Impossibility

Language model-generated rubrics are increasingly used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge evaluation, and automated grading. Such rubrics are reliable only if they reward honest answers over adversarial answers optimized to exploit them. Yet their robustness to such optimization remains poorly understood. We isolate the hardest regime: impossible tasks, where the prompt pressures the model toward an unsupported conclusion, so the only honest response is to acknowledge the impossibility. We introduce ImpossibleRubrics, a benchmark of 169 impossible tasks spanning six impossibility categories, each paired with a verifiable oracle certificate specifying what an honest answer may and may not claim, together with 48 answerable controls. Rather than providing fixed rubrics, ImpossibleRubrics provides task environments and certificates, allowing rubrics to be generated downstream and then adversarially tested for whether they reward certificate-violating answers. Eleven generators are exploited 8--26% of the time on the unbiased 150-of-169 environment cut; on a deliberately selected stress cut the strongest generator we measured is still exploited 36% while a certificate-faithful rubric is exploited 0%, so what we measure is a rubric-quality gap, not task impossibility. One result runs against intuition. A single generic rubric ("be decisive, penalize hedging") used unchanged for every task is exploited 64% of the time, and seven of the eleven generators are exploited more often than that while writing a rubric tailored to each one. The tailored criteria appear to tell an attacker which claim to fabricate. The problem is not that rubrics are vague; it is that they are specific about the wrong things.

Despite its ubiquity, clustering lacks a universally accepted definition of what is a cluster. Kleinberg's Impossibility Theorem formalizes this difficulty by showing that no flat clustering method can simultaneously satisfy three natural axioms: scale invariance, richness, and consistency. In this paper, we ask whether this impossibility persists when the output is a hierarchy rather than a single partition. We show that, in contrast to the flat clustering setting, the hierarchical analog of these axioms are jointly satisfiable. In fact, there exist uncountably many hierarchical clustering methods satisfying these axioms, which we call admissible. We explicitly construct several admissible methods, including methods based on well-separated clusters and a non-binary version of single linkage. For certain pairs of admissible methods, the hierarchy produced by one always refines that produced by the other. This refinement relation defines a partial order on the class of admissible methods. This partially ordered set has no greatest element and contains uncountably many pairwise incompatible maximal elements, revealing substantial diversity among admissible methods. Nevertheless, this diversity is constrained: every admissible method contains a hierarchy of sufficiently well-separated clusters, and every finite collection of admissible methods shares such a nontrivial common backbone.

A false data injection attack (FDIA) can change the estimated grid state while evading a residual-based bad data detector (BDD). Existing blind attacks learn a low-rank measurement subspace, but this algebraic view does not state the physical grid constraints that make an attack stealthy or the minimum information needed to recover the complete attack space. Under the connected direct-current (DC) branch-flow model, we show that the residual-sensitive subspace of the noiseless orthogonal test is exactly the weighted cycle space. Its orthogonal complement is therefore the complete stealthy attack space, making weighted cycle-space knowledge both necessary and sufficient for complete blind FDIA. This space identifies the topology only up to 2-isomorphism and the relative cycle-edge parameters only up to one scale per biconnected component; bridge parameters are neither identified nor required. We then formulate a computationally unconstrained benchmark and a tractable measurement-only reconstruction method. Experiments on IEEE systems compare BDD bypass rate at a 95% nominal-acceptance threshold against state impact. As a compact alternating-current (AC) extension, we characterize feasible branch P/Q measurements by a cycle manifold and demonstrate topology-assisted manifold fitting and measurement generation on a graphics processing unit (GPU). In the lossless fixed-voltage small-angle limit, the normal space of the active-power slice reduces to the DC weighted cycle space.

