The Rank-One Corner: How Much Value Equivalence Does a Task Need from a World Model?

# How Much Value Equivalence Does a Task Need from a World Model?
Source: [https://arxiv.org/html/2607.06640](https://arxiv.org/html/2607.06640)
Donna Vakalis Mila – Quebec AI Institute Université de Montréal Montréal, Québec donna\.vakalis@olympian\.org

###### Abstract

A learned world model is usually judged by how faithfully it reconstructs its observations or predicts reward, as though quality were something the model simply has or lacks\. But what a task actually needs from a model is narrower: the few predictive coordinates its queries depend on, which we call the*closure*\. We show that how much of that closure a latent comes to represent is set not by the model’s capacity or its observations but by the dimensionality of the objective it is trained against, and we measure this directly on a DreamerV3 stack in a controlled environment whose ground\-truth closure is known\. An aligned scalar value signal—the objective at the heart of value equivalence—installs only a one\-dimensional projection of a closure that needs several dimensions: read through a single linear probe, the recoverable structure rises fromR2=0\.10R^\{2\}=0\.10to0\.760\.76as the scalar is replaced by the full objective\. Sweeping the objective’s dimensionality from one to four installs exactly that many predictive directions through an auxiliary regression head, and the same staircase appears—at attenuated magnitude but the same rank—through the model’s own value head, so the dissociation is dimensional rather than an artifact of head form\. Capacity\-matched comparisons and in\-situ pressure checks rule out the obvious alternatives\. The law governs a regime, and we measure its boundary: on a companion closed\-loop task whose structure is directly observable frame by frame, reconstruction installs that structure and the scalar objective suffices—the objective decides what a latent represents precisely where cheaper training signals cannot already recover it\. Value equivalence is therefore not all\-or\-nothing but dimensional: the familiar single\-reward objective is its rank\-one corner, and a model installs as much of a task’s structure as the objective it is asked to predict\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2607.06640v1/x1.png)Figure 1:Value equivalence is dimensional\.Left: read through one linear probe, an aligned scalar \(single\-reward\) objective installs a fraction0\.100\.10of the query closure, while the full objective installs0\.760\.76in the same configuration\. Right: sweeping the objective’s target dimensionality installs exactly that many closure directions\. The single\-reward objective is the rank\-one corner of this law\.A learned world model is typically trained to reconstruct what it observes or to predict the reward it receives, and its quality is then reported as a single number: a reconstruction error, or a return under planning\. This treats fidelity as something a model either has or lacks\. But a model that reconstructs its inputs perfectly can still be useless for control, and a model that predicts reward well can fail to support the longer\-horizon reasoning a task demands\. What a world model should capture is not its observations in general but the structure its downstream queries actually depend on—the small set of collective coordinates that carry those queries and evolve together under the dynamics\. We call this set the queries’*closure*, and we take model quality to be how much of the closure a trained latent comes to represent\. Ask where the Moon will be at midnight: the closure is six slowly drifting orbital elements, no matter how many pixels of night sky we record\.

The objective a model is trained against is the lever that installs, or fails to install, that structure\. The dominant choice in model\-based reinforcement learning is value equivalence: rather than model the world faithfully, learn a model that predicts value or reward correctly, on the grounds that nothing else matters for acting well\. Value equivalence is attractive precisely because it is low\-dimensional—a scalar value signal is cheap to predict and is aligned with the quantity the agent ultimately cares about\. This raises a question that, to our knowledge, has not been put directly to a trained deep world model: does an aligned scalar objective install the structure a controllable task needs, or only a shadow of it?

We answer this on a DreamerV3 stack in a controlled environment, where a known low\-dimensional process is rendered into high\-dimensional observations so that the ground\-truth closure is available for comparison\. The answer is that a scalar objective installs only a one\-dimensional projection\. Training the latent against an aligned scalar value and then reading the closure out of it with a linear probe recovers a small fraction \(0\.100\.10\) of what the same probe recovers from a latent trained against the full, multi\-dimensional objective \(0\.760\.76\)\. The gap is not an artifact of the probe or the architecture—the network, the probe, and the training budget are identical—but of the objective’s dimensionality\. Sweeping that dimensionality from one to four installs exactly one, two, three, and four predictive directions in turn, and a one\-dimensional objective whose weight is raised to match the four\-dimensional objective’s total training pressure still installs a single direction\. The same staircase appears when the objective is the model’s own value head rather than an auxiliary regression, at attenuated magnitude, so the difference in head form does not account for it\. The structure a latent represents tracks the dimensionality of what it is asked to predict\.

This sharpens, rather than refutes, the value\-equivalence principle\. Proper value equivalence is defined over a*set*of value functions and may be high\-dimensional; what is insufficient is the ubiquitous single\-reward instantiation, the rank\-one corner of a law in which add\-dimensional objective installsddclosure directions\. Read constructively, the result gives a quantitative answer to how much value equivalence a task needs: at least the rank of its closure\. The law also has a boundary, and we measure it rather than assume it\. On a companion closed\-loop control stack whose closure is directly observable frame by frame, a single scalar reward matches a full value family in installed rank and in executed return\. Reconstruction without any reward installs the same closure: where the observations hand a task’s structure to reconstruction, the objective is moot \(Section[9](https://arxiv.org/html/2607.06640#S9)\)\. Objective dimensionality governs the complement—the regime in which reconstruction does not recover the closure—and the environment below is built to sit there\. We make no claim about natural images; the environment is a measurement instrument, built so that a failure to represent the closure is a matter of allocation rather than availability \(Section[2](https://arxiv.org/html/2607.06640#S2)\)\. Our contributions are:

- •a direct demonstration, on a learned deep world model, that the rank a latent installs equals the dimensionality of the objective it is trained against—swept from one to four and reproduced through both an auxiliary head and the model’s own value head;
- •the value\-equivalence rank\-one corner as a within\-system dissociation: an aligned scalar and a full objective, matched in stack, probe, and budget, install0\.100\.10and0\.760\.76of the same closure;
- •a measured boundary for the law: where the closure is frame\-observable, single\-reward value equivalence matches a full value family in rank and return, and reconstruction alone installs the closure; and
- •a reduced\-rank account of the ceiling, sharpened from an optimal\-readout bound to a gradient\-flow prediction, with a pre\-committed falsifier\.

We also connect the training\-side law to evaluation: the discounted Bellman residual decomposes by construction into a reward term and theγ\\gamma\-scaled value\-only operator error, so a Bellman\-residual score reads only the value slice of a model’s dynamics\. We give supporting evidence on released models \(Section[8](https://arxiv.org/html/2607.06640#S8)\) and defer the full treatment to a companion paper\(Vakalis,[2026](https://arxiv.org/html/2607.06640#bib.bib28)\)\.

## 2Closure and the controlled environment

A family of queries closes under the dynamics when a low\-dimensional set of coordinates carries every query in the family and evolves approximately autonomously, so that the queries’ future can be predicted from those coordinates alone\. The closure is the minimal such set, and the question of what a model represents is the question of which closure directions are recoverable from its latent state\.

Reconstruction targets the wrong object\. The dimension a signal occupies in covariance is not the dimension its prediction requires: a single oscillating mode fills one direction of the observation covariance but spans a two\-dimensional predictive subspace, because forecasting it requires both its phase components \(its sine and cosine components—equivalently, position and velocity\)\. A reconstruction or covariance objective therefore systematically under\-sizes the closure \(Figure[2](https://arxiv.org/html/2607.06640#S2.F2)\), which is the basic reason fidelity to the observation does not imply fidelity to the query\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x2.png)Figure 2:Covariance dimension is not closure dimension\.A single oscillatory mode occupies one direction of the observation covariance but spans a two\-dimensional predictive subspace \(position and velocity\), so a reconstruction objective under\-sizes the closure it is meant to capture\.We study this in a controlled environment built on a DreamerV3 categorical\-RSSM stack\(Hafner et al\.,[2023](https://arxiv.org/html/2607.06640#bib.bib16)\)\. A known process ofkkslowly varying latent coordinates is rendered through a fixed analytic warp into a64×6464\\times 64image observation\. Because the construction is ours, the closure rank is known and the latent coordinates are linearly decodable from the observation, atR2≈0\.85R^\{2\}\\approx 0\.85\. We make no natural\-image claim: the environment is a measurement device rather than a perception benchmark, and the linear decodability is what lets us attribute any failure to represent the closure to how the model allocates its latent, not to whether the information is there to be represented\.

## 3Method

We train the world model with its usual reconstruction objective augmented by an auxiliary head that predicts the query target with weightλ\\lambda, and we read what the latent has learned with a linear probe\. Two choices make the readout interpretable\. First, the auxiliary head and the probe both act on the stochastic latent alone; the deterministic recurrent state is excluded, so a query cannot be satisfied through a pathway the latent itself does not carry\. Second, we summarize what is represented by the*installed rank*: the number of closure coordinates a linear probe recovers from the latent above a threshold set by a reconstruction\-only null \(the per\-column null mean plus2\.52\.5standard deviations, fixed post hoc; the readouts are unchanged across the\+2\+2to\+3​σ\+3\\sigmaband\)\. Alongside the installed rank we report the total recoverable structure \(R2R^\{2\}of the full closure from the latent\) and a leakage check verifying that the closure is not instead readable from the excluded recurrent state\. Figure[3](https://arxiv.org/html/2607.06640#S3.F3)summarizes the instrument\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x3.png)Figure 3:The measurement instrument\.A known process ofkkslow latent coordinatesLtL\_\{t\}is rendered through a fixed analytic warp into a64×6464\\times 64image observation alongside a high\-variance distractorDtD\_\{t\}; a DreamerV3 categorical RSSM \(deterministichth\_\{t\}, stochasticztz\_\{t\}\) trains on reconstruction, which readshhandzz, plus a query head of weightλ\\lambdathat readszzonly, withhhexcluded from its forward pass\. Whatzzholds is read by a held\-out linear probe \(installed rank against a reconstruction\-only null threshold\), with a leakage check on the excludedhh\. BecauseLLis linearly decodable from the observation \(R2≈0\.85R^\{2\}\\approx 0\.85\), a failure to install it inzzis allocation, not availability\.
## 4The objective, not reconstruction, determines the representation

