present, from the drift of its own categories, the slowness of its own feedback, the fragmentation of its own platforms, and the collapse of its own quality under mis-set parameters. The same honesty that names the attacks has to name these.
This paper analyzes four non-adversarial failure modes. Each is a way the mechanism degrades from its internal dynamics rather than from an attack, and for each the paper specifies four things: the formal mechanism through which the failure operates, the conditions under which it dominates, the early warning that signals its onset, and the architectural defense if one exists within the current framework or the modification the framework requires if one does not. The point is not to argue that the mechanism is fragile. It is to locate the conditions under which it degrades, so that a deployment can watch for them and the architecture can carry its limits openly rather than discovering them in the field.
A disclosure is filed under a category, and the category carries an assumption: that disclosures filed under it are comparable, so that scoring one against another is valid. The assumption can fail as the constraint environment evolves.
Proposition NF-1 (Ontology drift). When the categories under which disclosures are filed fall out of alignment with the structure of the constraint space, disclosures filed under a category become non-comparable across time, and the track record built on that category loses scoring validity. The failure dominates in domains where the constraint structure evolves faster than the ontology. The early warning is a rising within-category variance in disclosure outcomes, and the defense is a versioned ontology with explicit mappings across versions.
Mechanism and defense. A category groups disclosures that are assumed to assess comparable phenomena, and a proper score compares a disclosure against the outcome on the assumption that other disclosures in the category are forecasting the same kind of thing. When a category spans phenomena that are in fact different, with different resolution timescales and different interaction structures, scoring across them is invalid, because the comparison treats forecasts of distinct quantities as forecasts of one. As the constraint environment evolves, a category that was once homogeneous can fragment: disclosures about a single labeled risk come to span physically distinct phenomena that the label no longer separates. The track record built on the drifting category aggregates scores that are no longer comparable, so it loses validity as a calibration signal. The early warning is observable: a category whose disclosures resolve with rising variance over time is a category that is splitting, because increasing outcome heterogeneity within a fixed label is the signature of a label that now spans distinct phenomena. The defense is to version the ontology, treating the category structure as a revisable object with explicit mappings between versions, so that disclosures filed under an old category can be re-sorted into the homogeneous subcategories that replace it, and the track record can be recomputed within the comparable groups (Beier, 2026aq, which specifies the constraint-domain ontology as a required component of the infrastructure). The emergence detection that the composition algebra performs depends on the ontology supplying comparable categories, so ontology drift also degrades that detection, which is a second reason the ontology must be maintained as the constraint space evolves (Beier, 2026, "The Composition Algebra," Proposition C-2). End.
The track record works by accumulating scored outcomes, and a dimension that resolves slowly accumulates them slowly. On the slowest dimensions, the accumulation can be too slow to be useful in time.
Proposition NF-2 (Scoring lag). When a constraint dimension resolves so slowly that the track record cannot reach useful length before the constraint binds, the mechanism cannot distinguish calibrated from uncalibrated analysts within the window of decision relevance. The failure dominates on long-resolution, irreversible dimensions such as climate trajectories, aquifer depletion, and institutional erosion. The defenses are partial.
Mechanism and partial defenses. The calibration threshold tightens as the track record lengthens, so a longer record distinguishes calibration more sharply (Beier, 2026v). On a dimension that resolves over decades, the record accumulates resolved outcomes once per resolution cycle, so it may not reach the length that sharp distinction requires before the constraint binds, which leaves the mechanism unable to tell the calibrated analyst from the lucky one in the window that matters. This is the architecture's weakest case, and the paper does not claim a full solution. Three partial defenses operate. Intermediate observables decompose a long-horizon claim into nearer-term checkpoints that resolve sooner, using the Scaling facet's trajectory to state intermediate predictions that can be scored before the terminal outcome, which shortens the effective resolution at the cost of scoring proxies rather than the final state. Cross-dimension transfer uses calibration on faster-resolving related dimensions as a partial predictor of calibration on the slow one, which helps to the extent the dimensions share structure and fails to the extent they do not, so the transfer is an assumption that itself carries uncertainty. The engagement signal carries claim-specific information about a long-horizon assessment's robustness where the track record is silent, which is exactly the long-resolution case the engagement paper identified as where engagement is most informative (Beier, 2026, "The Engagement Level," Proposition E-2). None of the three fully resolves the failure, and the honest statement is that on irreversible decadal dimensions the track-record discipline is at its weakest, which is the same boundary the threshold and survival analyses reach from their own directions: where outcomes resolve too slowly to score in time, the mechanism's internal correction is incomplete and external judgment must carry more of the weight. End.
The combinatorial advantage of the mechanism requires that disclosers read and build on one another's records. Incompatible platforms break the requirement.
Proposition NF-3 (Platform fragmentation). When incompatible disclosure platforms partition the disclosers into groups that cannot compose one another's records, the combinatorial advantage fragments by a factor equal to the number of incompatible platforms. The failure dominates when proprietary platforms with closed formats capture segments of the discloser population. The early warning is a rising share of disclosures in non-interoperable formats, and the defense is a shared open specification that lets records compose across implementations.
Mechanism and defense. The combinatorial discovery of constraint interactions requires disclosers to interact, reading and composing one another's records, which produces on the order of N-squared discovered interactions for N disclosers (Beier, 2026ag). When incompatible platforms partition N disclosers into a number of groups that cannot read across the partition, the interactions are confined within groups, and the count falls. For k equal incompatible platforms, each holds N over k disclosers and produces on the order of N over k squared interactions, so the total across platforms is on the order of N-squared over k, a factor-of-k loss against the N-squared the population would produce if it composed globally. The early warning is the rising share of disclosures locked in formats that do not interoperate, which is the leading indicator of the partition forming. The defense is structural and already in the architecture: the protocol is a specification, plain text with an ORCID binding and open storage, and any implementation that meets the specification interoperates with any other, so beierOS is one implementation rather than the platform, and a record produced under one conforming implementation composes with a record produced under another (Beier, 2026, "The Transmit Facet," Proposition D-4, on the anti-enclosure requirement). The composition algebra operates on records regardless of which implementation produced them, carrying the interface marker as metadata rather than as a barrier, so a population that discloses to a shared open specification retains the global combinatorial advantage even across many implementations (Beier, 2026, "The Composition Algebra"). The defense holds only as long as the open specification is the standard, which is why the architecture treats interoperability as a requirement and not a preference. End.
The mechanism's value to a reader depends on the signal-to-noise of the commons, and that ratio depends on parameters the mechanism sets. Set wrong, the parameters can collapse the quality.
Proposition NF-4 (Quality collapse). When the calibration threshold is set too low or the coverage weight too high, low-quality disclosures flood the commons, the value of reading falls below the cost of producing, and high-quality producers withdraw, which is a lemons trap. The failure dominates when the parameters are tuned for volume over quality. The early warning is a simultaneous decline in average disclosure quality and in readership, and the defense is a parameter regime that keeps reading worthwhile.
Mechanism and defense. Truth-telling is the dominant strategy under proper scoring for agents above a calibration threshold, and agents below it can find strategic or low-effort behavior profitable (Beier, 2026r). If the threshold is set too low, sub-threshold agents disclose profitably, and the commons fills with low-quality disclosures. If the coverage weight that rewards breadth is set too high, breadth is rewarded over accuracy, which invites shallow wide coverage that adds volume without value. Either mis-setting lowers the signal-to-noise of the commons, and a reader facing a noisy commons reads less, because the expected value of reading a disclosure falls as the share of low-value disclosures rises. As reading falls, the value of disclosing falls with it, since a disclosure read by fewer readers builds less track record and reaches less of the audience, and high-quality producers, who bear a real cost to produce, withdraw when the return drops below the cost, which is the lemons unraveling that the substrate paper analyzed for a different cause (Beier, 2026, "The Substrate Marker," Proposition S-3; Akerlof, 1970). The early warning is the joint decline of average quality and readership, the two moving together as the spiral begins. The defense is parametric: the calibration threshold and the coverage weight are policy levers, and keeping the threshold high enough that sub-threshold agents do not disclose profitably and the coverage weight low enough that accuracy is not sacrificed to breadth keeps the value of reading above the cost of producing for high-quality producers, which the incentive bounds parameterize (Beier, 2026r). The engagement-calibration discipline reinforces the defense by letting readers discount producers whose claimed depth does not predict accuracy, which raises the cost of flooding with shallow disclosures (Beier, 2026, "The Engagement Level," Proposition E-4). End.
The four non-adversarial failures complete the picture the adversarial paper began. The mechanism degrades, with no adversary present, when its categories drift, when its feedback is too slow, when its platforms fragment, and when its parameters are mis-set, and for three of the four the architecture carries a defense: a versioned ontology against drift, a shared open specification against fragmentation, and a parameter regime against quality collapse, each with an observable early warning. The fourth, scoring lag, is the architecture's weakest case, and the paper has not pretended otherwise: on irreversible long-resolution dimensions the track record cannot reach useful length in time, the defenses are partial, and external judgment must carry more of the weight there than the mechanism's internal correction can.
The analysis serves the architecture's credibility more than its design. A mechanism presented only with its strengths invites the suspicion that its failures were not examined, and the failures examined here are the ones a careful adopter would find, so naming them and stating their defenses and their limits is what lets the adopter trust the parts that work. The early-warning signals are the practical payoff: a live deployment can watch within-category variance for drift, the share of non-interoperable formats for fragmentation, and the joint movement of quality and readership for collapse, and can act before the failure compounds. The scoring-lag limit is the one to carry most openly, because it is the one no defense fully closes, and the domains where it bites are the long-horizon irreversible ones the architecture most wants to serve.
The Constraint Weight Geometry paper derives that the structure of a physical process leaves a measurable trace in the Fisher geometry of a model trained on it, and specifies a four-phase program to test the derivation. A specification is not a protocol. This paper converts the program into a pre-registered protocol, fixing the hypotheses, the data-generation procedures, the model architectures, the test statistics, the significance levels, the power, and the decision rules before any data is collected, which is the standard that separates a confirmatory test from an exploratory search through results. Three methodological hazards that the original specification leaves open are addressed directly. The overparameterization hazard, that an overparameterized model carries near-zero Fisher eigenvalues for reasons unrelated to any conservation law, is handled by testing a specific conservation direction against random control directions rather than by counting small eigenvalues, so the test asks whether the conservation direction is flatter than chance rather than whether flat directions exist. The multiple-testing hazard, that testing many geometric features inflates the chance of a spurious finding, is handled by controlling the false discovery rate across the features tested in the multi-feature phase. The architecture-confound hazard, that a geometric feature might reflect the model architecture rather than the process the model was trained on, is handled by requiring a feature to appear across several architectures before it is credited to the process. The four phases proceed from process-trained models through multiple processes and multiple features to language-trained models, and the final phase tests the prediction that the geometric trace survives in a text-trained model only when the training text carries the conservation densely enough, which is the prediction the language channel supplies.