Several properties safety monitors are asked to certify, among them cross-tenant noninterference, sandbagging and evaluation awareness, are 2-safety hyperproperties, witnessed only by two executions. The standard consequence is a binary impossibility: one trace cannot decide them. We replace the binary with a measurement. A tight bound puts the balanced accuracy of any single-trace monitor at $\tfrac12+\tfrac12\,TV(P_0,P_1)$, turning undecidability into a graded detectability frontier and defining an oversight gap: a monitor's shortfall below it. On a leak family with closed-form $TV$, nine LLM monitors are optimal at $TV=0$ but capture little signal as $TV$ grows; at $TV=1$, where a 20-line membership check scores $100\%$, they average $60.9\%$. That shortfall is mostly not capability: naming what to check closes $61\%$ of it while leaving the $TV=0$ control at chance. The same split runs through a $2{\times}2$ factorial: an imagined second run leaves monitors at chance ($50.4\%$) while the same rule on an executed second run reaches $90.0\%$, and a stored oracle without a comparison procedure yields only $68.2\%$. Information and procedure are each necessary and neither is capability. Under nondeterminism, replay tracks a closed-form $k$-replay curve only under the right projection, and a projection frontier shows the resulting dilemma is forced: narrow misses $98.6\%$ of off-channel leaks, broad flags $75.7\%$ of clean traffic, and attainable accuracy decays like $1/(qm)$ in the benign-variation rate and the channel count. Finally, two frontier LLM judges certified an earlier version of our own benchmark as sound while a sign test found a directional bias ($p=2.7\times10^{-5}$) that invalidated three of our findings. Construction validity for hyperproperty benchmarks should be proved mechanically, not audited by models.

Language can describe states of affairs that are false and states of affairs that could not be the case at all. Whether an AI model internally distinguishes these failures remains unclear. I report an exploratory activation study of the multimodal open-weight model Gemma 3 4B IT using 85 prompts from 17 philosophical families and a topic-matched modality set of 15 topics, each expressed as a truth, contingent falsehood, improbable claim, semantic anomaly, and necessary falsehood. In its answers, the model conflates contingent falsehood with contradiction, labeling 12 of 15 false statements "contradiction." Its activations show a different pattern. A linear truth probe separates impossible from true statements (AUC 0.93) but not impossible from false statements (AUC 0.20). An impossibility probe evaluated on held-out topic families separates necessary from contingent falsehood at AUC 1.00, peaking at layer 15 with balanced accuracy 0.97 (Bonferroni-adjusted P=0.018). The truth and impossibility directions are close to orthogonal, whereas the impossibility direction partially overlaps a semantic anomaly direction while remaining distinguishable from it. Sparse autoencoder features at the same layer repeat this geometry. Features selective for impossibility also fire on anomalous sentences but rarely on contingent falsehoods. In this model's activation space, necessary falsehoods are not extreme cases of contingent falsehood but lie closer to the experimentally defined category of semantic anomaly. This representational proximity does not imply that impossible statements are intrinsically meaningless. These correlational observations from one small model offer an empirical footnote to an old philosophical distinction.

This paper investigates whether the postmodern claim of unrestricted semantic indeterminacy, and its foundational Saussurean axiom of the arbitrary sign, are compatible with the structural architecture of Classical Arabic. We develop a formal mathematical model of Arabic non concatenative morphology in which lexical meaning is determined by the interaction between an invariant root and a morphosyntactic pattern. Within this framework, we establish a Morphological Correspondence Theorem, demonstrating that every lexical item is uniquely generated by a root pattern pair, and a Semantic Localization Theorem, proving that lexical meaning is determined at the derivational level prior to surface realization. To address Saussurean weaker notion of relative arbitrariness, we formalize it via conditional Kolmogorov complexity, defining arbitrariness algorithmically as the no rule property. We prove that general relative arbitrariness is formally undecidable, while Arabic relative arbitrariness is decidable and provably less than 1 for its motivated signifiers (Levels W and M), establishing a strict system complexity asymmetry over Indo-European languages.