Before asking which objectives install the closure, we establish that an objective installs it at all\. We place a high\-variance distractor in the observation alongside the low\-variance query coordinates, so that a model minimizing reconstruction error is drawn to the distractor and away from the query\. An auxiliary head allowed to read both the latent and the recurrent state leaves the query unrecoverable from the latent—but this is an artifact of architecture, not of the objective, because the head routes the query through the recurrent state and drains the latent\. Once the head is forced onto the latent alone, the picture reverses \(Figure[4](https://arxiv.org/html/2607.06640#S4.F4)\)\. A query\-aligned objective installs the closure in the latent even under the distractor, with recovery climbing as the objective weight increases \(from0\.550\.55to0\.880\.88\) and saturating at the level to which the closure is linearly present in the observation, while reconstruction remains intact\. The dependence is causal: shuffling the query labels during training collapses recovery to zero while the distractor stays readable, so it is the objective, not an incidental correlate, that determines what the latent holds\. A recovery gate fixed in advance confirmed that the latent could represent the closure before we measured how strongly the objective installs it\. This determination is exercised only where the environment forces it: where reconstruction already recovers the closure, the objective has nothing left to decide—the boundary we quantify in Section[9](https://arxiv.org/html/2607.06640#S9)\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x4.png)Figure 4:The objective, not reconstruction, determines what the latent represents\.When the query is forced through the stochastic latent, a query\-aligned objective installs the closure under a high\-variance distractor—recovery rises with the objective weightλ\\lambdaand saturates at the linear availability ceiling, while reconstruction stays intact\. A train\-time label shuffle abolishes the effect with the distractor still recoverable, establishing that the dependence is causal\.
## 5Value equivalence is the rank\-one corner

### 5\.1An aligned scalar objective installs a single dimension

We now replace the multi\-dimensional query target with an aligned*scalar*: a single value/reward signal of exactly the kind value equivalence prescribes, predicted through the model’s value head\. Although the scalar is perfectly aligned with the task, the latent it installs contains almost none of the query closure\. A linear probe recoversR2=0\.10R^\{2\}=0\.10of the full closure, far below the0\.370\.37install floor and well below the0\.760\.76the same probe recovers from a latent trained on the full objective in the same configuration \(k=4k=4,nz=4n\_\{z\}=4,λ=1\\lambda=1in both arms; Figure[5](https://arxiv.org/html/2607.06640#S5.F5)\)\.111Both arms are summarized over two training seeds\. We quote the scalar’s stronger seed \(0\.100\.10; the second installs0\.060\.06\) and the full arm’s two\-seed mean \(0\.760\.76, from0\.780\.78and0\.750\.75\); reported on a single convention—the two\-seed mean—the pair is0\.080\.08versus0\.760\.76, so quoting the scalar’s stronger seed is the conservative choice for the dissociation\. The0\.370\.37floor is the sum of a measured quantity and a pre\-committed one: the maximum recurrent\-state \(reward\-history\) leakage,*measured on the scalar arm itself*across decode windows \(0\.220\.22, peaking at a one\-step window\), plus a margin of0\.150\.15fixed in advance\.The scalar does install its own one\-dimensional projection \(0\.490\.49\), so the head is working as intended, and a placebo objective built from the distractor installs the distractor \(0\.870\.87\) but not the query—so the head is capable of installing structure when the structure it is asked for is present\. The shortfall is specific to the scalar’s dimensionality, not a failure of the probe or the head\. In plain terms: the model asked to predict value learned the value—and almost nothing else the task depends on\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x5.png)Figure 5:An aligned scalar objective installs a single dimension\.Through one linear probe, the scalar value objective recovers0\.100\.10of the closure while the full objective recovers0\.760\.76in the same configuration; the scalar installs its own one\-dimensional projection, and a placebo objective installs the distractor but not the query, so the head installs what it is asked to install\.
### 5\.2Installed rank tracks objective dimensionality

The scalar’s failure is the lower end of a graded law\. Holding the architecture and the head fixed, we vary only the dimensionalityddof the objective’s target from one to four\. The number of closure directions installed in the latent follows exactly: one, two, three, and four \(Figure[6](https://arxiv.org/html/2607.06640#S5.F6)\)\. The effect is dimensionality, not training pressure: raising a one\-dimensional objective’s weight 4\.5\-fold—to the four\-dimensional objective’s total target variance—still installs a single direction\. The ramp does strengthen that one direction \(its recovery rises to0\.920\.92\) but recruits no second: the unsupervised directions stay at the reconstruction null, so what tracks dimensionality is the*number*of installed directions, not their magnitude\. The latent holds as many directions as the objective asks for—no more, and asking harder for one does not buy a second\. The staircase reproduces across seeds at every step—three seeds at the interiord=2,3d=2,3, two at the endpoints, and every seed lands on installed rankdd\. Reduced\-rank regression supplies the reason \(Section[6](https://arxiv.org/html/2607.06640#S6)\): the rank of the structure an objective can install is bounded by the rank of its target, so add\-dimensional objective installs at mostdddirections, and empirically it installs exactlydd\. Single\-reward value equivalence sits atd=1d=1; the closure of a controllable task generally does not\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x6.png)Figure 6:Objective dimensionality sets the installed closure rank\.\(a, b\)Varying only the objective’s target dimensionality from one to four installs one, two, three, and four closure directions: \(a\) the per\-direction install matrix and \(b\) the seed\-replicated staircase, installed rank=d=dat every seed\. A one\-dimensional objective raised to4\.5×4\.5\\timesweight still installs a single direction and a label shuffle installs none, so the rank tracks dimensionality rather than gradient pressure; the per\-column threshold is not load\-bearing, as supervised and unsupervised columns are separated by a wide gap \(Figure[12](https://arxiv.org/html/2607.06640#A2.F12)\)\. The boundinstalled​rank≤d\\mathrm\{installed\\ rank\}\\leq dis a reduced\-rank\-regression corollary; the attainment=d=dis empirical\.\(c\)The same law read through the model’s own value head \(Section[5\.3](https://arxiv.org/html/2607.06640#S5.SS3)\): add\-dimensional value/reward family installs rankddon the seed\-mean staircase, with threshold\-free total recoveryR2​\(full​L←z\)=0\.11,0\.27,0\.47,0\.55R^\{2\}\(\\text\{full \}L\\\!\\leftarrow z\)=0\.11,0\.27,0\.47,0\.55strictly monotone indd, at about0\.7×0\.7\\timesthe auxiliary\-head magnitude—a head\-type effect that leaves the rank intact\. The law is claimed on this mean staircase and overlay, not on per\-seed counts; caveats \(thed=2d=2threshold fragility and the sweep’s pressure ramp\) in Section[5\.3](https://arxiv.org/html/2607.06640#S5.SS3)\.
### 5\.3The same law through the value head

The sweep above varies dimensionality through an auxiliary regression head, whose form differs from the value head that value equivalence actually uses\. To remove that difference we repeat the sweep through the model’s own value head, supervising add\-dimensional family of value/reward functions—the set over which proper value equivalence is defined—and reading the installed rank as before\. The law survives the change of head\. Averaged over seeds \(three atd=2d=2, two at each otherdd\), add\-dimensional value objective installsddclosure directions forddfrom one to four, and the rotation\-invariant total recovery climbs monotonically withdd\(0\.110\.11,0\.270\.27,0\.470\.47,0\.550\.55\), each dimension’s readout band separated from the next\. Because every cell of this sweep is installed by the same value head, head type no longer co\-varies with dimensionality—the confound the single matched comparison of Section[5\.4](https://arxiv.org/html/2607.06640#S5.SS4)could not fully remove\.

The recovered magnitudes are about0\.7×0\.7\\timesthe regression head’s at matcheddd\. The value head also warps its target throughtanh\\tanh, but on these slowly varying, near\-unit\-variance coordinates the warp is near\-linear \(a linear fit recoversR2≈0\.93R^\{2\}\\approx 0\.93of the target variance\), so this attenuation is attributable to head type, not to a distinct nonlinear target encoding—and it leaves the rank intact\. The per\-seedd=2d=2count is threshold\-fragile, so the law is stated on the three\-seed mean and the total\-recovery overlay rather than a per\-seed count\. The remaining caveats—the magnitude attenuation, the borderlined=2d=2coordinate, and the sweep’s uncontrolled aggregate pressure—are detailed in Appendix[B](https://arxiv.org/html/2607.06640#A2)\.

### 5\.4A matched\-dimension control

The scalar and the full objective compared above differ not only in dimensionality but in the form of their prediction head and target encoding\. To isolate dimensionality from these, we compare, at identical task dimension and capacity, a full multi\-dimensional target against the scalar: the former installs the closure \(0\.760\.76, a wide margin above the leakage ceiling222In this matched pair the closure is marginally readable from the excluded recurrent state \(R2​\(L←h\)≈0\.04R^\{2\}\(L\\\!\\leftarrow\\\!h\)\\approx 0\.04, just over our0\.030\.03purity bar—a ridge\-decode artifact arising where the target dimension approaches the probe’s effective degrees of freedom\); the sweep cells of Section[5\.2](https://arxiv.org/html/2607.06640#S5.SS2)satisfy the bar\.\) and the latter does not \(0\.100\.10\)\. This is consistent with the dimensionality reading and removes the cross\-task confound\. Head type and target encoding still co\-vary within this single comparison, but the head\-type difference is separated directly by the value\-head sweep of Section[5\.3](https://arxiv.org/html/2607.06640#S5.SS3), which installs rankddthrough the same value head at each interior dimension; the matched pair is retained as same\-regime support, no longer the load\-bearing evidence for that separation\.

## 6Why dimensionality bounds the representation

The ceiling is reduced\-rank regression in another guise\. If the objective’s target spansdddimensions, the cross\-covariance between target and latent has rank at mostdd, so the population\-optimal linear predictor occupies at mostdddirections of the latent, and*which*dddirections—the leading canonical directions of the target–latent cross\-covariance—is fixed by the same canonical\-correlation truncation that underlies our closure spectrum\(Eckart and Young,[1936](https://arxiv.org/html/2607.06640#bib.bib9); Anderson,[1951](https://arxiv.org/html/2607.06640#bib.bib3); Izenman,[1975](https://arxiv.org/html/2607.06640#bib.bib19)\)\. The bound is independent of how much capacity the latent carries: a wider latent does not install more of the objective’s structure, only the samedddirections\. This is the objective\-side counterpart of value equivalence\(Grimm et al\.,[2020](https://arxiv.org/html/2607.06640#bib.bib14)\)—the objective’s rank, not the model’s size, caps the closure a latent installs\.