Keywords: pre-registration, Fisher geometry, conservation laws, false discovery rate, architecture invariance, confirmatory test, OCOT
The Constraint Weight Geometry paper makes a testable claim and outlines how to test it (Beier, 2026as). The claim is that a conservation law in a data-generating process leaves the trained model flat along the weight-space direction that would encode a violation, that scaling exponents and propagation kernels leave their own traces, and that these traces can be read. The outline names four phases and gives enough detail for the first to be executed. What it does not do is fix the choices that determine whether the test is confirmatory or exploratory: the exact hypotheses, the data and model specifics, the statistics and their thresholds, and the rules for deciding whether each phase succeeded.
Fixing these before data collection is the difference between a test and a search. A program that chooses its statistics after seeing the data can find a significant result by trying enough of them, and the result carries little weight, because the search through analyses was itself a degree of freedom (Nosek et al., 2018). A program that commits to its hypotheses, statistics, thresholds, and decision rules in advance cannot mine the data for a favorable analysis, so a result that clears the pre-registered bar is evidence rather than artifact. This paper writes the commitment. It states the hypotheses, specifies the data generation and the architectures, fixes the test statistics with their significance levels and power, and sets the decision rules for proceeding from each phase to the next, and it addresses three methodological hazards that would otherwise undermine the test before the phases describe themselves.
The program tests four hypotheses, one principal hypothesis per phase, each a restatement of a Constraint Weight Geometry claim in testable form. The first is that a model trained on a process obeying a conservation law is flatter in Fisher geometry along the conservation-violating direction than along control directions. The second is that this flatness holds across multiple processes and multiple conservation laws, and that scaling exponents and propagation kernels leave their own measurable traces, so the correspondence is a property of constrained processes in general and not of one process. The third is that the full set of geometric features the theory predicts appears jointly in a single trained model, tested together under control of the false discovery rate. The fourth is that a model trained on text about a process carries the geometric trace when the text carries the conservation densely enough and does not when it does not, which is the language channel's prediction (Beier, 2026, "The Language Mediation Channel," Proposition LM-2).
Each hypothesis is paired with a pre-specified null. The null for the first is that the conservation direction is no flatter than a random control direction, and the test rejects the null when the conservation direction's Fisher quadratic form falls below the control distribution at the pre-set significance level. The nulls for the later phases are the corresponding statements that the predicted feature is absent, that the features do not appear jointly above the false-discovery threshold, and that the language-trained model's trace does not depend on the conservation density of its training text. Stating the nulls fixes what counts as a negative result, which a confirmatory program must specify as carefully as it specifies success.
Before the phases, three hazards that could produce a false positive or a false negative are addressed, because the design choices that handle them are pre-registration commitments that bind all four phases.
The overparameterization hazard is the first. An overparameterized model has many near-zero Fisher eigenvalues for reasons that have nothing to do with any conservation law, simply because the parameter count exceeds what the task constrains, so a test that looked for flat directions and counted them would find flatness everywhere and credit it to conservation falsely. The protocol does not count eigenvalues. It identifies the specific weight-space direction that encodes a violation of the conservation law in question, the conservation direction, using a constraint direction probe that maps the physical conservation to a direction in the model's parameter space, and it tests whether the Fisher quadratic form along that specific direction is smaller than the form along random control directions drawn from the same space. The test asks whether the conservation direction is flatter than chance, not whether flat directions exist, so the generic flatness of an overparameterized model is the control against which the conservation direction is compared rather than a confound that inflates the signal (Beier, 2026x, which supplies the noise-regime benchmark for the expected effect size).
The multiple-testing hazard is the second. The third phase tests many geometric features at once, several conservation laws, several scaling exponents, several kernels, and testing many features at a fixed per-test significance level would produce spurious findings in proportion to the number of features, because some features would clear the threshold by chance. The protocol controls the false discovery rate across the features tested in that phase using the Benjamini-Hochberg procedure at a pre-set level, which bounds the expected proportion of false findings among those declared significant rather than the per-test error, and is the appropriate control when many related hypotheses are tested together (Benjamini and Hochberg, 1995). The level is fixed in advance, so the multiplicity correction is not a choice made after seeing which features cleared an uncorrected threshold.
The architecture-confound hazard is the third. A geometric feature found in one trained model might reflect the model's architecture, the particular depth, width, and attention structure, rather than the process the model was trained on, in which case it would not be evidence for the correspondence the program tests. The protocol requires a feature to appear across several architectures before crediting it to the process, applying the multi-architecture invariance test that the symbolic-time-crystal detection protocol specifies: a feature that reflects the data-generating process should appear across architectures trained on the same process, while a feature that reflects the architecture should not survive the change of architecture (Beier, 2026ac; Beier, 2026at). A feature is credited to the process only when it is invariant across the architecture set, which separates the trace of the physics from the inductive bias of the model.
The first phase trains models on numerical output from simulations of a process obeying a single known conservation law. The data generation fixes the simulation, the integration scheme, the time span long enough for the conservation to manifest, the random seeds, and the train, validation, and held-out test splits, all recorded before training. The architecture is a transformer of pre-specified depth, width, and attention-head count, chosen to be large enough to fit the process and recorded with its rationale. The hyperparameters, the learning-rate schedule, the batch size, and the convergence criterion that defines when training stops, are fixed in advance. The test statistic is the Fisher quadratic form along the conservation direction relative to the control distribution, the power to detect the effect size that the noise-regime benchmark predicts is set to a pre-specified level at the chosen significance, and the decision rule is that the phase succeeds, and the program proceeds, when the conservation direction is significantly flatter than control across the pre-set number of seeds.
The second phase repeats the first across several processes and several conservation laws, and adds the tests for scaling exponents and propagation kernels, each with its own pre-specified statistic and threshold. The data generation spans the process set fixed in advance, the architectures are held to the same specification across processes so that the process is the only thing varying, and the decision rule is that the phase succeeds when the conservation flatness reproduces across the process set and the scaling and kernel traces clear their thresholds, which establishes the correspondence as a property of constrained processes rather than of one simulation.
The third phase tests the full predicted feature set jointly in a single trained model, under the false-discovery-rate control of Section 3 and the architecture-invariance requirement applied to each feature. The success criterion is pre-set: the phase succeeds when the proportion of predicted features that appear jointly, after false-discovery correction and the architecture-invariance filter, exceeds the threshold fixed in advance. This phase is where the multiplicity and architecture safeguards do their work, because it is the phase that tests many features at once and is therefore the phase most exposed to spurious joint findings.
The fourth phase trains models on text about a process rather than on the process data, and tests the language channel's prediction. The training corpora are constructed with measured conservation-relevant content density, the quantity the language channel identifies as governing whether the trace survives, and the corpora span a range of that density from low to high (Beier, 2026, "The Language Mediation Channel," Proposition LM-1). The test statistic is the same conservation-direction flatness, computed now in the language-trained models, and the decision rule tests the channel's specific prediction: the trace is detectable in the language-trained model when the corpus density clears the threshold and is absent when it does not, so the success criterion is not merely that the trace appears but that its appearance tracks the corpus density as the channel predicts. A flat trace at all densities, or a trace independent of density, would fail the phase even if some models showed the trace, because the channel's claim is about the dependence and not merely the presence.
The protocol commits, before any data is collected, to the four hypotheses and their nulls, the data-generation procedures and seeds, the architectures and their rationales, the hyperparameters and convergence criteria, the test statistics, the significance levels and power, the false-discovery and architecture-invariance safeguards, and the decision rules that determine success and progression for each phase. The commitment is the content of the pre-registration, and its function is to remove the analytic degrees of freedom that would let a program find a favorable result by searching, so that a result clearing the pre-registered bar is evidence for the correspondence and a result failing it is evidence against.
The protocol also fixes what a negative result would mean, which a confirmatory program must do as carefully as it fixes success. A failure of the first phase would mean the conservation trace is not present even in process-trained models, which would undercut the correspondence at its foundation. A failure of the fourth phase, with the earlier phases succeeding, would mean the trace is real in process-trained models but does not survive the passage through text, which would bound the correspondence's application to the language models the architecture most wants to reach and would send the language channel's preservation conditions back for revision. Stating these outcomes in advance is what lets the program learn from a negative result rather than explaining it away, and it is the discipline that turns the Constraint Weight Geometry claim from a derivation into a falsifiable prediction with a protocol that can falsify it.
Climate economics has a long disagreement about the rate at which to discount the welfare of future generations, with one position favoring a rate close to the market return and another favoring a rate close to zero on the ground that a future person's welfare counts as much as a present one's. The disagreement is real and this paper does not adjudicate it. It identifies a condition that is prior to it. Discounting a future cost presupposes that the cost can be seen, because a cost has to be known before any rate can be applied to it, and the costs that fall on constraint dimensions the price compresses away are not seen at all: the scalar carries no information about them, so they are not small after discounting but absent before it. The paper formalizes this as a consent structure. A party can consent to bearing a cost only if the cost is observable when the decision is made, and a future generation cannot consent to bearing a constraint-dimension cost that no signal at the time of the decision carried, which makes the cost-shifting non-consensual by the structure of the information rather than only by the absence of the future party. The accumulated cost of decisions made on the compressed signal, the reality debt, is in these terms the accumulated non-consensual intergenerational transfer. Disclosure changes the structure by making the constraint-dimension cost observable, which is the condition any discount rate presupposes, so disclosure resolves the information dimension of the problem without resolving the valuation dimension: it puts the cost on the table, and the discount-rate debate then operates on a complete table rather than a truncated one. Observability is necessary for the future to be represented in the decision and is not sufficient, and the paper marks the boundary.