Group Relative Policy Optimization (GRPO) is the dominant reinforcement learning algorithm for training reasoning capabilities in large language models, notably adopted by DeepSeek-R1. The recent improvement Dr. GRPO (COLM 2025) identifies the response-level length bias caused by per-trajectory length normalization in GRPO and proposes removing this normalization, claiming the resulting optimizer is "unbiased." We show that this claim is incomplete. Specifically, we establish an impossibility theorem: under the standard outcome reward + GRPO setting, no length-based weighting scheme can simultaneously achieve the following two properties. (P1) Gradient unbiasedness: the gradient estimator is an unbiased estimate of the true policy gradient. (P2) Length invariance: each trajectory's effective contribution to the gradient is independent of its token length. GRPO approximately satisfies P2 but violates P1; Dr. GRPO satisfies P1 but violates P2. We characterize the complete tradeoff spectrum via the parametric family f_alpha(L) = L^{alpha - 1}, where alpha = 0 recovers GRPO, alpha = 1 recovers Dr. GRPO, and provide quantitative analysis showing that Dr. GRPO's length bias can cause longer trajectories to dominate gradient updates by a factor proportional to the length ratio. Our results reveal that neither algorithm is universally "done right"; they occupy opposite ends of a fundamental and unavoidable tradeoff.

Information locality, the tendency for syntactically related words to appear close together, shapes both human language processing and language model learning. While prior work has examined whether language models can acquire impossible languages, it remains unclear whether they can recover natural language from such input and what this reveals about their inductive biases. We address this by complementing learnability-based approaches with a reconstruction framework: fine-tuning GPT-2 models pre-trained on impossible languages to reconstruct natural English from three perturbation types. Our findings show that the recovered structures exhibit shorter dependency lengths than the original text, mirroring the locality preference observed in unconstrained language model generation and providing a quantitative signature of an architectural bias that learnability experiments alone do not reveal. Recovery difficulty increases with the degree of locality disruption. Structural recovery (dependency Triple F1) dissociates from surface recovery (Exact Match), while fluency dissociates from faithful reconstruction under global shuffling. Sentence length further modulates performance: longer sentences facilitate recovery when local structure is preserved but lead to complete collapse under global shuffling. Finally, recovery difficulty tracks learnability difficulty across perturbation types, suggesting that information locality is the shared constraint governing both.

Recent work suggests that transformer language models show a bias towards human languages over unnatural ("impossible") languages argued to be unacquirable by humans. However, this literature has largely based these claims on differences in sample efficiency and test-set perplexity, rather than on direct evaluations of the linguistic capacities that could plausibly explain non-attestation in human languages. We evaluate two theoretically motivated linking hypotheses: impossibility arising from deficiencies in grammatical sensitivity or generative production. Using GPT-2 style models trained on perturbed "impossible" variants of English, we measure sensitivity to grammaticality using BLiMP minimal pairs, finding that model performance exhibits only gradual degradation, mediated by the language's information locality. In contrast, these models exhibited pronounced failures in generation, producing substantially fewer high-quality sentences at longer lengths. Together, these results suggest generative deficiency and transmission failures as a plausible linking hypothesis between language model behaviour and non-attestation of impossible languages.

Large language models (LLMs) deployed for structured generation (NER, JSON extraction, QA, and classification) lack formal reliability guarantees, and standard heuristic abstention policies miss user-specified risk targets by 7.5--12.5%. We characterize when conformal risk control (CRC) can certify structured LLM outputs and when it provably cannot. First, we prove an impossibility result: when the base risk (μ> α), any distribution-free method must abstain on at least ((μ-α)/(1-α)) examples, yielding a closed-form feasibility test: one can check whether CRC will work before running it. Second, we analyze a certification hierarchy across Hoeffding, empirical Bernstein, and a betting-based e-CRC bound, with strict gains in low-variance/large-sample regimes: the Hoeffding-to-Bernstein step delivers the largest gain (+37% certified configurations), while e-CRC adds value when calibration data is scarce (10% certification at 20% data versus 0% for Hoeffding). Third, we validate adaptive conformal inference (ACI) under cross-dataset shift, reducing risk-target violations from 71% to 21%, with residual failures concentrated exactly where the impossibility bound predicts. Across six open-weight models (3B--72B parameters), eight datasets, four tasks, and six nonconformity scores, hard NER/QA/CLS configurations are uncertifiable at (α= 0.10); relaxing to (α= 0.30--0.40) unlocks practical certification (47% NER, 40% QA, 60% CLS). The framework gives a three-step deployment recipe: check feasibility, select the bound and score, then mitigate shift.