The bound constrains the optimal readout; what training installs is a separate question, and in the linear case gradient descent answers it\. For a linear encoder and linear head trained by gradient flow on the query loss, the coupled dynamics learn the modes of the target–latent cross\-covariance one at a time, in order of their canonical correlation, converging to the rank\-ddreduced\-rank regressor—the top\-ddcanonical directions and no more\(Baldi and Hornik,[1989](https://arxiv.org/html/2607.06640#bib.bib4); Saxe et al\.,[2014](https://arxiv.org/html/2607.06640#bib.bib24)\)\. So attainment is not, in this case, a separate empirical fact but the gradient\-flow prediction: training installs*exactly*dddirections whenever the target covariance has rankddand allddcanonical correlations are strictly positive—a zero canonical correlation is a target direction the objective cannot reach, and would lower the count\. This turns the earlier “≤d\\leq dbound,=d=dempirical” into “=d=dis the gradient\-flow prediction,” which the experiments then verify one architecture up\.

We borrow this rigor rather than claim it, and two fences keep the borrowing honest\. First, the theory is linear—both the≤d\\leq dreadout bound and the=d=dgradient\-flow attainment are proved for a linear encoder and head, whereas every model here uses a nonlinear head; the transfer is to the rank a*linear probe*can read, and the gradient\-flow prediction is one we*verify*in the nonlinear stack rather than inherit, under a no\-spillover regularity—unsupervised target coordinates stay at the reconstruction null and a label shuffle collapses the readout—that stands in for the linear theory’s guarantees\. Second, the capacity\-independence is demonstrated synthetically; on the trained pixel stack the analogous install fails to generalize across distributions for reasons orthogonal to this bound \(Section[9](https://arxiv.org/html/2607.06640#S9)\)\.

The same algebra suggests a falsifiable companion\. Under a shared capacity budget the model must allocate latent directions between reconstruction and the query, and the reverse\-water\-filling intuition of rate–distortion theory\(Berger,[1971](https://arxiv.org/html/2607.06640#bib.bib6)\)predicts a greedy, marginal\-value ordering: tight capacity makes the query*displace*reconstruction, abundant capacity makes it*add*to it\. We state this as a prediction with a pre\-committed falsifier—the installed rank must plateau atddwith shielded coordinates at the null, and is refuted by a continued rise or a genuinely installed extra direction—and we do not derive the threshold itself \(Appendix[C](https://arxiv.org/html/2607.06640#A3)\)\. The same lens links the two sides of the paper: when the objective is value alone, the Bellman residual decomposes by construction into a reward term and theγ\\gamma\-scaled value\-only operator error, so value equivalence is the low\-dimensional slice on the evaluation side as well \(Section[8](https://arxiv.org/html/2607.06640#S8)\)\. The contribution throughout is the law’s behavior in a trained deep model and its falsifiable form, not a new theorem\. A stronger, ordinal reading of the same heuristic—installation*in proportion to*marginal value, not merely up to a count—survives only in part when tested one level below this ceiling: on a trained pixel stack of four graded\-predictability targets the converged install*magnitude*runs counter to the predictability order across three seeds \(seed\-mean rank correlation−0\.56\-0\.56; the second\-most\-predictable direction installs weakest in every seed\), while install*timing*partially recovers it\. This bounds the magnitude refinement of the allocation heuristic only; the rank ceiling, its attainment, and the confirmed rank\-ramp falsifier are untouched \(Appendix[C\.4](https://arxiv.org/html/2607.06640#A3.SS4)\)\.

## 7Calibration and a prospective test

Two checks anchor the readout against ground truth\. First, in a continuous\-latent version of the environment, the installed rank rises with capacity and then plateaus at the objective dimensiondd, holding through a latent several times larger than the closure; the plateau level tracksdd, not capacity, with the exact knee tolerance\-relative\. A categorical latent of the kind the main stack uses instead saturates at a fixed level regardless of size, as*predicted by*the rate–rank distinction:nzn\_\{z\}categorical variables overCCclasses carry aboutnz​log2⁡Cn\_\{z\}\\log\_\{2\}Cbits and can spread partial information over more thannzn\_\{z\}closure directions, so capacity binds as a rate, whereas a continuous latent of widthnzn\_\{z\}binds as a rank ofnzn\_\{z\}\(Appendix[C\.3](https://arxiv.org/html/2607.06640#A3.SS3); Figure[7](https://arxiv.org/html/2607.06640#S7.F7)\)\. This was registered as a prediction of the dimensionality law before it was run, and it is distinct from the trained stack’s continuous\-latent variant, whose generalization failure is reported in Section[9](https://arxiv.org/html/2607.06640#S9)\. Second, on planted environments whose closure rank is known by construction, a trained model’s minimal sufficient latent tracks that rank—exactly on the simplest family and monotonically once the observation map is varied \(Figure[8](https://arxiv.org/html/2607.06640#S7.F8)\)\. Both are synthetic calibrations and are labeled as such; the capacity term in particular is established only synthetically\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x7.png)Figure 7:Capacity is a rate, not a rank, for a categorical latent\.In a continuous\-latent version of the environment the installed rank rises with capacity and plateaus at the objective dimensiondd, whereas a categorical latent saturates at a fixed level regardless of size\. Synthetic calibration\.![Refer to caption](https://arxiv.org/html/2607.06640v1/x8.png)Figure 8:A trained latent tracks a known closure rank\.On planted environments with closure rank known by construction, a trained model’s minimal sufficient latent tracks that rank—exactly on the simplest family and monotonically as the observation map is varied\. Synthetic calibration\.
## 8The evaluation side

The same low\-dimensional character appears on the evaluation side, where a model is judged by how it rolls forward\. The discounted Bellman residual decomposes by construction into a reward term and theγ\\gamma\-scaled value\-only operator error \(Appendix[C\.7](https://arxiv.org/html/2607.06640#A3.SS7)\), so scoring a model by its Bellman residual reads only the value slice of its dynamics—and a full operator that also tracks non\-value structure can pull apart from it\. On released TD\-MPC2 checkpoints\(Hansen et al\.,[2024](https://arxiv.org/html/2607.06640#bib.bib17)\)the reward error is two orders of magnitude below the value term, and full\-operator error tracks executed return \(Spearman=−0\.90\\mathrm\{Spearman\}=\-0\.90;n=5n=5sizes on one task, correlational\) while reward\-prediction error does not \(Figure[9](https://arxiv.org/html/2607.06640#S8.F9)\)\. This is the evaluation\-side echo of the training\-side rank\-one install; we present it as supporting evidence and defer the full cross\-architecture treatment—a different axis, operator fidelity rather than representational content—to a companion paper\(Vakalis,[2026](https://arxiv.org/html/2607.06640#bib.bib28)\)\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x9.png)Figure 9:Value equivalence is the low\-dimensional slice on the evaluation side\.On released TD\-MPC2 models the full operator error tracks executed return \(a;Spearman=−0\.90\\mathrm\{Spearman\}=\-0\.90\) while reward\-prediction error stays in a narrow band that does not \(b\)\.n=5n=5checkpoints, single environment, correlational\.
## 9Scope and limitations

#### Where the objective matters\.

The law’s regime is not universal, and we measured its boundary directly rather than assuming it\. On a closed\-loop control stack whose four\-coordinate closure is directly observable frame by frame \(a single\-frame decode reachesR2≈0\.998R^\{2\}\\approx 0\.998per coordinate\), single\-reward value equivalence is indistinguishable from a full value family in installed rank—posterior and open\-loop rollout alike—and in executed return, and an action\-masked reconstruction arm installs the same closure with no reward signal at all \(Figure[10](https://arxiv.org/html/2607.06640#S9.F10)a; Appendix[B](https://arxiv.org/html/2607.06640#A2)\)\. Where the observations hand the closure to reconstruction, the objective is moot\. The corollary is a bounded sufficiency result for practice: when a task’s return\-relevant closure is frame\-observable, single\-reward value equivalence suffices in closed loop\. The rank\-one corner is a risk for passive prediction and for tasks whose closure is not observable, not a universal control deficiency\. Where the screen already shows the state, any objective—or none—will do\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x10.png)Figure 10:The scope boundary, measured\.\(a\)On a closed\-loop control task whose four\-coordinate closure is observable in every frame \(single\-frame decodeR2≈0\.998R^\{2\}\\approx 0\.998per coordinate\), arms differing only in objective—single\-reward value equivalence, a full value family, action\-masked reconstruction with no reward—install the same closure at the same rank, and the scalar arm’s executed return matches the family’s under common random numbers on every seed: the objective makes no difference here\.\(b\)When a cue is visible only transiently, the free reconstruction\-plus\-recurrence carry decays from near\-perfect at lag55to chance by lag4545on all three seeds \(gray: no objective; amber: an objective whose training window does not span the reveal\-to\-use gap\), while the arm whose objective and window span the gap stays at ceiling at every lag on all three seeds \(teal\)\. The durable arm is a supervised, ceiling\-saturated install \(durable to at least lag4545; its true decay rate is unmeasured\), and the contrast separates supervised from passive channels, not a distinct “value channel”; a pre\-registered single\-lag binary read of these data was seed\-fragile and was retired, disclosed in Appendix[B](https://arxiv.org/html/2607.06640#A2)\. Uniform\-CPU reads\.
#### The cheapest\-channel ordering\.