Keywords: discounting, intergenerational equity, consent, observability, reality debt, constraint dimensions, OCOT
The economics of long-horizon problems turns in large part on a single parameter, the rate at which future welfare is discounted to present terms. One tradition sets the rate near the observed return on capital, on the reasoning that future costs should be weighed as the market weighs them, and concludes that distant costs warrant modest present action (Nordhaus, 2007). Another sets the rate near zero, on the reasoning that the welfare of a person born later counts as much as the welfare of a person alive now, and concludes that distant costs warrant large present action (Stern, 2007). The disagreement is genuine, it is consequential, and it is not the subject of this paper, which takes no side on the rate.
The paper is about a condition the debate presupposes without stating. Applying any discount rate to a future cost requires that the cost be known, because the discounting operation maps a known future magnitude to a present one, and an unknown magnitude has nothing for the rate to act on. The costs that fall on the constraint dimensions a scalar price compresses away are not known at the time the decisions that produce them are made, because the price carries no information about those dimensions. They are not future costs that a high rate shrinks to insignificance. They are costs that no signal at the moment of decision represented at all, so the discount-rate debate, whatever it concludes, is a debate about the costs the price does carry and is silent about the costs it does not. This paper formalizes the silence, casts it as a failure of consent rather than only of valuation, and shows what disclosure changes and what it does not.
The first result locates the problem before the discount rate, in the structure of what can be agreed to.
Proposition CE-1 (Consent requires observability). A party can consent to bearing a cost only if the cost is observable when the decision that imposes it is made. A constraint-dimension cost that the price compresses away is not observable at the time of decision, so no party, present or future, can consent to bearing it. The cost-shifting onto a future generation is therefore non-consensual by the structure of the information, independent of the discount rate.
Proof sketch. Consent to bearing a cost is meaningful only when the consenting party knows the cost, since agreement to an unknown burden is not agreement to that burden but to something else. The price signal carries strictly positive excess distortion on the constraint dimensions, which is to say it carries no information about the costs that fall on them (Beier, 2026h). A constraint-dimension cost is therefore unknown at the time a decision is made on the price, and an unknown cost cannot be the object of consent. This holds for every party to the decision, not only the absent future one: even a present party cannot consent to a constraint-dimension cost the signal did not carry, because that party cannot see it either. The non-consensual character of the cost-shifting thus follows from the information structure, that the cost is unobservable in the signal the decision used, and not only from the familiar fact that the future generation was not present to object. The discount rate does not enter the argument, because the argument is about whether the cost can be known, which is prior to how a known cost is weighted. End sketch.
The result reframes the intergenerational problem. The standard concern is that future generations are not present to bargain, which is a problem of representation. The deeper problem the proposition identifies is that the constraint-dimension cost is not observable to anyone at the time of decision, which is a problem of information, and it would persist even if the future generation could somehow be represented, because there would be nothing for the representative to see. A cost that no signal carries cannot be priced into any bargain, present or intergenerational, and the price system's compression is what keeps the constraint-dimension cost out of every signal a bargain could use.
The second result places the discount-rate debate in relation to the consent structure.
Proposition CE-2 (The information floor). Discounting a future cost presupposes the cost is observable, so the choice of discount rate operates only on the costs the price carries and is silent on the constraint-dimension costs it does not. No choice of rate addresses a cost that is absent from the signal, and disclosure, by making the constraint-dimension cost observable, supplies the condition any rate presupposes.
Proof sketch. The discounting operation maps a future cost of known magnitude to a present value by applying the rate over the horizon, which requires the magnitude as input (Ramsey, 1928). For a constraint-dimension cost, the magnitude is not known at the time of decision, by Proposition CE-1, so the discounting operation has no input to act on and is undefined for that cost. The debate over the rate is therefore a debate about the costs whose magnitudes the signal supplies, and it is orthogonal to the constraint-dimension costs, which are lost from the calculation regardless of whether the rate is high or low. The constraint-dimension costs do not become small under a high rate and large under a low one; they are absent under both, because the signal carried nothing for either rate to discount. Disclosure changes this by carrying the constraint-dimension cost into an observable signal, which supplies the magnitude that discounting requires, so after disclosure the rate has an input to act on where before it had none. The information floor is the condition that a cost be observable before any rate applies to it, and disclosure is what lifts the constraint-dimension costs above that floor. End sketch.
The proposition says the discount-rate debate is well-posed only above the information floor. The debate assumes a complete ledger of future costs and argues about how to weight it, and the assumption fails for the constraint-dimension costs, which are not on the ledger at all. A high rate and a low rate disagree about how much to act on the costs the price shows, and they agree, by omission, in ignoring the costs the price hides. Disclosure does not settle the disagreement, and the paper does not claim it does. It completes the ledger the disagreement is about, which is a different and prior contribution: the rate can be argued about once the costs are visible, and the constraint-dimension costs were not visible until disclosure carried them.
The third result connects the consent structure to the accumulated-cost analysis of the corpus.
Proposition CE-3 (Reality debt as non-consensual transfer). The accumulated cost of decisions made on the compressed signal, the reality debt, is the accumulated non-consensual intergenerational transfer. Each decision made on the scalar imposes an unobserved constraint-dimension cost that the future bears without having consented, and the sum over decisions is the reality debt.
Proof sketch. Each decision made on the price signal imposes whatever constraint-dimension cost the decision generates, and by Proposition CE-1 that cost is non-consensual, because it was unobservable at the time of decision and so could not be agreed to by any party who would bear it. The cost falls on the future to the extent that constraint dimensions resolve over long horizons, which the long-resolution dimensions of climate, aquifer, and institutional erosion do. The accumulation of these costs across decisions is the reality debt, the accumulated cost of decisions made on compressed information that the corpus identifies (Beier, 2026af), and in the present terms each component of that debt is a non-consensual transfer to the future, so the reality debt is the accumulated non-consensual intergenerational transfer. Disclosure makes each constraint-dimension cost observable at the time of decision, which converts the cost from one no party could consent to into one a present party can see and weigh, and is the condition under which the future's interest can enter the decision at all. End sketch.
The result gives the reality debt a moral structure to go with its economic one. The debt is not only an accumulated inefficiency, capital allocated on a signal that hid the constraint-dimension costs; it is an accumulated transfer to parties who could not consent because the cost was invisible when the transfer was made. The two descriptions are the same quantity seen from two sides, the inefficiency from the allocation side and the non-consensual transfer from the intergenerational side, and disclosure addresses both by making the hidden cost observable, which both improves the allocation and lets the future's burden be seen by the present that imposes it.
The fourth result marks the boundary of what disclosure resolves.
Proposition CE-4 (Necessary, not sufficient). Making the constraint-dimension cost observable is necessary for the future to be represented in the decision and is not sufficient. A present generation that observes the cost can still choose to weight it lightly, so disclosure resolves the information dimension of the intergenerational problem and leaves the valuation dimension, the discount-rate question, operating now on an observable cost rather than an absent one.
Proof sketch. Observability is necessary by Proposition CE-1, since an unobservable cost cannot be represented in a decision at all. It is not sufficient, because a present generation that observes a future constraint-dimension cost retains the choice of how heavily to weight it, and may choose, under a high discount rate or a narrow conception of its obligations, to weight it lightly. Disclosure therefore determines the information set from which the present chooses and does not determine the choice. The valuation question, how heavily to weight an observable future cost, is the discount-rate question, and it returns after disclosure, now operating on a constraint-dimension cost that is on the ledger rather than absent from it. Disclosure resolves the prior question of whether the cost can be seen and hands the posterior question of how to weight it to the debate that was always about that, with the difference that the debate is now well-posed for the constraint-dimension costs that it could not previously reach. End sketch.
The boundary is the honest limit of the contribution. Disclosure does not make a present generation care about the future, and the paper does not claim it does. It makes the future's burden visible to a present generation that may or may not choose to act on it, which is the necessary condition for the choice to be made on complete information and is not the choice itself. The discount-rate debate, with all its unresolved disagreement, is what governs the choice, and disclosure's contribution is to put the constraint-dimension costs in front of that debate, where they were absent before. The information dimension and the valuation dimension are distinct, disclosure resolves the first, and the second remains exactly as contested as it was, now argued over a fuller account of what is at stake.
The consent analysis locates the disclosure architecture's bearing on the intergenerational problem precisely, neither overclaiming nor understating it. Disclosure does not settle the discount-rate debate, does not make the present generation weigh the future as the future would weigh itself, and does not by itself produce intergenerational justice. What it does is lift the constraint-dimension costs above the information floor that any discounting presupposes, converting them from costs no party could consent to because no signal carried them into costs a present party can observe and a debate can weigh. The reality debt, seen through this lens, is the accumulated transfer that the compression made non-consensual, and disclosure is the condition under which the transfer can be seen by those who make it.
The analysis suggests where the pension-fund case and the climate case meet. A fund with liabilities stretching decades into the future is, in the consent terms, a present party deciding under a signal that hides the constraint-dimension costs its own beneficiaries will bear, so the fund's interest in disclosure is not only the allocational one of seeing constraint risk but the representational one of seeing the costs its long-horizon obligations carry. The information floor is the same for the fund deciding on behalf of future beneficiaries and for the generation deciding on behalf of its successors, and disclosure lifts the costs above it in both cases, which is why the long-horizon investor and the intergenerational problem are the same problem seen at two scales.
The disclosure architecture works by forcing attention to constraint dimensions at the moment of decision, with the Scaling facet in particular requiring an analyst to attend to a systemic question the ordinary workflow leaves unasked. Attention at the decision point is finite, and this paper examines what the finiteness implies when adoption is partial. A disclosure requirement that covers some constraint dimensions but not others draws attention toward the covered dimensions, and because attention is bounded, the attention drawn to the covered dimensions is withdrawn from the uncovered ones, which is a negative externality that partial adoption imposes on the dimensions it does not yet reach. The externality has the structure of a Goodhart reallocation: the covered dimensions become the de facto target, observation bandwidth flows toward them, and the uncovered dimensions are starved exactly as bandwidth is starved when agents optimize on a scalar. The result is not a permanent feature but a transitional cost, because the externality falls only on uncovered dimensions, and as coverage expands the set of uncovered dimensions shrinks, so the externality decreases monotonically in coverage and vanishes at full coverage. The cost of the transition can be minimized by the order in which dimensions are brought under the requirement, and the order that minimizes the cumulative starvation prioritizes the dimensions whose neglect would be most costly. This sequencing objective can pull against the information-gain sequencing that the coupling analysis recommends, and the paper marks the tension rather than resolving it, because the rollout has two objectives that do not always align.