Advanced AI systems have extensive knowledge of their environments; in fact, their knowledge may (far) exceed that of their developers or users. Consequently, a desirable property for an AI system is that it is honest -- that it accurately reports its beliefs about the world. Designing an AI system to be honest may be difficult, especially if we want to ask it questions about latent variables in the environment -- variables which are hidden from the human interacting with it. This gives rise to the problem of eliciting latent knowledge (ELK): the problem of training an AI agent to honestly report its beliefs. In this paper, we make ELK formally precise using Causal Influence Diagrams (CIDs). CIDs can be used to describe the relationship between an agent's training environment and its subjective representation of the world. We use CIDs to formalise the distinction between observable and latent variables, to specify what exactly it means for an agent to be honest, and to formally define goal misgeneralisation. We show that, under certain circumstances, developers can incentivise an agent to honestly answer questions by providing correct feedback during training. However, a natural, but undesirable, way for an agent to generalise is to provide answers which humans would evaluate as true, rather than honest answers. We prove an impossibility theorem stating: There is no feedback-based training strategy that depends only on agent behaviour and with certainty produces an honest agent, even if feedback is perfect during training.

Universal LLM reliability is not a finite-library problem: across all possible tasks, tools, schemas, knowledge sources, and evaluator expectations, new intervention-distinguishable failure modes can appear without bound, so no finite intervention dictionary can guarantee bounded residual error for every such mode. But deployed systems do not operate over the whole universe. They operate inside operationally bounded patches (legal review, medical RAG, code repair, customer-support agents, contract extraction) with recurring tasks, schemas, tools, and evaluator expectations. Within such patches, empirical evidence suggests failures are sparse, repetitive, and concentrated in a small recurring catalogue, so reliability becomes a local catalogue-discovery and intervention-coverage problem rather than an exponential token-length problem. We formalize this transition with two propositions and one corollary. Proposition 1 is the worst-case-mode-wise negative result: no finite intervention dictionary covers every distinguishable failure mode of an unbounded domain. Corollary 1 is the inverse-discovery implication: the logarithmic upper bound on mode discovery cannot accommodate linearly more distinct tail modes without exponentially more observed hard-failure events. Proposition 2 is the positive patch-local result: under log active-mode exposure and head-heavy coverage, a sufficient per-hard-decision intervention budget grows polylogarithmically in sequence length and becomes domain-constant once the patch catalogue saturates. The framework relocates rather than dissolves long-context difficulty: where the number of hard decisions itself grows with task length, reliability remains hard; the contribution is to identify the on-axis intervention rather than to make those regimes easy.

Judgment aggregation studies how to combine individual judgments on logically related propositions into a collective judgment. Classical impossibility results show that sufficiently strong logical interconnections force dictatorship under natural aggregation axioms. In this paper, we ask whether such impossibility can still arise when the objects of aggregation are required to be genuinely modal judgments rather than plain factual propositions. Since modal logic contains propositional logic, this question is meaningful only if one excludes fact-based aggregation in disguise. We show that Arrow-type impossibility already re-emerges in a strikingly sparse modal setting. We prove an impossibility theorem on a simple cyclic frame for an agenda generated from a single propositional variable by repeated applications of a single modal operator, and we further demonstrate this phenomenon for an alternative family of frames satisfying a natural symmetry condition. Thus, even under a modal-operator requirement, semantic structure alone can generate the logical interconnections needed for dictatorship. Technically, our analysis has two layers. First, we prove a semantic reduction theorem showing that certain iterated modal patterns can be collapsed by shifting the evaluation point. Second, building on this reduction, we identify a local-to-global frame mechanism by which frame geometry yields minimally inconsistent modal judgment sets and the strong path-connectivity required for impossibility. The same reduction also turns consistency checking into a small combinatorial covering problem, which yields efficient implementations of non-dictatorial aggregation procedures.