The boundary appears to generalize by training channel\. A coordinate that is only*transiently*observable is also installed without objective pressure: the recurrent state carries it to its point of use, and this free carry decays to chance within about forty\-five steps while a supervised, objective\-installed coordinate stays durably at ceiling over the same lags—a threshold\-free contrast that replicates on all three seeds \(Figure[10](https://arxiv.org/html/2607.06640#S9.F10)b; Appendix[B](https://arxiv.org/html/2607.06640#A2)\)\. This suggests an ordering with a common currency\. Intuitively: if the model can see a coordinate, reconstruction learns it; if it just saw it, recurrence carries it; if it can infer it from how observations respond to actions, filtering completes it; only what none of these reach must be paid for by the objective\. Formally, fix the data\-collecting policyπ\\pi, and say a coordinate is*reached*by a channel when it is measurable with respect to that channel’s information set\. These sets are nested—the current\-frameσ\\sigma\-algebra \(what reconstruction sees\)⊆\\subseteqthe historyσ\\sigma\-algebra \(what recurrence carries\)⊆\\subseteqthe history closed under one\-step dynamics inversion \(what a recurrent filter can complete\)—so a coordinate installs through the cheapest channel that reaches it: reconstruction, then recurrence, then filtering, with the objective the channel of last resort\. Objective dimensionality then gates representation exactly for coordinates that are return\-relevant but measurable with respect to none of the three passive sets \(return\-relevant, unobservable, and unfilterable\) \(Figure[11](https://arxiv.org/html/2607.06640#S9.F11)\)\. We state the channel\-level observations \(reconstruction and recurrence install for free where they reach\) as replicated, and the ordering itself as a synthesis those observations support but do not yet establish\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x11.png)Figure 11:When the objective matters\.A coordinate installs through the cheapest training channel that reaches it: reconstruction if it is observable in the frame, recurrence if it is transiently observable, a filter over the action–response \(proposed; not directly tested here\), and the objective as the channel of last resort—so objective dimensionality governs representation exactly for coordinates that are return\-relevant, unobservable, and unfilterable, which is this paper’s regime\. The channel\-level entries are measured \(Figure[10](https://arxiv.org/html/2607.06640#S9.F10)\); the*ordering*is a proposed synthesis they support, not an established law\. The corner that no passive channel reaches is empty exactly when the data policy’s observations already excite every return\-relevant direction, and open where they do not \(dual control\); we claim nothing about it here\.The sharpest question the ordering poses is whether any coordinate exists that*only*the objective can install\. Under the fixed policyπ\\pi, a coordinate is passively reachable exactly when it is measurable with respect toΣobs​\(π\)\\Sigma\_\{\\mathrm\{obs\}\}\(\\pi\), the observationσ\\sigma\-algebra generated underπ\\pi\(the union of the three nested passive sets above\)\. The corner is therefore empty precisely when every return\-relevant direction lies in that set—the completeness statementΣval⊆Σobs​\(π\)\\Sigma\_\{\\mathrm\{val\}\}\\subseteq\\Sigma\_\{\\mathrm\{obs\}\}\(\\pi\)—and this is a stronger condition than certainty equivalence: a return\-relevant coordinate thatπ\\pi’s observations never excite is unreachable by any passive channel regardless of whether the value function factors through the belief\. Where completeness fails—dual control, where acting to gather information has value the passive prediction problem never sees—the corner can be non\-empty; that edge is open, and we claim nothing about it here\.

#### Limitations\.

Our evidence comes from a single architecture, a single environment family, and a synthetic observation, chosen so that ground truth is available; we do not claim the law holds unchanged on natural images, and the capacity term is established only synthetically\. The continuous\-latent variant carries an honest negative: the closure it installs is genuine on the training distribution but does not transfer to fresh rollouts, even though the closure remains linearly present in those rollouts—a generalization failure we have not been able to attribute to a mechanism\. On a public world\-model benchmark\(Warrier et al\.,[2025](https://arxiv.org/html/2607.06640#bib.bib30)\), our linear closure estimator reaches only about a third of the relevant structure and the test is underpowered at the sixteen available environments, so we report it as a boundary on the instrument rather than a test the thesis passes or fails\.

## 10Related work

Value equivalence and its proper, set\-valued refinement frame our question\(Grimm et al\.,[2020](https://arxiv.org/html/2607.06640#bib.bib14),[2021](https://arxiv.org/html/2607.06640#bib.bib15)\); we sharpen rather than overturn them\. Value\-aware model learning derives model objectives from the value functions they must serve\(Farahmand et al\.,[2017](https://arxiv.org/html/2607.06640#bib.bib11); Farahmand,[2018](https://arxiv.org/html/2607.06640#bib.bib10); Abachi et al\.,[2020](https://arxiv.org/html/2607.06640#bib.bib1)\), and MuZero is the scaled instance of training a model only on reward, value, and policy\(Schrittwieser et al\.,[2020](https://arxiv.org/html/2607.06640#bib.bib25)\); our result measures how much of a task’s structure that family of objectives installs, as a function of its dimensionality\.

On the representation side, auxiliary prediction tasks have long been used to shape what an agent’s features carry\(Jaderberg et al\.,[2017](https://arxiv.org/html/2607.06640#bib.bib20)\), the geometry of a set of value functions is known to determine the representation they induce\(Bellemare et al\.,[2019](https://arxiv.org/html/2607.06640#bib.bib5); Dabney et al\.,[2021](https://arxiv.org/html/2607.06640#bib.bib8)\), and task\-informed world models discard observation structure a target deems irrelevant\(Gelada et al\.,[2019](https://arxiv.org/html/2607.06640#bib.bib13); Fu et al\.,[2021](https://arxiv.org/html/2607.06640#bib.bib12); Wang et al\.,[2022](https://arxiv.org/html/2607.06640#bib.bib29)\); what we add is the closure as the reference object and the rank law for how much of it an objective of a given dimensionality installs\.

The bound we use is reduced\-rank regression\(Anderson,[1951](https://arxiv.org/html/2607.06640#bib.bib3); Izenman,[1975](https://arxiv.org/html/2607.06640#bib.bib19); Eckart and Young,[1936](https://arxiv.org/html/2607.06640#bib.bib9)\), and the closure object is the stochastic\-realization and canonical\-correlation reading of minimal predictive state\(Akaike,[1975](https://arxiv.org/html/2607.06640#bib.bib2); Larimore,[1990](https://arxiv.org/html/2607.06640#bib.bib21); Moore,[1981](https://arxiv.org/html/2607.06640#bib.bib23)\), with predictive\-state representations supplying the rank side\(Littman et al\.,[2001](https://arxiv.org/html/2607.06640#bib.bib22); Singh et al\.,[2004](https://arxiv.org/html/2607.06640#bib.bib26)\)\. Our contribution is empirical—a demonstrated mechanism in a learned deep world model, and a falsifiable form of it—rather than a new theorem\.

## 11Conclusion

In the regime where reconstruction does not recover a task’s closure, what a world model represents is set by the dimensionality of the objective it is trained against—not by its capacity, its observations, or the magnitude of a scalar reward; where reconstruction does recover it, the objective is moot, a boundary we measured rather than assumed\. Value equivalence is dimensional, and its familiar single\-reward form is the rank\-one corner of that law\. How much value equivalence a task needs is, to first order, the rank of its closure—and whether it needs any depends on whether that closure is already in view\. The Moon we began with was, by this light, the easy case: all six of its coordinates are offered to any patient observer, no objective required\. The rank\-one corner is reserved for the structure the sky keeps to itself\.

## Acknowledgments and Disclosure of Funding

We gratefully acknowledge financial support from the Canada CIFAR AI Chairs program, and compute resources from Mila \(mila\.quebec\)\.

## References

- Abachi et al\. \[2020\]R\. Abachi, M\. Ghavamzadeh, and A\. Farahmand\.Policy\-aware model learning for policy gradient methods, 2020\.
- Akaike \[1975\]H\. Akaike\.Markovian representation of stochastic processes by canonical variables\.*SIAM Journal on Control*, 13\(1\):162–173, 1975\.doi:10\.1137/0313010\.
- Anderson \[1951\]T\. W\. Anderson\.Estimating linear restrictions on regression coefficients for multivariate normal distributions\.*Annals of Mathematical Statistics*, 22\(3\):327–351, 1951\.doi:10\.1214/aoms/1177729580\.
- Baldi and Hornik \[1989\]P\. Baldi and K\. Hornik\.Neural networks and principal component analysis: Learning from examples without local minima\.*Neural Networks*, 2\(1\):53–58, 1989\.
- Bellemare et al\. \[2019\]M\. G\. Bellemare, W\. Dabney, R\. Dadashi, A\. A\. Taiga, P\. S\. Castro, N\. L\. Roux, D\. Schuurmans, T\. Lattimore, and C\. Lyle\.A geometric perspective on optimal representations for reinforcement learning\.In*Advances in Neural Information Processing Systems 32 \(NeurIPS\)*, 2019\.
- Berger \[1971\]T\. Berger\.*Rate Distortion Theory: A Mathematical Basis for Data Compression*\.Prentice\-Hall, Englewood Cliffs, NJ, 1971\.
- Brillinger \[1969\]D\. R\. Brillinger\.The canonical analysis of stationary time series\.In P\. R\. Krishnaiah, editor,*Multivariate Analysis – II*, pages 331–350\. Academic Press, New York, 1969\.
- Dabney et al\. \[2021\]W\. Dabney, A\. Barreto, M\. Rowland, R\. Dadashi, J\. Quan, M\. G\. Bellemare, and D\. Silver\.The value\-improvement path: Towards better representations for reinforcement learning\.In*AAAI Conference on Artificial Intelligence*, 2021\.
- Eckart and Young \[1936\]C\. Eckart and G\. Young\.The approximation of one matrix by another of lower rank\.*Psychometrika*, 1\(3\):211–218, 1936\.doi:10\.1007/BF02288367\.
- Farahmand \[2018\]A\. Farahmand\.Iterative value\-aware model learning\.In*Advances in Neural Information Processing Systems 31 \(NeurIPS\)*, pages 9072–9083, 2018\.
- Farahmand et al\. \[2017\]A\. Farahmand, A\. Barreto, and D\. Nikovski\.Value\-aware loss function for model\-based reinforcement learning\.In*Proceedings of the 20th International Conference on Artificial Intelligence and Statistics \(AISTATS\)*, volume 54 of*PMLR*, pages 1486–1494, 2017\.
- Fu et al\. \[2021\]X\. Fu, G\. Yang, P\. Agrawal, and T\. Jaakkola\.Learning task informed abstractions\.In*International Conference on Machine Learning \(ICML\)*, 2021\.
- Gelada et al\. \[2019\]C\. Gelada, S\. Kumar, J\. Buckman, O\. Nachum, and M\. G\. Bellemare\.DeepMDP: Learning continuous latent space models for representation learning\.In*Proceedings of the 36th International Conference on Machine Learning \(ICML\)*, volume 97 of*PMLR*, pages 2170–2179, 2019\.
- Grimm et al\. \[2020\]C\. Grimm, A\. Barreto, S\. Singh, and D\. Silver\.The value equivalence principle for model\-based reinforcement learning\.In*Advances in Neural Information Processing Systems 33 \(NeurIPS\)*, 2020\.
- Grimm et al\. \[2021\]C\. Grimm, A\. Barreto, G\. Farquhar, D\. Silver, and S\. Singh\.Proper value equivalence\.In*Advances in Neural Information Processing Systems 34 \(NeurIPS\)*, 2021\.
- Hafner et al\. \[2023\]D\. Hafner, J\. Pasukonis, J\. Ba, and T\. Lillicrap\.Mastering diverse domains through world models, 2023\.
- Hansen et al\. \[2024\]N\. Hansen, H\. Su, and X\. Wang\.TD\-MPC2: Scalable, robust world models for continuous control\.In*International Conference on Learning Representations \(ICLR\)*, 2024\.
- Hsu et al\. \[2012\]D\. Hsu, S\. M\. Kakade, and T\. Zhang\.A spectral algorithm for learning hidden Markov models\.*Journal of Computer and System Sciences*, 78\(5\):1460–1480, 2012\.doi:10\.1016/j\.jcss\.2011\.12\.025\.Conference version in COLT 2009\.
- Izenman \[1975\]A\. J\. Izenman\.Reduced\-rank regression for the multivariate linear model\.*Journal of Multivariate Analysis*, 5\(2\):248–264, 1975\.doi:10\.1016/0047\-259X\(75\)90042\-1\.
- Jaderberg et al\. \[2017\]M\. Jaderberg, V\. Mnih, W\. M\. Czarnecki, T\. Schaul, J\. Z\. Leibo, D\. Silver, and K\. Kavukcuoglu\.Reinforcement learning with unsupervised auxiliary tasks\.In*International Conference on Learning Representations \(ICLR\)*, 2017\.
- Larimore \[1990\]W\. E\. Larimore\.Canonical variate analysis in identification, filtering, and adaptive control\.In*Proceedings of the 29th IEEE Conference on Decision and Control \(CDC\), Honolulu*, pages 596–604, 1990\.doi:10\.1109/CDC\.1990\.203665\.
- Littman et al\. \[2001\]M\. L\. Littman, R\. S\. Sutton, and S\. Singh\.Predictive representations of state\.In*Advances in Neural Information Processing Systems 14 \(NeurIPS\)*, pages 1555–1561, 2001\.
- Moore \[1981\]B\. C\. Moore\.Principal component analysis in linear systems: controllability, observability, and model reduction\.*IEEE Transactions on Automatic Control*, 26\(1\):17–32, 1981\.doi:10\.1109/TAC\.1981\.1102568\.
- Saxe et al\. \[2014\]A\. M\. Saxe, J\. L\. McClelland, and S\. Ganguli\.Exact solutions to the nonlinear dynamics of learning in deep linear neural networks\.In*International Conference on Learning Representations \(ICLR\)*, 2014\.
- Schrittwieser et al\. \[2020\]J\. Schrittwieser, I\. Antonoglou, T\. Hubert, K\. Simonyan, L\. Sifre, S\. Schmitt, A\. Guez, E\. Lockhart, D\. Hassabis, T\. Graepel, T\. Lillicrap, and D\. Silver\.Mastering Atari, Go, chess and shogi by planning with a learned model\.*Nature*, 588\(7839\):604–609, 2020\.doi:10\.1038/s41586\-020\-03051\-4\.
- Singh et al\. \[2004\]S\. Singh, M\. R\. James, and M\. R\. Rudary\.Predictive state representations: A new theory for modeling dynamical systems\.In*Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence \(UAI\)*, pages 512–518, 2004\.
- Tassa et al\. \[2018\]Y\. Tassa, Y\. Doron, A\. Muldal, T\. Erez, Y\. Li, D\. de Las Casas, D\. Budden, A\. Abdolmaleki, J\. Merel, A\. Lefrancq, T\. Lillicrap, and M\. Riedmiller\.DeepMind Control Suite, 2018\.
- Vakalis \[2026\]D\. Vakalis\.Operator\-on\-F complements value\-equivalence: A planning\-time diagnostic for latent world models\.In*Reinforcement Learning Conference 2026 Workshop on Model\-based RL in the Era of Generative World Models*, 2026\.URL[https://arxiv\.org/abs/2607\.04464](https://arxiv.org/abs/2607.04464)\.
- Wang et al\. \[2022\]T\. Wang, S\. S\. Du, A\. Torralba, P\. Isola, A\. Zhang, and Y\. Tian\.Denoised MDPs: Learning world models better than the world itself\.In*International Conference on Machine Learning \(ICML\)*, 2022\.
- Warrier et al\. \[2025\]A\. Warrier, D\. Nguyen, M\. Naim, M\. Jain, Y\. Liang, K\. Schroeder, C\. Yang, J\. B\. Tenenbaum, S\. Vollmer, K\. Ellis, and Z\. Tavares\.Benchmarking world\-model learning with environment\-level queries, 2025\.