Keywords: attention, partial adoption, Goodhart reallocation, coverage, sequencing, transition cost, OCOT
The architecture's mechanism is in large part a mechanism for directing attention. The six-facet format requires an analyst to attend to the evidence, the method, the conclusion, the systemic effect, the source, and the disposition, and the Scaling facet in particular forces attention to a question the analyst's ordinary work does not pose, the question of what happens at system scale, which is why the Scaling facet was shown to be the uniquely performative one, generating reasoning that was not otherwise occurring (Beier, 2026, "The Scaling Facet as a Distinct Information Type," Proposition SF-2). The architecture surfaces constraint-dimension reality by making analysts look at it, and looking is the scarce input.
Attention at the moment of decision is finite, a point established since the recognition that a wealth of information produces a poverty of attention (Simon, 1971; Kahneman, 2011). The finiteness has a benign implication under full adoption, where every constraint dimension is disclosed and the format directs attention across all of them. It has a less benign implication under partial adoption, where only some dimensions are brought under the disclosure requirement, because the attention the requirement directs to the covered dimensions has to come from somewhere, and a bounded attention budget means it comes from the dimensions the requirement does not cover. This paper examines that implication: it shows that partial adoption imposes a negative attention externality on the uncovered dimensions, that the externality is a Goodhart reallocation, that it is transitional rather than permanent, and that its cost during the transition depends on the order in which coverage expands.
The first result establishes the externality.
Proposition AE-1 (The attention externality). A disclosure requirement that covers some constraint dimensions but not others draws attention toward the covered dimensions, and because attention at the decision point is bounded, attention drawn to the covered dimensions is withdrawn from the uncovered ones. Partial adoption therefore imposes a negative attention externality on the dimensions it does not cover.
Proof sketch. An analyst has a bounded budget of attention to allocate across constraint dimensions at the moment of decision. A disclosure requirement on a set of dimensions raises the salience and the cost of inattention to those dimensions, because the analyst must produce a structured assessment of them that will be scored, which directs attention toward them. Under a bounded budget, attention directed toward the covered dimensions is unavailable for the uncovered ones, so the uncovered dimensions receive less attention than they would absent the requirement. The reduction in attention to the uncovered dimensions is a cost imposed on those dimensions by the requirement on the covered ones, which is a negative externality, since the decision to require disclosure on one set of dimensions degrades the observation of another set that the decision did not address. The externality exists precisely because adoption is partial: a requirement that covered all dimensions would direct attention across all of them without an uncovered set to starve. End sketch.
The result identifies a cost that an enthusiastic partial rollout could overlook. Bringing a few constraint dimensions under a disclosure requirement looks like pure gain, because those dimensions become observed where they were not, and the gain is real. The cost is that the analyst's finite attention now concentrates on the newly required dimensions, and the dimensions still outside the requirement receive less attention than before, so a partial rollout can improve observation on its target dimensions while degrading it on the rest. The net effect depends on how the gain on the covered dimensions compares with the loss on the uncovered ones, and a rollout that ignored the loss would overstate its benefit.
The second result identifies the externality's structure, which connects it to a mechanism the corpus already analyzed.
Proposition AE-2 (Goodhart structure). The attention externality is a Goodhart reallocation. The covered dimensions become the de facto target, observation bandwidth flows toward them, and the uncovered dimensions are starved, which is the same reallocation that occurs when agents optimize on a scalar target and bandwidth shifts toward the targeted dimension at the expense of the rest.
Proof sketch. Goodhart reallocation is the result that when agents optimize on a target derived from a multi-dimensional source, their observation bandwidth reallocates toward the targeted dimension, raising the neglect of the others (Beier, 2026i, the Goodhart derivation; the Universal Goodhart Theorem of the companion formal work). A partial disclosure requirement makes the covered dimensions a target, because the analyst is scored on them and not on the uncovered ones, so the analyst optimizes attention toward the dimensions that carry consequence. Bandwidth therefore reallocates toward the covered dimensions and away from the uncovered ones, which is the Goodhart reallocation restricted to the covered subset. The attention externality of Proposition AE-1 is thus not a new failure mode but the Goodhart mechanism applied to partial adoption, with the covered dimensions playing the role of the scalar target and the uncovered dimensions playing the role of the starved remainder. End sketch.
The identification matters because it places the externality in a known structure and tells the rollout what to watch for. The Goodhart reallocation is the same pathology the architecture exists to remedy at the level of the price, where the scalar is the target and the constraint dimensions are starved. A partial disclosure rollout reproduces the pathology in miniature, with the covered dimensions as the new target, and the irony is that a mechanism designed to relieve attention starvation can, while its adoption is partial, induce a smaller version of the same starvation on the dimensions it has not yet reached. Recognizing the externality as Goodhart reallocation is what makes it predictable and, by the next result, bounded.
The third result shows the externality is a feature of the transition and not of the architecture at rest.
Proposition AE-3 (Monotone decline in coverage). The attention externality decreases monotonically as coverage expands and vanishes at full coverage. Because the externality falls only on uncovered dimensions, the shrinking of the uncovered set as coverage grows shrinks the externality, which reaches zero when no uncovered dimension remains.
Proof sketch. The externality of Proposition AE-1 is the attention starvation imposed on the uncovered dimensions. As coverage expands, dimensions move from uncovered to covered, so the set of dimensions that can be starved shrinks. The total externality, the sum of the starvation over the uncovered dimensions, therefore decreases as the uncovered set shrinks, which is monotone decline in coverage. At full coverage the uncovered set is empty, there are no dimensions left to starve, and the externality is zero. The externality is therefore a transitional cost that is largest when coverage is smallest and disappears when coverage is complete, rather than a permanent property of the disclosure architecture. End sketch.
The result is the reassurance the externality requires, stated without overstating it. The attention externality is real, and a rollout that ignored it would understate its own transitional cost, but the externality is not a reason against the architecture, because it is a cost of being partway through adoption rather than a cost of the architecture itself. The fully adopted architecture has no uncovered dimensions and so imposes no attention externality; the partially adopted architecture imposes one that shrinks as adoption proceeds. The honest framing is that the externality raises the cost of a slow or stalled rollout, because the longer adoption remains partial, the longer the uncovered dimensions are starved, which is an argument for completing adoption rather than an argument against beginning it.
The fourth result concerns the transition's cost and how to minimize it.
Proposition AE-4 (Sequencing to minimize starvation). The cumulative attention externality over the transition depends on the order in which dimensions are brought under the disclosure requirement, and the order that minimizes the cumulative starvation prioritizes the dimensions whose neglect during the transition would be most costly. This objective can conflict with the information-gain ordering that the coupling analysis recommends.
Proof sketch. The cumulative externality over a transition is the starvation imposed on the not-yet-covered dimensions integrated over the time they remain uncovered. A dimension that is covered early is starved for less of the transition than a dimension covered late, so the cumulative starvation depends on which dimensions are covered early and which late. To minimize the cumulative starvation, the dimensions whose neglect is most costly, those carrying the largest constraint cost per unit of attention withdrawn, should be covered earliest, so that they spend the least time in the starved uncovered set. This is a greedy ordering by the cost of neglect. It need not coincide with the ordering the coupling analysis recommends, which connects domains in the order that maximizes the cross-domain information recovered per domain added (Beier, 2026, "Joint Dynamics of Coupled Constraint States," Proposition CD-4). The most attention-vulnerable dimension and the most information-rich connection are different criteria, and a dimension can rank high on one and low on the other, so the two sequencing objectives can pull apart. End sketch.
The tension is worth stating plainly rather than resolving by fiat. A rollout has at least two reasons to prefer one order of coverage over another: to recover the most cross-domain information per step, which the coupling analysis addresses, and to minimize the attention starvation during the transition, which this proposition addresses. The two need not agree, because the dimension whose coverage recovers the most coupling information is not necessarily the dimension whose neglect would be most costly, and a rollout that optimized only for information gain might leave a highly attention-vulnerable dimension starved for the length of the transition. The paper does not supply a single optimal order, because the order depends on the relative weight a deployment places on information recovery against attention protection, which is a deployment-specific judgment. What it supplies is the recognition that both criteria exist and can conflict, so that a rollout chooses its order deliberately rather than discovering the starvation after the fact.
The attention externality is a transitional cost of partial adoption with a clear structure and a clear remedy. It arises because attention at the decision point is finite and a partial requirement reallocates that attention toward the covered dimensions, it has the form of a Goodhart reallocation with the covered dimensions as the target, it decreases monotonically as coverage expands and vanishes at full coverage, and its cost during the transition depends on a coverage order that can be chosen to minimize starvation. None of this argues against the architecture, and the paper has been careful to say so: the externality is a property of being partway through adoption, not of the adopted system, and the fully covered architecture imposes none.
What the externality tells a rollout is to attend to its own transition. A rollout that treats partial adoption as pure gain on the covered dimensions will understate its cost on the uncovered ones, and a rollout that drifts in a partial state will hold the uncovered dimensions in starvation longer than necessary. The practical guidance is to complete adoption rather than rest in a partial state, and to sequence the coverage deliberately, weighing the information-gain order against the attention-protection order according to the constraint dimensions most at risk during the transition. The externality is the reason a disclosure rollout should be planned as a transition with its own costs rather than executed as a series of independent additions, each treated as gain without regard to what it withdraws from the dimensions still waiting.
The coupling analysis showed that cross-domain coordination requires both domains to disclose, and that cross-domain disclosure recovers the coupling up to the weaker side. It assumed that an analyst in one domain can read an analyst's disclosure in another. Reading across a domain boundary is not free. A hydrologist's disclosure stated in hydrological terms is not directly actionable by a financial analyst, who must translate it into the terms of investment risk before acting on it, and the translation has a cost. This paper develops the economics of that cost. Reading a disclosure is the application of the reader's analytical framework to the record, and a cross-domain reader must first map the disclosure's domain-specific terms into the reader's framework, a translation whose cost rises with the distance between the two domains' frameworks. The translation cost is a second barrier to cross-domain coordination, distinct from the disclosure barrier: the coupling recovery the coupling analysis derived is realized only when a reader actually reads across the boundary, which a rational reader does only when the value of the recovered coupling exceeds the translation cost, so a coupling can stay unrecovered even when both sides disclose if translating between them is too expensive. Structured disclosure lowers the translation cost relative to unstructured cross-domain prose, because the six-facet format supplies a common schema that a reader projects from rather than a free-form text whose structure the reader must first reconstruct. A translation layer that translates a domain's disclosures once and serves many readers amortizes the cost across them, which makes cross-domain flow viable at a scale where pairwise translation would not be, and is the role the coproduction engine and cross-domain intermediaries play.