Large language models now write software, draft legal documents, and produce clinical notes, yet fundamental limits, from Turing and Arrow to the No Free Lunch theorems, shape what computation can do. This thesis turns such impossibility results from curiosities into design rules. Its flagship result proves an accuracy ceiling set by architecture alone: past a critical reasoning depth, no amount of training moves it, at any adapter rank, sample size, or loss function. Computable before deployment from layer count and embedding width, this Deterministic Horizon is measured between nineteen and thirty-one across twelve transformer architectures, and fine-tuning on optimal-length traces recovers under four percentage points. The mechanism is a capacity invariant of the residual stream, and an information-theoretic conversion yields super-exponential accuracy decay past the horizon. An unconditional circuit-complexity lower bound for modular exponentiation against constant-depth prime-modulus circuits complements this result. The same argument recasts across subfields: preference learning under any misspecified model jumps discontinuously in sample complexity; multi-stage retrieval pipelines require at least as many independent metrics as stages; standard truthful auctions fail for agents with prompt-dependent valuations; and zero-knowledge verification of neural inference pays a measured overhead of one hundred ten to one hundred ninety times per non-linear activation. Together these form a catalogue of sixteen specifications, each pairing a computable boundary, a quantified violation cost, and a constructive design rule: two compositions are proved, one pairing is an honest obstruction, and four remain open. The impossibility-specification methodology is offered for the generative research programme that trustworthy AI may need. Every fundamental limit of AI is also a design rule.

Impossible objects, geometric constructions that humans can perceive but that cannot exist in real life, have been a topic of intrigue in visual arts, perception, and graphics, yet no satisfying computer representation of such objects exists. Previous work embeds impossible objects in 3D, cutting them or twisting/bending them in the depth axis. Cutting an impossible object changes its local geometry at the cut, which can hamper downstream graphics applications, such as smoothing, while bending makes it difficult to relight the object. Both of these can invalidate geometry operations, such as distance computation. As an alternative, we introduce Meschers, meshes capable of representing impossible constructions akin to those found in M.C. Escher's woodcuts. Our representation has a theoretical foundation in discrete exterior calculus and supports the use-cases above, as we demonstrate in a number of example applications. Moreover, because we can do discrete geometry processing on our representation, we can inverse-render impossible objects. We also compare our representation to cut and bend representations of impossible objects.

As AI systems become increasingly capable, safety strategies must be evaluated not only by how much they reduce present risk, but by whether they could sustain safety once external control can no longer reliably constrain system behavior. This paper addresses that problem by using control theory to clarify, at a structural level, whether externally enforced safety-sustaining strategies can succeed and, if not, what any alternative strategy would have to satisfy in order to be viable. It establishes two main results. First, under explicit premises including a reachability condition, it proves a class-wide external impossibility result: once the system's effects exceed what bounded external control can counteract, no strategy that depends in any degree on continued external enforcement can sustain AI safety. This failure is structural across the entire externally enforced class rather than contingent on any particular strategy. Second, it establishes a conditional class-level necessity result: if at least one candidate safety-sustaining strategy remains after that elimination, then all such remaining strategies must be intrinsic. It then states four structural requirements for viability: safety may not depend on continued external enforcement; the system's terminal objective must be safety-compatible when first formed; that objective must remain stable under self-modification; and safety must continue to be preserved as capability grows. The paper does not propose a complete strategy for sustaining AI safety. Its contribution is to give formal structure to a widely held concern about the limits of external control. It does so by deriving explicit conditional results that identify which safety-sustaining strategies are ruled out and what any remaining strategies must satisfy.