## Appendix AMethod and readout conventions

We train a DreamerV3 categorical\-RSSM world model\[Hafner et al\.,[2023](https://arxiv.org/html/2607.06640#bib.bib16)\]with its standard reconstruction objective augmented by an auxiliary query head of weightλ\\lambda, and we read what the latent has learned with a linear probe applied after training\. Both the auxiliary head and the probe act on the stochastic latentzzalone: the deterministic recurrent statehhis excluded from the head’s forward pass and from the probe, so a query cannot be satisfied through a pathway the latent itself does not carry\. This is the single most consequential design choice—a head free to read\[h,z\]\[h,z\]routes the query throughhhand leaveszzholding the distractor, an architecture artifact documented in Appendix[B](https://arxiv.org/html/2607.06640#A2)\.

We summarize what a latent represents by its*installed rank*: the number of closure coordinates a linear probe recovers fromzzabove a per\-column threshold\. The probe is a ridge regression with held\-out regularization; for each closure coordinateLiL\_\{i\}we report the held\-out, gap\-correctedR2​\(Li←z\)R^\{2\}\(L\_\{i\}\\\!\\leftarrow z\)\(recovery in excess of a same\-capacity baseline\), and count a coordinate as installed when this exceedsτcol\\tau\_\{\\mathrm\{col\}\}, the reconstruction\-only per\-column null mean plus2\.52\.5standard deviations\. The anchor is fixed by inspecting the null pool rather than pre\-registered, and the installed\-rank readouts are unchanged across the\+2\+2to\+3​σ\+3\\sigmaband \(Appendix[B](https://arxiv.org/html/2607.06640#A2)\)\. Alongside the installed rank we always report the total recoverable structureR2​\(full​L←z\)R^\{2\}\(\\text\{full \}L\\\!\\leftarrow z\), the full convention panel across thresholds, and a leakage checkR2​\(L←h\)R^\{2\}\(L\\\!\\leftarrow h\)confirming the closure is not instead readable from the excluded recurrent state\.

A linear probe reports only the*linearly\-readable*rank, a lower bound on what the head may have installed nonlinearly; where the distinction matters we cross\-check with a one\-hidden\-layer probe\. Before measuring the dose\-response of Section[4](https://arxiv.org/html/2607.06640#S4)we verified, with a pre\-registered recovery gate, that the latent can represent the closure at all in the absence of the distractor \(recovery0\.870\.87–0\.890\.89\); only then did we measure how strongly the objective installs it under the distractor\.

## Appendix BAdditional controls and negative results

We treat the negative and confounded results as part of the evidence; this appendix collects the controls behind the main claims\.

#### The architecture confound and its fix\.

Our first auxiliary head read bothzzand the recurrent statehh\. Under the distractor this left the closure unrecoverable fromzz, but as a control it is confounded: the head routes the query throughhhand drainszzeven without the distractor\. We therefore force the head ontozzalone \(Appendix[A](https://arxiv.org/html/2607.06640#A1)\), which makes the leak structurally impossible \(R2​\(L←h\)≤0\.04R^\{2\}\(L\\\!\\leftarrow h\)\\leq 0\.04\)\. The honest reading is that the permissive head is a well\-verified but architecture\-confounded null and thezz\-only head is the clean test\.

#### Causal control\.

To confirm that the objective, not an incidental correlate, installs the closure, we shuffle the query labels across the training set with everything else fixed\. Recovery collapses to−0\.040\-0\.040\(the reconstruction null\) while the distractor stays readable \(R2​\(D←z\)=0\.395R^\{2\}\(D\\\!\\leftarrow z\)=0\.395\): the install is driven by the supervised target\.

#### Dimensionality vs\. pressure\.

The pressure\-matched control of Section[5\.2](https://arxiv.org/html/2607.06640#S5.SS2)raises a one\-dimensional objective’s weight to the four\-dimensional objective’s total target variance \(4\.5×4\.5\\times\)\. This matches the objective*weight*, not the achieved gradient magnitude: the realised gradient norm of the one\-dimensional control \(0\.0690\.069\) did not reach the four\-dimensional objective’s \(0\.3160\.316\)\. The pressure confound is therefore ruled out not by this control alone but directly, in situ: under the ramp the already\-installed direction*does*strengthen in magnitude—its recovery rises from0\.840\.84to0\.920\.92\(Figure[6](https://arxiv.org/html/2607.06640#S5.F6)a\)—but no second direction lights up, the unsupervised coordinates staying at the reconstruction null\. It is the*number*of installed directions, not their magnitude, that tracks dimensionality; extra pressure on one direction does not recruit another\. Pressure, in short, cannot pull rank\. \(The magnitude of the installed directions is roughly flat across thedd\-sweep as well, but that is a weaker observation and we do not lean on it\.\)

#### Threshold robustness\.

The per\-column thresholdτcol\\tau\_\{\\mathrm\{col\}\}is anchored post hoc to the reconstruction\-only null; the installed\-rank staircase is unchanged across the\+2\+2to\+3​σ\+3\\sigmaband, so the readout does not depend on the particular anchor within that range\. The threshold is not load\-bearing for a more basic reason \(Figure[12](https://arxiv.org/html/2607.06640#A2.F12)\): every supervised column is recovered atR2≈0\.8R^\{2\}\\approx 0\.8and every unsupervised column at≈−0\.05\\approx\-0\.05, soτcol\\tau\_\{\\mathrm\{col\}\}and its whole band fall inside a∼\\sim0\.80\.8\-wide empty gap between the two—any anchor in that gap yields the same staircase\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x12.png)Figure 12:The per\-column threshold is not load\-bearing\.Every supervised column of the objective\-dimensionality sweep is recovered atR2≈0\.8R^\{2\}\\approx 0\.8and every unsupervised column at≈−0\.05\\approx\-0\.05; the re\-anchored thresholdτcol\\tau\_\{\\mathrm\{col\}\}and its full\[\+2​σ,\+3​σ\]\[\+2\\sigma,\+3\\sigma\]band sit inside the∼\\sim0\.80\.8\-wide gap between them, so the installed\-rank staircase of Figure[6](https://arxiv.org/html/2607.06640#S5.F6)does not depend on where in that gap the threshold is drawn\.
#### Leakage in the matched pair\.

In the single matched\-dimension comparison of Section[5\.4](https://arxiv.org/html/2607.06640#S5.SS4)the closure is marginally readable from the excluded recurrent state \(R2​\(L←h\)≈0\.04R^\{2\}\(L\\\!\\leftarrow h\)\\approx 0\.04, just over our0\.030\.03purity bar\)—a ridge\-decode artifact where the target dimension approaches the probe’s effective degrees of freedom; the sweep cells of Section[5\.2](https://arxiv.org/html/2607.06640#S5.SS2)satisfy the bar\.