Keywords: translation, cross-domain reading, analytical framework, boundary objects, intermediaries, coupling recovery, OCOT
A hydrologist discloses an assessment of an aquifer: the drawdown trajectory, the recharge dynamics, the compaction threshold, stated in the terms a hydrologist uses. A financial analyst managing exposure to assets in the region needs that assessment, because the aquifer's trajectory bears on the assets' value, but the analyst cannot act on the hydrologist's disclosure as written. The disclosure speaks of recharge rates and storage coefficients, and the analyst's decision turns on asset values and risk premia, so the analyst must translate the hydrological assessment into the language of investment risk before it can enter a portfolio decision. The translation is work, and the work has a cost.
The coupling analysis established that cross-domain coordination requires both domains to disclose and that cross-domain disclosure recovers the coupling up to the weaker side (Beier, 2026, "Joint Dynamics of Coupled Constraint States," Proposition CD-3). It assumed the reader could read. This paper examines the cost of reading across a boundary, which the assumption hid. It shows that reading is the application of the reader's framework to the disclosure, that a cross-domain reader pays a translation cost that rises with the distance between the frameworks, that this cost is a second barrier to cross-domain coordination distinct from the disclosure barrier, that structured disclosure lowers it, and that a translation layer amortizes it across readers.
The corrected tensor gives reading a precise form, which the translation cost builds on.
Proposition TR-1 (Translation is cross-domain framework application). Reading a disclosure is the application of the reader's analytical framework to the record, and a reader in one domain reading a disclosure from another must first map the disclosure's domain-specific terms into the reader's framework. The cost of this translation rises with the distance between the two domains' frameworks, the dissimilarity of their variables, units, and constructs.
Proof sketch. A disclosure is a vector in the facet space, and a reader extracts decision-relevant content by applying an analytical framework, a linear map from the facet space to the reader's decision-relevant representation (Beier, 2026, "The Six-Facet Beier Tensor," Definition 2). For a reader native to the disclosure's domain, the framework applies directly, because the reader's representation and the disclosure's terms share a domain. For a cross-domain reader, the disclosure's terms belong to a domain the reader's framework does not natively represent, so the reader must first map those terms into terms the framework can act on, which is the translation. The cost of the mapping rises with the dissimilarity of the two domains' frameworks: when the domains share constructs, units, and variables, the mapping is short, and when they share little, the mapping must bridge a larger gap, so the translation cost is increasing in the framework distance. End sketch.
The result locates translation within the architecture's existing structure rather than treating it as an extra. The transformation law of the corrected tensor already says that different readers project the same disclosure through different frameworks, recovering different content, and the same-domain case of that law is reading without translation. The cross-domain case is reading with translation, where the projection must first bridge the gap between the disclosure's domain and the reader's. Translation is therefore the cross-domain instance of the framework application that reading always is, and its cost is the cost of bridging a framework gap that the same-domain reader does not face.
The second result shows that the translation cost can block coordination that disclosure alone would otherwise enable.
Proposition TR-2 (Translation gates coupling recovery). The coupling recovery that cross-domain disclosure provides is realized only when a reader reads across the boundary, which a rational reader does only when the value of the recovered coupling exceeds the translation cost. A coupling can therefore stay unrecovered even when both domains disclose, if the translation cost between them exceeds the value of recovering the coupling.
Proof sketch. The coupling recovery of Proposition CD-3 presupposes that the reader incorporates the other domain's disclosure into the reader's assessment, which requires the reader to translate that disclosure at the cost of Proposition TR-1. A reader who weighs the cost against the benefit incorporates the cross-domain disclosure only when the value of the recovered coupling exceeds the translation cost, since a reader does not pay a cost that exceeds its return. When the translation cost exceeds the value of the recovered coupling, the reader does not translate, the cross-domain disclosure is not incorporated, and the coupling stays unrecovered despite both domains having disclosed. The disclosure barrier of the coupling analysis, that both domains must disclose, is therefore necessary but not sufficient for coupling recovery; the translation cost must also be below the value of the recovered coupling, which is a second barrier the disclosure barrier did not capture. End sketch.
A second barrier appears where the coupling analysis assumed costless reading. That analysis treated cross-domain disclosure as recovering the coupling once both sides disclose, and the recovery is real, but only if the cross-domain reading actually happens, which the translation cost can prevent. Two domains can both disclose substantively and still fail to coordinate, because the cost of translating each other's disclosures exceeds the value either places on the coupling, in which case the disclosures sit in their respective domains unread across the boundary. The coupling between water and finance can be disclosed on both sides and remain unrecovered if no financial analyst finds the hydrological translation worth its cost, which makes lowering that cost a precondition for the coordination the disclosure was meant to enable.
The third result shows that the format does work on the translation cost, not only on the within-domain content.
Proposition TR-3 (Structured disclosure lowers translation cost). A six-facet disclosure has a lower cross-domain translation cost than unstructured prose about the same content, because the format supplies a common schema that the reader projects from, whereas unstructured prose requires the reader to reconstruct the disclosure's structure before translating it.
Proof sketch. The translation cost decomposes into the cost of recovering the disclosure's structure and the cost of projecting that structure into the reader's framework. For unstructured cross-domain prose, the structure-recovery cost is high, because the reader must infer how the prose is organized, what its claims are, and how its evidence relates to its conclusions before any translation can begin. For a six-facet disclosure, the structure is supplied by the format: the reader knows the disclosure carries an input, a method, an output, a scaling assessment, an identity, and a disposition, in that order, so the structure-recovery cost is near zero and only the framework-projection cost remains. The six-facet format thus functions as a boundary object, a shared structure that heterogeneous communities can each interpret in their own terms without renegotiating the structure each time (Star and Griesemer, 1989). The total translation cost of the structured disclosure, framework projection alone, is therefore lower than that of unstructured prose, structure recovery plus framework projection, so the format lowers the translation cost. End sketch.
The result adds a second sense in which the format pays for itself. The format makes within-domain assessments legible and scorable, which is its primary function, and it also lowers the cost of reading those assessments across domain boundaries, because a reader who knows the schema does not have to reconstruct it. The common schema is what lets a financial analyst find the hydrologist's output facet without parsing a hydrological essay to locate the conclusion, and what lets the analyst's framework project from a known structure rather than an inferred one. The format is, in this respect, an instrument for cross-domain coordination as much as for within-domain legibility, because the structure it imposes is the structure a cross-domain reader would otherwise have to recover at cost.
The fourth result shows how cross-domain flow becomes viable at scale.
Proposition TR-4 (Amortization through a translation layer). A translation layer that translates a domain's disclosures once and serves many readers amortizes the translation cost across them, which makes cross-domain flow viable at a scale where pairwise translation would not be. The amortized arrangement dominates pairwise translation when the marginal cost of serving a translated assessment is below the cost of translating it.
Proof sketch. Suppose a number of readers in one domain each need to read a disclosure from another. Under pairwise translation, each reader pays the translation cost, so the total is the number of readers times the translation cost. Under a translation layer that translates the disclosure once and serves the translated version to all readers, the total is the one-time translation cost plus the marginal serving cost times the number of readers. The layer is cheaper than pairwise translation when the serving cost is below the translation cost, because then the per-reader marginal cost under the layer is below the per-reader cost under pairwise translation, and the gap grows with the number of readers. Since serving a pre-translated assessment is in general cheaper than translating it from scratch, the layer dominates at scale, and cross-domain flow that would not occur under pairwise translation, because each reader's translation cost exceeds the reader's value, becomes viable under the layer, because the amortized per-reader cost falls below that value. End sketch.
The result identifies the role of intermediaries in cross-domain coordination. A coproduction engine that translates a domain's disclosures into the terms of other domains, or a cross-domain intermediary that maintains the mappings between framework pairs, pays the translation cost once and serves it to many readers, which is the structure that makes cross-domain reading affordable at scale. The arrangement has a precedent in the data intermediaries that let firms combine information they could not each process pairwise, and it is the role the coproduction engine plays when it surfaces a domain's assessments in forms other domains can read (Beier, 2026, "The Composition Algebra," on cross-domain combination). The translation layer is what turns the lowered per-disclosure cost of structured disclosure into a per-reader cost low enough that the cross-domain coupling actually gets read.
The translation analysis completes the account of what cross-domain coordination requires. The coupling analysis named the disclosure barrier, that both domains must disclose, and this analysis names the second barrier, that the cross-domain reading must be affordable, which the translation cost governs. Cross-domain coupling is recovered only when both domains disclose and the translation between them is cheap enough that readers cross the boundary, and the architecture works on both barriers: structured disclosure lowers the translation cost by supplying a common schema, and a translation layer amortizes the cost across readers, so the coupling that both sides disclose actually gets read.
The analysis sharpens the case for the coproduction engine and for the format that feeds it. The format earns its place not only because structured assessments are legible and scorable within a domain but because they are cheaper to read across domains, and the engine earns its place not only because it detects emergence within the commons but because it amortizes the cost of translating between domains that would otherwise read each other only at prohibitive cost. The water-energy-food nexus that the coupling analysis identified as the first network is also the first translation problem, three domains whose mutual reading the engine must make affordable, and the translation economics is what determines whether the coupling those three domains disclose is recovered in practice or left sitting unread across the boundaries between them.
A corrupted disclosure, whether an adversarial fabrication, an honest error, or a calibration failure, enters the scored commons and propagates to the analysts who read and build on it. Because the commons is a network in which disclosers interact combinatorially, a corrupted signal can spread, and the question is when it spreads and when it dies out. This paper models the propagation with the apparatus of epidemiology, assigning a corrupted signal a basic reproduction number equal to the rate at which it is read, times the probability a reader incorporates it, times the window during which it spreads before scoring corrects it. Three results follow. On constraint dimensions where outcomes resolve within the propagation window, proper scoring drives the reproduction number below one, because the corrupted signal's track record degrades as outcomes reveal its departure from truth, which lowers the rate at which it is read and the probability it is incorporated, so the signal is self-limiting and its own propagation destroys its credibility. On dimensions where the scoring lag exceeds the propagation window, the reproduction number can exceed one, which identifies the vulnerable domains as the long-resolution ones, the climate trajectories, institutional-resilience predictions, and deep aquifer dynamics where outcomes resolve too slowly for scoring to catch the signal before it spreads. Three architectural properties hold the reproduction number below one even under scoring lag: reading weighted by track record, which caps the reach of unestablished sources; cross-referencing against independently scored disclosures, which flags a signal that contradicts a consensus its own echoes cannot manufacture; and multi-channel detection, which catches a corrupted signal that any single detection channel would miss. The vulnerable domains are the same ones the scoring-lag failure mode identifies, and the containment architecture is what the commons relies on where scoring alone is too slow.