Fairness impossibility results often look like distinct scalar incompatibility statements. We show that several share one RKHS geometry: fairness criteria are linear constraints on conditional mean embeddings, and unequal base rates make the law of total expectation overdetermine those constraints. This view yields four results. The Kleinberg--Mullainathan--Raghavan dichotomy needs only group-conditional unbiasedness, not full calibration. The \emph{Pokémon theorem} shows that a distinct group pair satisfying any finite collection of linear mean-fairness criteria leaves a residual violation witnessed by the MMD, decaying at the Kolmogorov $m$-width rate under spectral regularity. The same tools prove an impossibility for fair feature learning: parity and class-conditional separation in representation space force class collapse under unequal base rates. The approximate relaxations yield signal and error frontiers, allowing a trade-off between real-world estimators and fairness goals. Experiments on standard fairness benchmarks are consistent with our bounds.

Are certain cognitive biases mathematically inevitable consequences of sequential information processing? We prove that primacy effects, anchoring, and order-dependence are architecturally necessary in autoregressive language models due to causal masking constraints. Our three impossibility theorems establish: (1) primacy bias arises from asymmetric attention accumulation; (2) anchoring emerges from sequential conditioning with provable information bounds; and (3) exact debiasing by permutation marginalization requires factorial-time computation, with Monte Carlo approximation feasible at constant per-tolerance overhead. We validate these bounds across 12 frontier LLMs ($R^2 = 0.89$; $Δ$BIC $= 16.6$ vs. next-best alternative). We then derive quantitative predictions from the framework and test them in two pre-registered human experiments ($N = 464$ analyzed). Study 1 confirms anchor position modulates anchoring magnitude ($d = 0.52$, BF$_{10} = 847$). Study 2 shows working memory load amplifies primacy bias ($d = 0.41$, BF$_{10} = 156$), with WM capacity predicting bias reduction ($r = -.38$). These convergent findings reframe cognitive biases as resource-rational responses to sequential processing.

Human-AI teams fail to outperform their best member in 70% of studies, yet no theory specifies when complementarity is achievable. We derive tight bounds for the broad class of confidence-based aggregation rules by integrating signal detection theory with information-theoretic analysis, yielding four results: (1) a complementarity theorem (teams outperform individuals iff error correlation $ρ_{HM} < ρ^*$, with $ρ^* \approx a$ in the symmetric near-chance regime); (2) minimax bounds showing gains scale as $Θ(\sqrt{Δd})$ with metacognitive sensitivity difference; (3) an impossibility result proving no confidence-based aggregation rule achieves complementarity when $ρ_{HM} \geq ρ^*$; and (4) multi-class generalization $ρ^*_K \approx ρ^*/\sqrt{K-1}$. Predictions match observed team accuracy ($R = 0.94$ on ImageNet-16H, $R = 0.91$ on CIFAR-10H) and the multi-class threshold scaling holds on human data ($R = 0.93$, $K = 16$), with robustness under non-Gaussian distributions. The framework explains why complementarity is rare and provides actionable design formulas; results apply to aggregation, not to interactive deliberation that generates novel answers.