#### The value\-head sweep\.

Repeating the dimensionality sweep through the model’s own value head over add\-dimensional value/reward family \(d=1d=1–44\) separates objective dimensionality from head form\. The seed\-mean staircase \(three seeds atd=2d=2, two at each otherdd\) installs rank1,2,3,41,2,3,4with total recovery0\.11,0\.27,0\.47,0\.550\.11,0\.27,0\.47,0\.55, about0\.7×0\.7\\timesthe regression head’s magnitudes at matcheddd\. The per\-seed count atd=2d=2is threshold\-fragile: one seed’s leading unsupervised coordinate \(0\.0260\.026\) grazes the null threshold \(≈0\.025\\approx 0\.025\) and changes sign across seeds \(\+0\.026,−0\.018,−0\.043\+0\.026,\-0\.018,\-0\.043; mean−0\.012\-0\.012\), so the law is stated on the three\-seed mean and the total\-recovery overlay, not a per\-seed count\. The aggregate training pressure is not held constant across the sweep—the value\-head loss rises roughly ninefold fromd=1d=1tod=4d=4—so dimensionality is separated from pressure by the in\-situ proxies and the pressure\-matched control of Section[5\.2](https://arxiv.org/html/2607.06640#S5.SS2), as in the regression sweep, rather than by a matched design here\.

#### The knee recalibration\.

In the planted\-rank calibration \(Appendix[E](https://arxiv.org/html/2607.06640#A5)\) the exact “saturates atkk” claim is convention\-relative: under a stricter knee rule the trained capacity under\-reads\. We retain the diffeomorphism\-invariant statement \(minimal capacity tracks the closure rank monotonically\) and treat the convention\-dependence of the exact knee as a disclosed property of the estimator, not a hidden failure \(Figure[13](https://arxiv.org/html/2607.06640#A2.F13)\)\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x13.png)Figure 13:The knee is convention\-relative; the ordering is not\.Only the intrinsic rank iskkfor every observation map: the linear population spectrum over\-counts the high\-frequency swirl lift \(a: no sharp knee;90%→1390\\%\\to 13,95%→2395\\%\\to 23against plantedk=6k=6\), and the learned knee under\-counts the saturating and entangled lifts \(b\); the robust invariant is the monotone trend \(pooled Spearman0\.9670\.967\)\. Panel \(a\): Monte\-Carlo predictable\-variance spectra, recomputed atτ~=0\.10\\tilde\{\\tau\}=0\.10; panel \(b\): learneddsatd\_\{\\mathrm\{sat\}\}under the pre\-registered knee convention\.
#### Reconstruction sufficiency \(the scope boundary\)\.

The three arms behind Figure[10](https://arxiv.org/html/2607.06640#S9.F10)a—single\-reward value equivalence, a full value family, and action\-masked reconstruction with no reward—are trained on the frame\-observable control task quantified in Section[9](https://arxiv.org/html/2607.06640#S9)\(single\-frame decodeR2≈0\.998R^\{2\}\\approx 0\.998per coordinate\)\. All install the same rank\-44closure in posterior and open\-loop rollout alike, and the value\-equivalence arm’s return matches the family’s under common random numbers on every seed, so none of the main\-text dissociations appear where reconstruction already recovers the closure\.

#### The recurrence carry and its decay\.

In a partially observed variant in which a cue is visible only transiently, arms trained with reconstruction alone still carry the cue to its point of use: the recurrent state holds it with no objective pressure\. Read at increasing lags, this free carry decays from near\-perfect to chance by lag4545on all three seeds, while an arm that receives the cue through a supervised objective stays at ceiling across the same lags on all three seeds \(Figure[10](https://arxiv.org/html/2607.06640#S9.F10)b\)\. The contrast is a decay curve, not a single\-lag binary—a pre\-registered single\-lag threshold read of the same data proved seed\-fragile and was retired as a scorer, which we report rather than quote—and it separates supervised from passive channels, not a distinct “value channel” from supervision\. It is the evidence behind the ordering suggested in Section[9](https://arxiv.org/html/2607.06640#S9)\.

## Appendix CReduced\-rank ceiling and the allocation law

This appendix collects the borrowed rigor behind the bound of Section[6](https://arxiv.org/html/2607.06640#S6)\. None of it is a new theorem: the linear\-algebraic backbone is classical, cited in each statement, and we use it only to account for the*shape*of an empirical law in a trained deep model\. We first recall the closure\-spectrum identities the bound rests on, then state the ceiling and its proof sketch, then the rate–distortion allocation law \(separating what is derivable from what is borrowed intuition\), and finally the two scope fences, the pre\-committed falsifier, and the evaluation\-side reduction\.

### C\.1Closure spectrum: the linear backbone

Take a linear–Gaussian systemxt\+1=A​xt\+wtx\_\{t\+1\}=Ax\_\{t\}\+w\_\{t\}with a linear queryy=C​xy=Cx, and stack the horizon\-HHfuture queryft=\[yt\+1;…;yt\+H\]=OH​xt\+Ntf\_\{t\}=\[y\_\{t\+1\};\\dots;y\_\{t\+H\}\]=O\_\{H\}x\_\{t\}\+N\_\{t\}, whereOHO\_\{H\}is the observability map andNt⟂xtN\_\{t\}\\perp x\_\{t\}the irreducible future noise\. With reachability GramianWc=Cov​\(xt\)W\_\{c\}=\\mathrm\{Cov\}\(x\_\{t\}\)andM:=Wc1/2​OH⊤M:=W\_\{c\}^\{1/2\}O\_\{H\}^\{\\top\}, three facts organize everything downstream\.

#### Lemma 1 \(spectrum identity\)\.

The population canonical correlationsρi∈\[0,1\]\\rho\_\{i\}\\in\[0,1\]between past and future query are the singular values ofM​Σf​f−1/2M\\,\\Sigma\_\{ff\}^\{\-1/2\}, and the deterministic Hankel singular valuesσi\\sigma\_\{i\}are the singular values ofMMitself\. The two share rank and range exactly—\#​\{ρi\>0\}=\#​\{σi\>0\}=dim\(closure\)\\\#\\\{\\rho\_\{i\}\>0\\\}=\\\#\\\{\\sigma\_\{i\}\>0\\\}=\\dim\(\\text\{closure\}\), with the same closure subspace—and share ordering under isotropic predictable\-coordinate noise; the magnitudes coincide only up to the noise normalization\. That past/future canonical correlations are the stochastic\-realization Hankel singular values is classical canonical variate analysis\[Anderson,[1951](https://arxiv.org/html/2607.06640#bib.bib3)\]; we use only its rank/range content\.

#### Theorem 1 \(rank\-dderror==floor\+\+tail\)\.

In the whitened query metric, the best rank\-ddpredictor’s relative error decomposes exactly as

err​\(d\)2=1−1q​H​∑iρi2⏟floor2\+1q​H​∑i\>dρi2,\\mathrm\{err\}\(d\)^\{2\}\\;=\\;\\underbrace\{1\-\\tfrac\{1\}\{qH\}\\textstyle\\sum\_\{i\}\\rho\_\{i\}^\{2\}\}\_\{\\mathrm\{floor\}^\{2\}\}\\;\+\\;\\tfrac\{1\}\{qH\}\\textstyle\\sum\_\{i\>d\}\\rho\_\{i\}^\{2\},each added dimension removing exactlyρd\+12/q​H\\rho\_\{d\+1\}^\{2\}/qHof error\. The tail bookkeeping is reduced\-rank regression / canonical\-correlation truncation\[Eckart and Young,[1936](https://arxiv.org/html/2607.06640#bib.bib9), Brillinger,[1969](https://arxiv.org/html/2607.06640#bib.bib7), Izenman,[1975](https://arxiv.org/html/2607.06640#bib.bib19)\]; the named, co\-equal floor term and the exact orthogonal split are what the present program treats as a first\-class axis\.

#### Corollary 1 \(minimal sufficiency\)\.

The number of canonical correlations above a toleranceτ\\tauis the minimal model dimension past which each further direction buys less thanτ2/q​H\\tau^\{2\}/qHof query variance—a property of the dynamics, query, horizon, exploration, and feature class, not of any model\. This is the Hankel\-rank result shared by balanced truncation\[Moore,[1981](https://arxiv.org/html/2607.06640#bib.bib23)\]and spectral learning of predictive states\[Hsu et al\.,[2012](https://arxiv.org/html/2607.06640#bib.bib18)\]; we add only its relativity to the full tuple\.

### C\.2The reduced\-rank ceiling

WriteL∈ℝkL\\in\\mathbb\{R\}^\{k\}for the closure the query depends on andz∈ℝnzz\\in\\mathbb\{R\}^\{n\_\{z\}\}for a learned latent\. Suppose an objective supervises the firstd≤kd\\leq kcoordinatesLd=U⊤​LL\_\{d\}=U^\{\\top\}L\(U∈ℝk×dU\\in\\mathbb\{R\}^\{k\\times d\}\) of the closure throughzz, and writeinstalledL\\mathrm\{installed\}\_\{L\}for the rank a linear probe can read out ofzz\.

#### Statement \(linear head\)\.

For a linear head forming the population\-optimal prediction ofLdL\_\{d\}fromzz, the relevant cross\-covarianceΣLd,z=U⊤​ΣL,z\\Sigma\_\{L\_\{d\},z\}=U^\{\\top\}\\Sigma\_\{L,z\}has rank at mostrank​\(U⊤\)=d\\mathrm\{rank\}\(U^\{\\top\}\)=d, so the optimal predictor occupies at mostdddirections ofzz\. With the empirical premise that reconstruction installs essentially none ofLLin the regime where the distractor is the reconstruction\-salient direction \(verified: the reconstruction\-only readout ofLLcollapses to the null\), this gives

installedL≤d,independently ofnz\.\\boxed\{\\;\\mathrm\{installed\}\_\{L\}\\;\\leq\\;d,\\quad\\text\{independently of \}n\_\{z\}\.\\;\}The rank\-ddceiling is Eckart–Young\[Eckart and Young,[1936](https://arxiv.org/html/2607.06640#bib.bib9)\]; the load\-bearing content—*which*dddirections install, namely the leading canonical directions of the\(L,z\)\(L,z\)cross\-covariance—is the canonical\-correlation truncation of Lemma 1 / Theorem 1\[Anderson,[1951](https://arxiv.org/html/2607.06640#bib.bib3), Izenman,[1975](https://arxiv.org/html/2607.06640#bib.bib19)\]\. The “≤\\leq” is the theorem; equality \(that the installed rank attainsddrather than falling short\) is*empirical*\. The result is the objective\-side counterpart of value equivalence\[Grimm et al\.,[2020](https://arxiv.org/html/2607.06640#bib.bib14)\]: the objective’s rank caps the installed closure\.