Keywords: information epidemiology, reproduction number, corrupted signals, scoring lag, track record, containment, OCOT
A disclosure enters the scored commons carrying a conclusion that is wrong. It may be wrong because an adversary fabricated it, because an honest analyst erred, or because a poorly calibrated analyst believed it, and the cause does not change what happens next: other analysts read it, some incorporate it into their own assessments, and those assessments are read in turn. The commons is a network in which disclosers interact and build on one another, which is the source of its combinatorial information production (Beier, 2026ag), and the same network that propagates good signals propagates bad ones. A corrupted disclosure can spread through it.
Whether it spreads or dies out is a question epidemiology already knows how to ask. An infectious agent spreads when each case produces more than one further case on average and dies out when each produces fewer, a threshold captured by the basic reproduction number (Kermack and McKendrick, 1927; Anderson and May, 1991). A corrupted signal in the commons has an analogous reproduction number, and this paper computes it, shows when scoring drives it below the threshold, identifies the domains where scoring cannot, and specifies the architectural properties that contain the signal where scoring is too slow. The corruption is treated as a contagion through the disclosure network, and the question is what keeps the contagion from becoming an epidemic.
Let a corrupted disclosure be read by analysts at some rate, incorporated by a reader with some probability, and spread for some window before scoring corrects it. Define its basic reproduction number as
R_info = beta times kappa times w,
where beta is the rate at which the signal is read, kappa is the probability that a reader who reads it incorporates it into a disclosure of their own, and w is the propagation window, the time during which the signal spreads before scoring degrades its credibility. The product is the expected number of further corrupted disclosures that one corrupted disclosure produces: each reader who incorporates the signal produces a further carrier, the reading rate and incorporation probability set how many readers do so, and the window sets how long the spreading continues.
The threshold is the same as in epidemiology. When R_info is below one, each corrupted disclosure produces fewer than one further corrupted disclosure on average, and the signal dies out. When R_info exceeds one, each produces more than one, and the signal spreads through the commons until something checks it. The propagation window w is closely tied to the scoring lag, because scoring is what corrects the signal, so a dimension whose outcomes resolve quickly has a short window and a dimension whose outcomes resolve slowly has a long one. The three results that follow concern when scoring drives R_info below the threshold, when it cannot, and what holds the signal below the threshold when scoring is too slow.
The first result shows that on fast-resolving dimensions the signal destroys its own ability to spread.
Proposition IE-1 (Self-limitation under scoring). On a constraint dimension where outcomes resolve within the propagation window, proper scoring drives the reproduction number below one. The corrupted signal's track record degrades as outcomes reveal its departure from truth, which lowers the rate at which it is read and the probability it is incorporated, so the signal's propagation destroys its own credibility.
Proof sketch. The rate at which a disclosure is read and the probability it is incorporated both depend on the source's track record, because a higher track record draws more readers and more incorporation through the track-record metric that weights disclosures by cumulative scoring performance (Beier, 2026, "The Six-Facet Beier Tensor," Definition 3; Beier, 2026z, BT-2). A corrupted signal departs from ground truth, so when outcomes resolve they score it poorly, and its track record degrades in proportion to its departure. When the resolution time is within the propagation window, this degradation occurs while the signal is still spreading, so the reading rate beta and the incorporation probability kappa fall during the propagation, which lowers the product beta times kappa times w below the threshold. The signal is therefore self-limiting on fast-resolving dimensions: the act of spreading exposes it to scoring, the scoring degrades its track record, and the degraded track record reduces its further spread, so the reproduction number falls below one and the signal dies out. End sketch.
The result is the commons's first line of defense and the cleanest one. A corrupted signal on a dimension that resolves quickly cannot establish itself, because spreading it means being read and scored, and being scored against quickly resolving outcomes destroys the track record that spreading depends on. The signal's propagation is its own undoing on these dimensions, which is why the commons needs no special machinery to contain corruption where outcomes resolve fast: the scoring that the architecture already performs drives the reproduction number below the threshold as a matter of course. The defense weakens only where the scoring is too slow, which the next result addresses.
The second result identifies where the first defense fails.
Proposition IE-2 (Epidemics under scoring lag). On a constraint dimension where the scoring lag exceeds the propagation window, the reproduction number can exceed one, so a corrupted signal can spread. The vulnerable dimensions are the long-resolution ones, where outcomes resolve too slowly for scoring to degrade the signal's track record before it spreads.
Proof sketch. The self-limitation of Proposition IE-1 depends on the corrupted signal's track record degrading while it spreads, which requires outcomes to resolve within the propagation window. When the scoring lag exceeds the window, outcomes have not resolved by the time the signal has spread, so the track record has not degraded, and the reading rate and incorporation probability remain at the levels the signal's stated confidence and apparent quality command rather than at the degraded levels its falsity would eventually produce. The reproduction number beta times kappa times w is then computed at the undegraded reading rate and incorporation probability and at a long window, so it can exceed one, and the signal spreads. The vulnerable dimensions are those with long resolution times, since they are the ones where the scoring lag exceeds the propagation window: climate trajectories, institutional-resilience predictions, and deep aquifer dynamics resolve over years to decades, so a corrupted signal on them can spread for years before scoring corrects it. End sketch.
The vulnerable domains are exactly the ones the scoring-lag failure mode identified from a different direction. The non-adversarial analysis found that on long-resolution dimensions the track record cannot reach useful length before the constraint binds, which leaves calibrated and uncalibrated analysts indistinguishable in the window that matters (Beier, 2026, "Non-Adversarial Failure Modes," Proposition NF-2). The epidemiological analysis finds that on the same dimensions a corrupted signal can spread before scoring corrects it. The two are the same weakness seen as a failure of discrimination and as a failure of containment, and they coincide because both depend on scoring catching up in time, which it does not on the slowest dimensions. These are the dimensions where the commons cannot rely on scoring alone, and where the containment architecture of the next result has to carry the load.
The third result specifies the defenses that hold even when scoring is slow.
Proposition IE-3 (Containment architecture). Three architectural properties hold the reproduction number below one even under scoring lag. Reading weighted by track record caps the reach of unestablished sources. Cross-referencing against independently scored disclosures flags a signal that contradicts a consensus its own echoes cannot manufacture. Multi-channel detection catches a corrupted signal that any single detection channel would miss. Each lowers the reading rate or the incorporation probability enough to offset a long window.
Proof sketch. Reading weighted by track record makes the reading rate beta a function of the source's established track record, so a corrupted signal from an unestablished source has a low reading rate regardless of the scoring lag, because the source has not built the track record that wide reading requires; this caps the reach of new corrupted sources before any outcome resolves. Cross-referencing against independently scored disclosures reduces the incorporation probability kappa for a signal that contradicts the consensus of analysts with disjoint lineages, because a reader who checks a signal against independent assessments incorporates it less when it conflicts with them, and the independence requirement is what makes the check valid, since agreement among sources confirms only when their lineages are disjoint and a corrupted signal cannot manufacture confirmation by spawning its own echoes (Beier, 2026, "The Composition Algebra," Proposition C-3). Multi-channel detection reduces kappa further by catching corrupted signals that fail any of several detection channels, the economic, computational, mathematical, and information-theoretic channels of the adversarial analysis together with the provenance channel of the substrate marker, which have no shared blind spot, so a signal that evades one channel is caught by another (Beier, 2026y; Beier, 2026, "The Substrate Marker"). Each property reduces beta or kappa, and together they reduce the product beta times kappa enough that the reproduction number stays below one even when the window w is long, which contains the signal where scoring alone would not. End sketch.
The containment architecture is what the commons relies on precisely where it is most exposed. On fast-resolving dimensions, scoring contains corruption by itself, and the architecture is a backstop. On the long-resolution dimensions where scoring lags, the architecture is the primary defense, and its three properties attack the reproduction number through its controllable factors: the reach of a source through track-record-weighted reading, the incorporation of a contradicted signal through independent cross-referencing, and the incorporation of a detectable signal through multi-channel detection. The independence requirement is the load-bearing one, because it is what prevents a corrupted signal from manufacturing its own apparent confirmation, and it is the same requirement that keeps the convergence signal honest, so the property that distinguishes consensus from echo is also the property that contains contagion.
The epidemiological model locates the commons's exposure to corruption precisely. A corrupted signal spreads when its reproduction number exceeds one, scoring drives the number below one on dimensions that resolve quickly, and the vulnerable dimensions are the long-resolution ones where scoring lags, which are the same dimensions the scoring-lag failure mode identifies as the architecture's weakest case. Where scoring is too slow, three architectural properties hold the reproduction number below one, and the independence requirement among them is the one that prevents a signal from confirming itself through its own echoes.
The model also tells a deployment where to spend its containment effort. On fast-resolving dimensions, the scoring the architecture already performs is sufficient, and a deployment can rely on it. On long-resolution dimensions, the deployment must invest in the containment architecture, in track-record-weighted reading, in independent cross-referencing, and in multi-channel detection, because scoring will not catch a corrupted signal before it spreads. The same long-resolution dimensions that the consent analysis identified as where the future bears unobserved costs, and that the failure analysis identified as where the track record cannot discriminate in time, are the dimensions where corruption can spread before correction, so the architecture's hardest problems concentrate on the slow-resolving constraints, and the containment architecture is part of what the commons must build to serve them.