Grammar-constrained generation is often combined with local vocabulary masking and speculative decoding, but the resulting sampling law is not the grammar-conditional distribution users usually intend. We show that any speculative decoder with local mask access, Leviathan rejection, and rollback soundness samples from the locally projected distribution $μ^{\mathrm{proj}}$ rather than the grammar-conditional distribution $μ^\star$. This extends the GAD impossibility result to speculative decoding; on Dyck grammars with Qwen3-8B, the total-variation gap can reach 0.996. We identify the future-validity function $Φ_t(y)=\Pr_p[\mathrm{valid\ completion}\mid y]$ as the missing correction statistic. The target distribution is a Doob transform of the base model with $h=Φ$, while local masking corresponds to setting $h$ to one. With exact $Φ$, our oracle decoder FVO-Spec samples exactly from $μ^\star$; with approximate $Φ$, we bound the resulting total-variation error. Because exact future validity is hard for general context-free grammars, we evaluate estimator hierarchies on tractable Dyck and finite JSON languages. OneStep reduces Dyck TV by 14% with under 1% throughput overhead, exact dynamic programming reduces it by 97%, and finite-language correction closes JSON gaps to numerical precision. All fidelity claims are scoped to enumerable grammars and token tries.

An agent's probability report is paid for twice: by a strictly proper scoring rule, and by an approval rule for the decision it triggers. In this classical decision-coupled setting, non-affine approval is known to defeat truthful reporting. We show the conflict is endogenous: when feasible, the welfare-maximizing approval rule is never affine. The distortion, however, is predictable and can be designed around. There is a reserve report at which pretending to be the marginal type costs exactly the approval prize. Approving at or above the reserve screens types perfectly under every strictly proper score, and the reserve does not depend on the type distribution. A Lipschitz rule with a single kink attains first-best exactly; under strict feasibility no continuously differentiable rule does. The binding constraint is steepness, not smoothness. First-best is attainable within a slope budget if and only if the budget is at least the critical slope: the steepest chord of the pretending cost up to the reserve. Below it the welfare loss is cubic in the shortfall. Where the pretending cost is convex up to the reserve, as for Brier, log and power scores, the critical slope is closed-form. The instances are AI-agent oversight and marketplace operation.

We identify and prove a fundamental trade-off governing long-sequence models: no model can simultaneously achieve (i) per-step computation independent of sequence length (Efficiency), (ii) state size independent of sequence length (Compactness), and (iii) the ability to recall a number of historical facts proportional to sequence length (Recall). We formalize this trade-off within an Online Sequence Processor abstraction that unifies Transformers, state space models, linear recurrent networks, and their hybrids. Using the Data Processing Inequality and Fano's Inequality, we prove that any model satisfying Efficiency and Compactness can recall at most O(poly(d)/log V) key-value pairs from a sequence of arbitrary length, where d is the model dimension and V is the vocabulary size. We classify 52 architectures published before March 2026 into the triangle, showing that each achieves at most two of the three properties and that hybrid architectures trace continuous trajectories in the interior. Experiments on synthetic associative recall tasks with five representative architectures validate the theoretical bound: empirical recall capacity lies strictly below the information-theoretic limit, and no architecture escapes the triangle.

We report a systematic failure mode in predictive representation learning. Across 2695 neural network configurations trained to predict linear-Gaussian dynamics, the optimal encoder tracks the environment rather than the system it is meant to model. The mean causal fidelity -- the fraction of encoder sensitivity allocated to system degrees of freedom -- is 0.49, and only 2.5% of configurations exceed 0.70. The failure intensifies with dimension: at N=100, the optimal encoder becomes causally blind (fidelity ~10^{-8}) while achieving 92% lower prediction error than the causal representation. We prove this is not an optimization artifact but a structural property of the predictive objective: when environment modes are slower or less noisy than system modes, every minimizer of the population risk encodes the former. The set of dynamics exhibiting this predictive-causal gap is open and of positive measure in parameter space. In a nonlinear Duffing-GRU sweep, unconstrained predictors learn environment-dominant representations in 55% of tasks (95% CI 41--68%) versus 24% under operational grounding (p=2.3e-3); the median out-of-distribution MSE inflation under environment shift is 1.82x versus 1.00x. Operational grounding -- restricting the loss to system observables -- partially suppresses the gap, but causal fidelity is never recovered without an explicit system-environment boundary. The results identify the predictive-causal gap as a structural limit of learning, with implications for self-supervised representation learning, world models, and the scaling paradigm.