#### Proof sketch\.

Supervising onlyLd=U⊤​LL\_\{d\}=U^\{\\top\}Lmakes the target–latent cross\-covariance factor throughU⊤U^\{\\top\}, of rank≤d\\leq d; the best rank\-constrained linear map fromzzto that target is its reduced\-rank / canonical\-correlation truncation, whose row space has dimension≤d\\leq d\[Eckart and Young,[1936](https://arxiv.org/html/2607.06640#bib.bib9), Anderson,[1951](https://arxiv.org/html/2607.06640#bib.bib3)\]\. Capacitynzn\_\{z\}enters only as the ambient dimension ofzzand cannot raise the cross\-covariance rank, so the bound is capacity\-independent\. The reconstruction term, by the empirical premise, contributes no additionalLL\-direction\.

#### From an optimal\-readout bound to a gradient\-flow prediction\.

The statement above bounds an optimal linear*readout*; the empirical claim is about what training*installs*\. In the linear case the two coincide, and this upgrades “=d=dis empirical” to a training\-dynamics prediction\. For a linear encoder and linear head trained by gradient flow from small initialization, learning proceeds mode by mode in order of the canonical correlations of the target–input cross\-covariance, converging to exactly its top\-ddcanonical directions—the deep\-linear learning dynamics ofBaldi and Hornik \[[1989](https://arxiv.org/html/2607.06640#bib.bib4)\], Saxe et al\. \[[2014](https://arxiv.org/html/2607.06640#bib.bib24)\], whose fixed point is the reduced\-rank / canonical\-correlation truncation already invoked\. So for a linear stackinstalledL=d\\mathrm\{installed\}\_\{L\}=dis not only the ceiling but the gradient\-flow limit, and our nonlinear\-stack measurement \(installed rank=d=dat everydd\) is the verification that the prediction carries over\. Equality \(rather thaninstalledL<d\\mathrm\{installed\}\_\{L\}<d\) requires two conditions, which sharpen the falsifier: the supervised target’s covariance must have full rankdd, and allddof its canonical correlations with the recoverable latent must be strictly positive; a rank\-deficient target or a vanishing canonical correlation installs strictly fewer thandddirections\.

### C\.3The allocation law: derivable ordering vs\. borrowed intuition

The latent has finite capacity, so a model trained on reconstruction\-plus\-query solves a constrained allocation\. We separate two registers\.

#### Derivable\.

At a constrained optimum the marginal unit of capacity goes to whichever direction yields the largest marginal loss reduction—a greedy marginal\-value ordering\. This much is the standard convex\-allocation argument and is all that is genuinely derivable here\.

#### Borrowed intuition \(not derived\)\.

The reverse\-water\-filling solution of rate–distortion theory\[Berger,[1971](https://arxiv.org/html/2607.06640#bib.bib6)\]supplies the picture for that ordering, but it is a single\-source, single\-distortion, single\-budget result and does*not*instantiate the present two\-objective problem \(reconstruction distortion versus query distortion sharing one budget\); reducing the two to one water\-filling problem requires a common marginal\-value currency on a jointly\-Gaussian basis, which is a modeling assumption, not the theorem\. The regime statements are therefore*predictions*the heuristic makes, labeled as such: under tight capacity the query*displaces*reconstruction \(a one\-sided shedding, with the two readouts not summing to a conserved budget\); under abundant capacity the query*adds*, withinstalledL=d\\mathrm\{installed\}\_\{L\}=dand reconstruction retained\. On a continuous latent, capacity binds as a*rank*\(nzn\_\{z\}independent directions\); on a categorical latent it binds as a*rate*\(aboutnz​log2⁡Cn\_\{z\}\\log\_\{2\}Cbits\), so a single categorical direction can carry partial information about more thannzn\_\{z\}closure directions at once and the rank form takes the wrong functional shape—which is why the capacity claims are made on a continuous \(or synthetic\-continuous\) latent and why a categorical latent saturates by finer quantization rather than by adding directions\. The threshold at which displacement turns into addition is itself tolerance\-relative and ordinal, never a sharp integer; we do not derive it\.

### C\.4The marginal\-value ordering, tested one level below the ceiling

The rank ceiling fixes*how many*directions install and the canonical\-correlation truncation names*which*ones; the allocation heuristic adds a stronger, ordinal claim—that under a flat objective the directions install*in proportion to*their marginal value, i\.e\. their dynamical predictability\. We tested this refinement on a trained pixel stack supervising four modal targets of distinct predictability \(ρ=0\.98,0\.90,0\.80,0\.65\\rho=0\.98,0\.90,0\.80,0\.65\), reading both each direction’s installed magnitude and its install timing\. Unlike the scoped\-out generalization negative of Appendix[C\.5](https://arxiv.org/html/2607.06640#A3.SS5), the install here*is*present on fresh episodes—that is the premise of the test—so what is at issue is not whether the directions install but how strongly\.

The magnitude half of the prediction does not hold\. Across three seeds the converged install magnitude runs*counter*to the predictability order \(per\-seed rank correlations−0\.67,−0\.67,−0\.33\-0\.67,\-0\.67,\-0\.33; mean−0\.56\-0\.56\), the stable anomaly being that the*second*\-most\-predictable direction \(ρ=0\.90\\rho=0\.90\) installs weakest in every seed\. Install*timing*, by contrast, partially recovers the predictability order \(first\-crossing rank correlation\+0\.33\+0\.33, all three seeds\), so it is the converged*magnitude*, not the arrival order, that departs\. The obvious confounds are ruled out in this stack: a same\-architecture, dynamics\-free convolutional control recovers all four directions atR2≈0\.999R^\{2\}\\approx 0\.999\(so the effect is not encoder readability\), the per\-direction target variances are isotropic to within3\.5%3\.5\\%\(not a scale artifact\), and the loss has plateaued \(not under\-training\)\. We do not identify the mechanism here; a separate line probes it interventionally\.

This recalibrates the borrowed*intuition*, not the theorem, and along an axis the paper has already found fragile\. Just as the planted\-rank saturation knee is estimator\-relative while its monotone ordering survives \(Appendix[B](https://arxiv.org/html/2607.06640#A2)\), here a magnitude—the relative install strength—is stack\-relative; the parallel is that a magnitude recalibrates in both, not that the ordering is equally kept—the install*timing*order here is only sign\-stable \(τ=\+0\.33\\tau=\+0\.33across three seeds\), far short of the knee’s near\-perfect ordinal survival\. The rank bound \(installedL≤d\\mathrm\{installed\}\_\{L\}\\leq d\) and its attainment \(=d=d\) are untouched: all four directions clear the reconstruction\-only per\-column null\. This magnitude result is moreover distinct from—and does not dent—the confirmed ordinal*shape*of the rank falsifier \(Appendix[C\.6](https://arxiv.org/html/2607.06640#A3.SS6), the sub\-ddramp rising into a plateau atdd\): that shape concerns how the*count*fills as capacity grows, not the relative strength of thedddirections at fixed capacity\. The invariant the paper rests on is the rank; the relative magnitude with which its directions install is not\.

### C\.5Two scope fences

#### Linear head vs\. nonlinear head\.

The ceiling is proved for a linear head; every experiment here uses a nonlinear head, whose input–output Jacobian is not rank\-≤d\\leq d\. The bound therefore transfers only to the*linearly\-readable*rank—what a linear probe extracts—and only under an empirical no\-spillover regularity: under ramped query pressure the unsupervised coordinates ofLLstay at the reconstruction null, a label shuffle collapses the readout, and reconstruction alone installs none ofLL\. We never present the nonlinear\-head experiments as instances the linear theorem proves\.

#### Synthetic vs\. pixel\.

The capacity\-independent half—installed rank rising then plateauing atddas the latent grows past the closure—is demonstrated on a synthetic continuous latent\. The trained pixel stack is a separate, scoped\-out generalization*negative*: there the install is real on the training distribution but fails to generalize to fresh rollouts even though the signal is present in the fresh observations, an encoder\-generalization failure with no positively identified mechanism\. It is orthogonal to the rank\-axis bound and is not explained or rescued by this appendix\.

### C\.6The pre\-committed falsifier

So that a flexible law cannot absorb any outcome, we pre\-commit, before the run and on a continuous low\-rate latent: the rank form is*confirmed*iff the installed rank plateaus atddthrough a latent at least twice the closure rank, with the unsupervised coordinates at the reconstruction null and non\-overlapping supervised/unsupervised readout bands; it is*refuted*by a continued rise with capacity or by a genuinely installed extra direction in shielded coordinates\. The invariant is the ordinal*shape*—a sub\-ddramp rising into a plateau atdd—rather than an exact integer count, since the count inside the ramp is tolerance\-relative\. The synthetic calibration shows this criterion is testable rather than vacuous: it resolves a plateau atd=3d=3from a plateau atd=4d=4by the objective dimension, not by capacity\.

### C\.7The evaluation\-side reduction

The same low\-dimensional reading appears on the evaluation side\. At an anchor\(st,at,st\+1\)\(s\_\{t\},a\_\{t\},s\_\{t\+1\}\)letz^′\\hat\{z\}^\{\\prime\}be the model’s one\-step latent rollout andz′=enc​\(st\+1\)z^\{\\prime\}=\\mathrm\{enc\}\(s\_\{t\+1\}\)the encoded true next state, with reward errorΔ​r=rmodel−rtrue\\Delta r=r\_\{\\text\{model\}\}\-r\_\{\\text\{true\}\}and value\-only operator errorΔ​v=V​\(z^′\)−V​\(z′\)\\Delta v=V\(\\hat\{z\}^\{\\prime\}\)\-V\(z^\{\\prime\}\)\. The discounted Bellman residual decomposes*by construction*:

\|\(rmodel\+γ​V​\(z^′\)\)−\(rtrue\+γ​V​\(z′\)\)\|=\|Δ​r\+γ​Δ​v\|,\|Bellman residual−γ​\|Δ​v\|\|≤\|Δ​r\|\.\\big\|\(r\_\{\\text\{model\}\}\+\\gamma V\(\\hat\{z\}^\{\\prime\}\)\)\-\(r\_\{\\text\{true\}\}\+\\gamma V\(z^\{\\prime\}\)\)\\big\|\\;=\\;\|\\Delta r\+\\gamma\\,\\Delta v\|,\\qquad\\big\|\\,\\text\{Bellman residual\}\-\\gamma\\,\|\\Delta v\|\\,\\big\|\\;\\leq\\;\|\\Delta r\|\.When the reward head is exact \(Δ​r=0\\Delta r=0\) the Bellman residual equalsγ\\gammatimes the value\-only operator error*exactly*; when the reward error is small relative to the value term, the two are near\-identical and agree in rank\. This is elementary—the affine one\-step backup telescoped into its reward and value parts—and we present it as a standard observation, not a theorem\. The identity is written for a single sampled transition; for stochastic dynamics the Bellman residual carries an expectation over next states \(r\+γ​𝔼s′​V​\(z′\)r\+\\gamma\\,\\mathbb\{E\}\_\{s^\{\\prime\}\}V\(z^\{\\prime\}\)\) and the same telescoping holds in expectation, while our per\-anchor operator error is the sampled form, applied anchor\-by\-anchor\. On released models the reward error is two orders of magnitude below the value term, so the \(unnormalized\) value\-only operator error and the Bellman residual are empirically near\-identical; the normalized\{r,V\}\\\{r,V\\\}slice, a different aggregate, is not\. Value equivalence is thus the low\-dimensional slice on the evaluation side as it is on the training side\.