The corpus has noted, more than once, a structural parallel between the minimum-dissipation function of the coordination theory and the free energy of statistical mechanics, with a water level playing a role like temperature. The parallel is suggestive, and a strong reading of it would mean the coordination theory touches physics rather than being optimization theory with rich structure. This paper tests the strong reading and rejects it. An isomorphism between two theories requires agreement at four levels: the optimization structure, the state-function structure expressed through Legendre duality, the fluctuation structure expressed through a fluctuation-dissipation relation, and the phase-transition structure expressed through universality class. The paper shows that the parallel holds at the first two levels and that holding there is no evidence, because Legendre duality and the Maxwell relations it generates are generic consequences of convex optimization over eigenvalues with a sum constraint, which both theories happen to be, rather than signatures of shared physics. It then shows that the two non-generic levels, where a real isomorphism would have to be demonstrated and not inherited, do not support the claim: the coordination theory has no derived fluctuation-dissipation relation, only a postulated analogy through the reflexive observation drift, and the coordination phase transition is a saddle-node bifurcation whose universality class differs from the symmetry-breaking transitions usually invoked in thermodynamics. The conclusion is negative on the strong isomorphism and positive on the weak parallel: the two theories share mathematical form at the optimization level and diverge at the fluctuation and universality levels, so the coordination theory's contact with physics, if it has any, does not run through classical thermodynamic isomorphism and would have to run through the quantum-level conjectures of the observation-floor work, which are a separate and more speculative bet.
Keywords: rate-distortion theory, statistical mechanics, Legendre duality, Maxwell relations, fluctuation-dissipation, universality class, negative result, OCOT
The coordination theory has a function that behaves like a thermodynamic potential. The minimum dissipation given an observation capacity, written D*(M), is the value of a convex optimization that allocates limited capacity across constraint dimensions, and it has a water-filling structure in which a level rises until the capacity is exhausted, the dimensions below the level being observed and those above it being starved (Beier, 2026i). The level plays a role that looks like temperature, the dissipation looks like free energy, and the resemblance has been noted across the corpus as suggestive of something deeper, with the observation-floor work extending it to conjectures about a connection at the quantum scale (Beier, 2026m).
A strong reading of the resemblance would be consequential. If the coordination theory and statistical mechanics were isomorphic, the coordination theory would not merely borrow the language of thermodynamics but share its structure, which would mean the theory touches physics rather than being an optimization theory whose objects happen to resemble physical ones. The claim is worth wanting and therefore worth testing carefully, because a resemblance one wants to be deep is exactly the kind a careful analysis should be most skeptical of. This paper tests it. It states what an isomorphism would require, shows that the parts of the resemblance that hold are the parts that would hold for any theory of the same mathematical type, shows that the parts that would distinguish a real isomorphism do not hold or are not demonstrated, and concludes against the strong claim while preserving the weak one.
Two theories are isomorphic in the relevant sense when they agree not only in the form of their optimization but in the structure that the optimization induces, at four levels. The first is the optimization structure itself, that both minimize or maximize a function of the same type subject to a constraint of the same type. The second is the state-function structure, that both admit a Legendre duality swapping conjugate variables, so that a potential expressed in one variable transforms into a potential expressed in its conjugate, as free energy in temperature transforms into entropy in energy. The third is the fluctuation structure, that both obey a fluctuation-dissipation relation tying the size of fluctuations around the optimum to the system's response to perturbation, which in statistical mechanics relates the variance of an observable to its susceptibility. The fourth is the phase-transition structure, that both undergo transitions of the same universality class, sharing critical exponents and qualitative form at the transition.
The four levels are not equally informative about an isomorphism, and the distinction is the heart of this paper. The first two are generic: any theory that minimizes a smooth convex function subject to a constraint has an optimization structure and a Legendre dual, because convex duality is a property of convex problems and not of any particular subject matter (Rockafellar, 1970). Agreement at the first two levels therefore tells us only that both theories are convex optimizations, which is not in dispute and not evidence of shared physics. The second two levels are non-generic: a fluctuation-dissipation relation and a universality class are properties a theory has to earn through its specific dynamics, and two theories that agree on them agree on something their convexity does not force. An isomorphism worth the name has to be demonstrated at the non-generic levels, and the generic levels, however striking the agreement looks, carry no weight.
The first result establishes that the most-cited piece of the resemblance is the piece that proves nothing.
Proposition TH-ISO-1 (Generic Legendre duality). The minimum-dissipation function D*(M) has a Legendre dual, the rate-distortion function, and the transform between them is structurally identical to the thermodynamic transform between free energy and entropy. This identity is a generic consequence of convex optimization and is not evidence of a connection to physics.
Proof sketch. The minimum-dissipation function is the value of a convex optimization, and its Legendre dual is the rate function R*(D) obtained by transforming with respect to the capacity variable, which for the Gaussian model takes the water-filling form summing the positive parts of the log-ratios of eigenvalues to distortions (Cover and Thomas, 2006). The transform pairs the capacity with its conjugate, the water level, exactly as the thermodynamic transform pairs energy with temperature, so the two transforms have the same structure. The structure is generic, because every closed convex function is its own biconjugate and admits a Legendre dual that swaps the function's variable for its conjugate, which is the content of convex duality and holds for any convex value function regardless of what it describes (Rockafellar, 1970). The thermodynamic pair and the rate-distortion pair are two instances of this generic duality, and they share it because both are convex optimizations, not because they share a physical mechanism. The identity of the transforms is therefore a consequence of convexity that any two convex theories would exhibit, and it carries no information about whether the coordination theory touches physics. End sketch.
The result deflates the resemblance's most quoted feature. That the dissipation function and the free energy are Legendre duals of their respective rate and entropy functions looks like a deep correspondence, and it is the kind of fact a strong reading leans on, but it is the kind of fact every pair of convex optimizations exhibits. A theory of portfolio optimization and a theory of optimal transport would display the same Legendre structure, and no one would conclude that portfolios obey thermodynamics. The coordination theory's Legendre duality with rate-distortion is the duality convexity guarantees, and the agreement with thermodynamics at this level is the agreement any convex theory has with any other, which is to say no agreement of substance at all.
The second result addresses the next piece of the resemblance and finds it equally generic.
Proposition TH-ISO-2 (Maxwell relations from smoothness). The coordination theory admits relations among the cross partials of its dissipation function that are analogous to the thermodynamic Maxwell relations, and these relations hold as algebraic identities for the Gaussian model. They hold because the dissipation function is a smooth convex potential whose mixed partials commute, not because of any physical content, so their agreement with the thermodynamic Maxwell relations is generic.
Proof sketch. The thermodynamic Maxwell relations are the equalities among the cross partial derivatives of a state function, following from the commutativity of mixed second partials of a smooth potential. The coordination analog is the corresponding equality among the cross partials of the dissipation function with respect to its arguments, the capacity and the number of observed dimensions, and computing both sides for the Gaussian water-filling model shows the equality holds. The equality holds because the dissipation function is twice continuously differentiable as a convex value function, and the mixed second partials of a twice continuously differentiable function commute by the symmetry of second derivatives, so the relation is forced by smoothness alone. The thermodynamic Maxwell relations have the same origin, the smoothness of the thermodynamic potential, so the two sets of relations agree because both potentials are smooth and not because they describe the same physics. The agreement is therefore generic, a consequence of differentiability that any smooth potential exhibits, and it provides no more evidence of an isomorphism than the Legendre duality did. End sketch.
The two generic results together account for the bulk of the resemblance and dispose of it as evidence. The Legendre duality and the Maxwell relations are the features a strong reading points to, and both are properties that the coordination theory has by virtue of being a smooth convex optimization, which is what it is and not in question. A resemblance built on these features is a resemblance every smooth convex theory shares with thermodynamics, and the impression of depth it creates is the impression convexity always creates when its generic consequences are mistaken for specific ones. What would distinguish a real isomorphism is agreement at the levels convexity does not force, and the next result examines those.
The third result turns to the two levels where an isomorphism would have to be earned and finds the evidence absent or negative.
Proposition TH-ISO-3 (Failure at the non-generic levels). At the two levels that a generic convexity does not force, the isomorphism is not supported. The coordination theory has no derived fluctuation-dissipation relation, only a postulated analogy through the reflexive observation drift, and the coordination phase transition is a saddle-node bifurcation whose universality class differs from the symmetry-breaking transitions usually invoked in thermodynamics.
Argument. The fluctuation-dissipation level requires a relation tying the fluctuations around the coordination optimum to the system's response to perturbation, derived from the theory's dynamics rather than asserted by analogy. The candidate for the fluctuation is the reflexive observation drift, the noise that the system's own observation introduces, and the candidate for the response is the dissipation, but the relation between them in the coordination theory is postulated by analogy to the physical fluctuation-dissipation theorem and is not derived from a coordination-theoretic first principle. Absent a derivation, there is no positive evidence at this level, and the burden of proof for an isomorphism, which falls on the claim and not on its denial, is unmet. The universality level requires that the coordination phase transition share a universality class with a thermodynamic transition. The coordination transition occurs at the stability ratio crossing one and is a saddle-node bifurcation, a fold with no symmetry breaking, which the corpus's bistability and basin results establish (Beier, 2026e, the phase transition; the basin-transition analysis). The continuous thermodynamic transitions usually invoked in the resemblance are symmetry-breaking, of pitchfork or comparable type, with critical exponents a saddle-node does not share. A saddle-node and a symmetry-breaking transition are in different universality classes, so the coordination transition and the canonical thermodynamic transitions do not match at this level, which is negative evidence rather than absent evidence. The two non-generic levels therefore yield, respectively, no support and contrary support, and an isomorphism that cannot be demonstrated at either non-generic level is not established. End argument.
The result completes the case. The levels at which the coordination theory and thermodynamics agree are the levels at which any two convex theories agree, and the levels at which a real isomorphism would have to show itself are the levels at which the coordination theory either has not been shown to agree or has been shown to differ. The fluctuation-dissipation relation, which would be the strongest single piece of evidence, is postulated rather than derived, so it is a hope and not a result. The universality class, which would be the second, points the other way, because a fold is not a symmetry breaking. The strong isomorphism is unsupported where it would matter and generic where it looks impressive.
The conclusion is negative on the strong claim and positive on the weak one. The coordination theory and statistical mechanics share mathematical form at the optimization level, the Legendre duality and the Maxwell relations, because both are smooth convex optimizations over eigenvalues with a sum constraint, and this shared form is real and is not evidence of a deep connection, because every theory of that type shares it. The theories diverge at the fluctuation and universality levels, where the coordination theory has no derived fluctuation-dissipation relation and where its phase transition is a saddle-node rather than a symmetry breaking, so the structure that would make the resemblance an isomorphism is not present. The honest statement is that the coordination theory is optimization theory with rich structure, that the structure resembles thermodynamics at the level convexity guarantees, and that the resemblance does not reach the level that would make it physics.