## Appendix DEvaluation\-side protocol

The evaluation\-side result of Section[8](https://arxiv.org/html/2607.06640#S8)is reported in full in a companion paper\[Vakalis,[2026](https://arxiv.org/html/2607.06640#bib.bib28)\]; we summarize the protocol here\. For a model with encoderenc\\mathrm\{enc\}andkk\-step latent rolloutOkO\_\{k\}, and an observable subsetFF, the operator error at an anchor\(st,at:t\+k−1,st\+k\)\(s\_\{t\},a\_\{t:t\+k\-1\},s\_\{t\+k\}\)compares the model’s rolled\-forward latentz^t\+k=Ok​\(zt,at:t\+k−1\)\\hat\{z\}\_\{t\+k\}=O\_\{k\}\(z\_\{t\},a\_\{t:t\+k\-1\}\)to the encoded true next statezt\+k′=enc​\(st\+k\)z^\{\\prime\}\_\{t\+k\}=\\mathrm\{enc\}\(s\_\{t\+k\}\), by reading each functionalϕ∈F\\phi\\in Foff both through a shared ridge probe and aggregating the normalized per\-anchor errors\. The*value slice*setsF=\{r,V\}F=\\\{r,V\\\}using the model’s own reward and value heads; the*full*aggregate adds a per\-anchor PCA basis on the encoded next\-state geometry, fit on a held\-out half of the anchors so it cannot adapt to operator error\.

We evaluate released TD\-MPC2 checkpoints\[Hansen et al\.,[2024](https://arxiv.org/html/2607.06640#bib.bib17)\]on the DeepMind Controlcheetah\-runtask\[Tassa et al\.,[2018](https://arxiv.org/html/2607.06640#bib.bib27)\]at five parameter counts \(11M–317317M\), at a matched five\-step horizon\. Across the sweep, reward\-prediction error stays in\[0\.028,0\.091\]\[0\.028,0\.091\]while the full\-operator error spans0\.280\.28to2\.622\.62; the full\-operator error tracks executed return at Spearman−0\.90\-0\.90\(anchor\-bootstrap CI\[−0\.90,−0\.70\]\[\-0\.90,\-0\.70\], leave\-one\-out≤−0\.80\\leq\-0\.80;n=5n=5sizes, correlational\), whereas the Bellman residual and reward error track return only weakly \(−0\.10\-0\.10and−0\.30\-0\.30\)\. The unnormalized value\-only operator error is rank\-identical to the Bellman residual across the sweep \(Spearman\+1\.00\+1\.00\), consistent with the decomposition of Appendix[C\.7](https://arxiv.org/html/2607.06640#A3.SS7); the normalized value slice is not \(\+0\.10\+0\.10\)\. The full cross\-architecture comparison and additional controls are deferred to the companion paper\[Vakalis,[2026](https://arxiv.org/html/2607.06640#bib.bib28)\]\.

## Appendix EEnvironment construction and synthetic protocols

#### The controlled testbed\.

The main experiments use a controlled latent\-recovery environment: a known low\-dimensional process ofkkslowly varying latent coordinates is rendered through a fixed analytic \(cubic\) warp into a64×6464\\times 64image observation, with a high\-variance distractor added alongside the low\-variance query coordinates\. The query coordinates modulatekkGaussian blobs at fixed, well\-separated positions and the distractor modulates full\-frame low\-frequency fields, spatially entangled so that every pixel carries both contributions and a convolutional encoder cannot ignore the distractor by receptive\-field masking\. The closure rank is therefore known by construction, and the query coordinates are linearly decodable from the observation atR2≈0\.85R^\{2\}\\approx 0\.85, so any failure to represent them reflects how capacity is allocated rather than a missing signal\. The world model is the DreamerV3 categorical\-RSSM stack\[Hafner et al\.,[2023](https://arxiv.org/html/2607.06640#bib.bib16)\], trained to convergence\. The dose–response of Section[4](https://arxiv.org/html/2607.06640#S4)usesk=2k=2closure coordinates; the corner comparison and both dimensionality sweeps of Section[5](https://arxiv.org/html/2607.06640#S5)usek=4k=4with latent capacitynz=4n\_\{z\}=4\.

#### Capacity \(continuous\-latent\) calibration\.

To separate capacity from objective dimensionality we replace the categorical latent with a continuous one and sweep its width past the closure rank\. The installed rank rises with capacity and plateaus at the objective dimensiondd\(the plateau*level*tracksdd; the exact knee is tolerance\-relative\), holding through a latent several times the closure rank; a categorical latent of the kind the main stack uses instead saturates at a fixed level regardless of width, consistent with its capacity binding as a rate rather than a rank \(Appendix[C\.3](https://arxiv.org/html/2607.06640#A3.SS3)\)\. This run was pre\-registered as a test of the rank\-form prediction, with the falsifier of Appendix[C\.6](https://arxiv.org/html/2607.06640#A3.SS6)fixed before it was executed\.

#### Planted\-rank calibration\.

As a ground\-truth check on reading rank off a trained model, we plant environments whose closure rankkkis known and ask where a trained model’s minimal sufficient latentdsatd\_\{\\mathrm\{sat\}\}saturates\. On the simplest \(linear\) familydsat=kd\_\{\\mathrm\{sat\}\}=kexactly \(Spearman1\.0001\.000\); as the observation map is varied across analytic lifts \(cube, tanh, and mixtures\) the minimal capacity trackskkmonotonically \(pooled Spearman0\.9670\.967\), with the exact knee level convention\-relative \(Appendix[B](https://arxiv.org/html/2607.06640#A2)\)\. All synthetic and planted results are labeled as calibration and are never read as natural\-image rank results\.

#### Covariance vs\. closure\.

The teaching panel of Section[2](https://arxiv.org/html/2607.06640#S2)uses a single oscillatory mode, which occupies one direction of the observation covariance but spans a two\-dimensional predictive subspace; shuffling or replacing the dynamics with noise collapses the closure on both axes\.

## Appendix FCapacity reallocation

At tight capacity the allocation law of Appendix[C\.3](https://arxiv.org/html/2607.06640#A3.SS3)predicts that installing the query*displaces*the distractor rather than adding to it\. At a latent width equal to the closure rank, as the query weightλ\\lambdarises the recovered distractor falls monotonically \(SpearmanρD​\(λ\)=−1\.0\\rho\_\{D\}\(\\lambda\)=\-1\.0\) while the recovered query rises \(ρL​\(λ\)=\+0\.94\\rho\_\{L\}\(\\lambda\)=\+0\.94\); the two do not sum to a conserved budget, consistent with one\-sided shedding rather than a fixed\-budget rank trade\. The displacement is reproduced across seeds at the clean interior weight, where the install is partial and the latent carries both \(R2​\(L←z\)≈0\.42R^\{2\}\(L\\\!\\leftarrow z\)\\approx 0\.42whileR2​\(D←z\)≈0\.29R^\{2\}\(D\\\!\\leftarrow z\)\\approx 0\.29\); the full monotone shape at largerλ\\lambdais single\-seed and entangled with degrading reconstruction\. We therefore report displacement as a confirmed tight\-capacity prediction and the precise high\-λ\\lambdatrajectory as suggestive\.

## Appendix GBenchmark probe

As a check on whether the closure estimator says anything on a real combinatorial benchmark, we apply it to the publicly runnable subset of AutumnBench\[Warrier et al\.,[2025](https://arxiv.org/html/2607.06640#bib.bib30)\]\. From a fixed uniform\-random exploration protocol we estimate, per \(environment, query\-family\) cell, the closure’s effective rank and linear error floor, and ask whether these separate the environments the released models scale on from those they saturate on\. They do not: across the sixteen publicly runnable environments the per\-family separations are consistent with chance \(stratified within\-family permutationp=0\.53/0\.75/0\.98p=0\.53/0\.75/0\.98on the floor, effective\-rank, and action\-corrected excess axes; AUC≈0\.5\\approx 0\.5\), while shuffle and i\.i\.d\. controls collapse as expected \(Figure[14](https://arxiv.org/html/2607.06640#A7.F14)\)\.

![Refer to caption](https://arxiv.org/html/2607.06640v1/x14.png)Figure 14:The benchmark probe is a bounded null within the instrument’s reach\.\(a\)The action\-corrected excess floor does not separate the environments the released models scale on from those they saturate on \(AUC=0\.56=0\.56, stratified permutationp=0\.98p=0\.98;n=16n=16public environments—underpowered, so a moderate effect would not be distinguishable from noise\)\.\(b, c\)The high floors decompose into an exogenous action\-entropy bound plus a closure part: of the4848\(environment, family\) cells,1414keep a large excess floor \(nonlinear\-closure candidates the linear estimator cannot test\),2626are indeterminate at the pre\-registered budget,77are linearly learnable, and11is action\-noise\-dominated\. The linear estimator reaches only1717of4848closures, which is the reach bound that motivates a deeper estimator; this figure bounds the instrument, not the thesis\.Two bounds make this a statement about the instrument rather than a test the thesis passes or fails\. First, it is underpowered: only sixteen of the forty\-three scored environments are publicly runnable, so a moderate effect could not be distinguished from noise\. Second, the linear estimator reaches only about a third of the relevant closures \(roughly seventeen of forty\-eight cells have a recoverable linear floor\); the rest would need a nonlinear estimator to be tested at all\. We therefore report AutumnBench as the boundary at which our linear instrument stops being informative on a discrete combinatorial benchmark, and as motivation for a deeper estimator—not as evidence for or against the dimensionality law\.