The result has a consequence for where the corpus's contact with physics, if it has any, would have to be sought. It is not in the classical thermodynamic isomorphism, which this paper has rejected. If the coordination theory touches physics, the contact would have to run through the observation-floor conjectures, which place the connection at the quantum scale, relating the observation level to the Planck constant rather than to classical thermodynamic potentials (Beier, 2026m). Those conjectures are a separate and more speculative bet than the classical isomorphism, and this paper does not evaluate them; it only locates the contact there by eliminating the classical route. The discipline the result imposes on the corpus is to stop leaning on the thermodynamic resemblance as evidence of unification with physics, because the resemblance is generic, and to treat the physics question as open and resting on the quantum conjectures, which have their own and heavier burden of proof.
Akerlof, G. A. (1970). "The Market for Lemons: Quality Uncertainty and the Market Mechanism." Quarterly Journal of Economics, 84(3), 488 through 500.
Amari, S. (1985). Differential-Geometrical Methods in Statistics. Berlin: Springer.
Anderson, R. M., and May, R. M. (1991). Infectious Diseases of Humans: Dynamics and Control. Oxford: Oxford University Press.
Beier, G. C. (2026). "Basin Transition Economics: How Constraint Dimensions Cross Thresholds, Why the Price Cannot See Them Coming, and What the Disclosure Mechanism Changes." Working Paper.
Beier, G. C. (2026). "Joint Dynamics of Coupled Constraint States: How Domain Boundaries Destroy Information and How Cross-Domain Disclosure Recovers It." Working Paper.
Beier, G. C. (2026). "Non-Adversarial Failure Modes: Ontology Drift, Scoring Lag, Platform Fragmentation, and Quality Collapse." Working Paper.
Beier, G. C. (2026). "Strike-Time and Transmit-Time: The Information Economics of the Two-Timestamp Record." Working Paper.
Beier, G. C. (2026). "Sustainability as an Emergent Aggregate: Why It Is Read From the Scaling Facets and Is Not a Fourth Domain." Working Paper.
Beier, G. C. (2026). "The Composition Algebra: Lineage, Independence, and the Detection of Emergence." Working Paper.
Beier, G. C. (2026). "The Engagement Level: Depth of Work as a Reader's Signal, Not a Scorer's Input." Working Paper.
Beier, G. C. (2026). "The Language Mediation Channel: Under What Conditions Does Text About Physics Preserve the Physics?" Working Paper.
Beier, G. C. (2026). "The Scaling Facet as a Distinct Information Type: Why Interaction Effects at Aggregate Scale Cannot Be Derived from Individual-Transaction Assessment." Working Paper.
Beier, G. C. (2026). "The Six-Facet Beier Tensor: A Formal Definition of the Extended Price with Disposition, Temporal, and Composition Structure." Working Paper.
Beier, G. C. (2026). "The Substrate Marker: Provenance, Self-Monitoring, and the Mixed Human-Machine Commons." Working Paper.
Beier, G. C. (2026). "The Transmit Facet: Disposition, Scope, and Why the Public Option Cannot Be Removed." Working Paper.
Beier, G. C. (2026aa). "The Economics of Cognitive Trade: Three Propositions on the Human-AI Exchange Rate." Working Paper. Zenodo.
Beier, G. C. (2026ab). "The Performative Proposition: How Structured Disclosure Formats Generate Information That Unstructured Formats Cannot Elicit." Working Paper. Zenodo.
Beier, G. C. (2026ac). "Formal Foundations for Symbolic Time Crystals: Three Propositions on Temporal Invariance in Learned Representations." Working Paper. Zenodo.
Beier, G. C. (2026ae). "The Solow Residual as Coordination Quality: How Information Architecture Determines Total Factor Productivity." Working Paper. Zenodo.
Beier, G. C. (2026af). "Reality Debt: The Accumulated Cost of Decisions Made on Compressed Information." Working Paper. Zenodo.
Beier, G. C. (2026ag). "Factorial Growth Through Coordination: Why the Interaction of N Disclosers Produces Combinatorial Information, Not Linear." Working Paper. Zenodo.
Beier, G. C. (2026ah). "Scaling Risk as the Fifth Consideration: Why Interaction Effects at Scale Constitute a Distinct Information Type the Price System Cannot Carry." Working Paper. Zenodo.
Beier, G. C. (2026ai). "The Mathematical Structure of Interaction Effects Across Scale." Working Paper. Zenodo.
Beier, G. C. (2026al). "How the Improvement Mechanism Scales Across Populations: The Economics of Collective Calibration Improvement." Working Paper. Zenodo.
Beier, G. C. (2026am). "When Constraint Domains Interact: The Economics of Cross-Domain Information Failure." Working Paper. Zenodo.
Beier, G. C. (2026an). "Adversarial Disclosure Economics: Attacks on the Judgment Data Stream and the Cross-Disciplinary Defense That Unification Provides." Working Paper. Zenodo.
Beier, G. C. (2026aq). "The Institutional Substrate for Knowledge Coordination: What Must Exist Before Disclosure Can Work." Working Paper. Zenodo.
Beier, G. C. (2026as). "The Geometry of Constraint Knowledge: How Physical Laws Impose Structure on Learned Representations, and What This Means for Coordination." Working Paper. Zenodo.
Beier, G. C. (2026at). "Symbolic Time Crystals: Temporal Invariants in Human-Generated Data Streams and Their Detection Through Weight-Space Geometry." Working Paper. Zenodo.
Beier, G. C. (2026b). "The Disclosure Incentive: How Track Record Capital Resolves the Grossman-Stiglitz Paradox and Reveals the Competition-Cooperation Synthesis Across Four Disciplines." Working Paper. Zenodo.
Beier, G. C. (2026e). "Supple: How Beierspace Extends Eigenspace." Working Paper. Zenodo.
Beier, G. C. (2026h). "Excess Distortion in the Price System: A Rate-Distortion Analysis of Scalar Compression in Multi-Dimensional Constraint Environments." Working Paper. Zenodo.
Beier, G. C. (2026i). "Observation Is Maintenance: Axiomatic Foundations for Coordination Thermodynamics." Working Paper. Zenodo.
Beier, G. C. (2026m). "The Observation Floor: Is Planck's Constant the Universe's Water Level?" Working Paper. Zenodo.
Beier, G. C. (2026n). "Five Axioms for Coordination Under Compression." Working Paper. Zenodo.
Beier, G. C. (2026r). "The Incentive Compatibility of the Disclosure Incentive: A Mechanism Design Foundation for the Judgment Market." Working Paper. Zenodo.
Beier, G. C. (2026t). "The Extropy-Distortion Bridge: How Beier Extropy Maps to Rate-Distortion Improvement in the Price System." Working Paper. Zenodo.
Beier, G. C. (2026u). "Compositional Improvement Beyond the Gaussian Case." Working Paper. Zenodo.
Beier, G. C. (2026v). "A Tight Bound on the Calibration Threshold for Honest Disclosure: Order Statistics Refinement of the Strategic Selection Analysis." Working Paper. Zenodo.
Beier, G. C. (2026x). "Conservation Geometry Under Measurement Noise: A Quantitative Convergence Rate for Weight-Space Rank Deficiency." Working Paper. Zenodo.
Beier, G. C. (2026y). "Adversarial Detection Superadditivity and Geometric Distinguishability." Working Paper. Zenodo.
Beier, G. C. (2026z). "The Beier Tensor: A Formal Geometric Definition of the Deep Price." Working Paper. Zenodo.
Benjamini, Y., and Hochberg, Y. (1995). "Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing." Journal of the Royal Statistical Society Series B, 57(1), 289 through 300.
Brier, G. W. (1950). "Verification of Forecasts Expressed in Terms of Probability." Monthly Weather Review, 78(1), 1 through 3.
Cover, T. M., and Thomas, J. A. (2006). Elements of Information Theory, 2nd ed. Hoboken: Wiley.
Ericsson, K. A., Krampe, R. T., and Tesch-Romer, C. (1993). "The Role of Deliberate Practice in the Acquisition of Expert Performance." Psychological Review, 100(3), 363 through 406.
Hsieh, C.-T., and Klenow, P. J. (2009). "Misallocation and Manufacturing TFP in China and India." Quarterly Journal of Economics, 124(4), 1403 through 1448.
Jaynes, E. T. (1957). "Information Theory and Statistical Mechanics." Physical Review, 106(4), 620 through 630.
Kahneman, D. (2011). Thinking, Fast and Slow. New York: Farrar, Straus and Giroux.
Kermack, W. O., and McKendrick, A. G. (1927). "A Contribution to the Mathematical Theory of Epidemics." Proceedings of the Royal Society A, 115(772), 700 through 721.
Nordhaus, W. D. (2007). "A Review of the Stern Review on the Economics of Climate Change." Journal of Economic Literature, 45(3), 686 through 702.
Nosek, B. A., Ebersole, C. R., DeHaven, A. C., and Mellor, D. T. (2018). "The Preregistration Revolution." Proceedings of the National Academy of Sciences, 115(11), 2600 through 2606.
Ramsey, F. P. (1928). "A Mathematical Theory of Saving." Economic Journal, 38(152), 543 through 559.
Rockafellar, R. T. (1970). Convex Analysis. Princeton: Princeton University Press.
Romer, P. M. (1990). "Endogenous Technological Change." Journal of Political Economy, 98(5), S71 through S102.
Simon, H. A. (1971). "Designing Organizations for an Information-Rich World." In M. Greenberger, ed., Computers, Communications, and the Public Interest, 37 through 72. Baltimore: Johns Hopkins Press.
Solow, R. M. (1957). "Technical Change and the Aggregate Production Function." Review of Economics and Statistics, 39(3), 312 through 320.
Star, S. L., and Griesemer, J. R. (1989). "Institutional Ecology, Translations and Boundary Objects." Social Studies of Science, 19(3), 387 through 420.
Stern, N. (2007). The Economics of Climate Change: The Stern Review. Cambridge: Cambridge University Press.
Strogatz, S. H. (1994). Nonlinear Dynamics and Chaos. Reading: Addison-Wesley.
Gregory Caldwell Beier
Susarb LLC, Cambridge, Massachusetts
July 2026