1 Introduction↩︎

Foundation-model evaluation routinely reports per-question accuracy, calibration, and proper-scoring statistics; formal, instance-wise guarantees on system-level performance under composition of multiple model calls are comparatively scarce. In multi-component agents, planners route retrieval, arithmetic, and probability assessment to specialist subagents or tools, and each component handles only part of a joint question. Even when every component is calibrated and internally coherent on its assigned questions, the aggregated belief need not be: a research component emitting \(P(\text{Republican}){=}0.6\) and a forecasting component emitting \(P(\text{Democrat}){=}0.6\) produce a \(1.2\)-mass quote that no probability measure can assign, inducing a Dutch-book exposure [1] that arises strictly between components. Per-component coherence does not, in general, repair the composed system; we call this regime locally coherent, globally incoherent; 1 shows a concrete instance from the planner-discretion harness of §5.3.

Per-component calibration, self-consistency [2], and conformal prediction [3] act on individual model outputs and preserve only per-output coherence properties; cross-component logical constraints are invisible to all three. We identify the compositional residual \(\varepsilon^\star\), the \(L_2\) distance from the composed quote to the joint coherent polytope, as a runtime, distribution-free certificate of system-level coherence under owner-selected aggregation; it is computable from the composed quote and the cross-component constraints alone. As a constructive repair preserving specialist routing, we use a hierarchical Boyle–Dykstra projection. We pair these with closed-form bounds, an anytime-valid sequential test, and a decomposable benchmark whose hardness is predicted by the theory.

A product-structure dichotomy (1) characterises when local coherence suffices: under owner-selected aggregation (each component owns a subset of joint coordinates; the aggregator just selects what each owner already wrote), component coherence guarantees system coherence for all inputs if and only if the joint polytope factorises as a Cartesian product of local polytopes. Under any tighter coupling, locally coherent component forecasts exist whose composition is globally incoherent, and \(\varepsilon^\star\) certifies this failure mode (Cor. 1). Cor. 5 converts the existential dichotomy into a closed-form magnitude prediction (Rayleigh-quotient form) computable from the specialist panel covariance alone.

Operationally, \(\varepsilon^\star\) yields a bound \(\mathrm{Exposure}^\star\!\le\!\sqrt{m^\star}\,\varepsilon^\star\) (Cor. 2) on system-level Dutch-book exposure under the Fundamental Theorem of Asset Pricing (FTAP); the hierarchical Boyle–Dykstra repair (3) drives this bound to the numerical floor, and the corresponding e-process (a sequential test controlling false positives at any stopping time) gives an anytime-valid coherence test (§3.7, 5). The convex-analytic machinery is classical; the contribution is the operational reframing of \(\varepsilon^\star\) as a runtime certificate, the dichotomy of 1, the falsifiable magnitude prediction of Cor. 5, and the decomposable benchmark whose hardness ordering the theory predicts.

Empirically, we evaluate \(1{,}876\) ensemble cliques drawn from Paleka [4] and Polymarket across four contemporary foundation models, stratified by four logical-relation classes. The empirical hardness ordering (partition \(>\) negation \(>\) disjunction \(>\) conjunction) tracks how tightly each polytope constrains the joint quote, as predicted by 1, and Cor. 5’s magnitude prediction matches the observed expected residual within \(7\%\) on three of the four relation classes (the conjunction under-shoot at \(0.83\!\times\) is itself predicted by the Cor. 5 interior-\(\bar\Pi\) regime). The exposure bound \(\sqrt{m^\star}\varepsilon^\star\) averages \(0.137\) under naive composition (2b) and is reduced to the QP solver floor by hierarchical projection. We report per-cell Brier with Diebold–Mariano significance (the standard paired test for forecast comparison), and validate with a same-model decoupling control, leakage filtering, and a tool-augmented A/B; full protocols and a reproducibility manifest are in the appendices.

2 Preliminaries↩︎

Throughout, starred objects denote the composed system, while subscripts denote individual components.

2.0.0.1 Cliques and the coherent polytope.

A clique \(C\) is a tuple \((Q_1,\dots,Q_m, R)\) of \(m\) Bernoulli questions and logical relations \(R\). By [1], the set \(\mathcal{M}_C= \{r\in[0,1]^m : \exists\,\mu\in\Delta(\{0,1\}^m)\; \text{consistent with } R\}\) is a closed convex polytope. We denote the \(L_2\) projection \(\Pi_C: [0,1]^m \to \mathcal{M}_C\) and the residual \(\varepsilon_C(\hat{p}) := \|\hat{p}- \Pi_C(\hat{p})\|_2\).

2.0.0.2 \(K\)-sample empirical marginal.

A forecaster supplies \(K\) i.i.d.samples per question, yielding \(\hat{p}\in [0,1]^m\) with \(\hat{p}_j = K^{-1}\sum_k Y_{j,k}\) and population limit \(p_F:= \lim_{K\to\infty}\hat{p}\). Generically \(\hat{p}\notin \mathcal{M}_C\) even when \(p_F\in\mathcal{M}_C\) (due to finite-sample fluctuation).

2.0.0.3 Within-component JCD.

Joint-Coherent Decoding is the \(L_2\) projection of \(K\)-sample LLM marginals onto the local coherent polytope \(\mathcal{M}_C\). The compositional layer treats JCD as the local repair operator; it inherits a Pythagorean Brier improvement (Theorem 4, App. 7) and a finite-state no-Dutch-book bound via the Fundamental Theorem of Asset Pricing (FTAP; Cor. 6) bounded by \(m\cdot\varepsilon^{\mathrm{KKT}}\!\le\!1.4\!\times\!10^{-5}\) at the default tolerance of OSQP, the quadratic-program solver we use throughout.

2.0.0.4 Multi-component agent.

The agent has \(k\) sub-models indexed by \(a \in [k]\). Sub-model \(a\) emits an empirical marginal \(\hat{p}^{(a)} \in [0,1]^{m_a}\) on its local question set \(\mathcal{Q}_a\) with local polytope \(\mathcal{M}_a\subseteq [0,1]^{m_a}\); component-level JCD produces \(\Pi_a(\hat{p}^{(a)}) \in \mathcal{M}_a\). The agent’s joint question set is \(\mathcal{Q}^\star = \bigcup_a \mathcal{Q}_a\) with size \(m^\star = |\mathcal{Q}^\star|\) and joint polytope \(\mathcal{M}^{\star}\subseteq [0,1]^{m^\star}\), which respects (i) each local relation lifted to \(\mathcal{Q}^\star\) and (ii) all cross-component coupling constraints \(\mathcal{C}\): identifications of shared questions, cross-component logical relations like \(Q^\star = Q_a \wedge Q_b\), and partitions spanning components.

2.0.0.5 Aggregator.

An aggregator \(\mathcal{A}: \prod_a [0,1]^{m_a} \to [0,1]^{m^\star}\) assembles the agent’s quoted marginal vector. The aggregator \(\mathcal{A}\) is the owner-selected coordinate selector of Definition 2; this covers the specialist-routing and tool-call composition patterns used in our experiments. We focus on owner-selection because it is the practical pattern under compute budgets: each component owns a subset of joint coordinates and there is no redundant elicitation across specialists. A coord-wise averaging aggregator \(\mathcal{A}^\mathrm{avg}(\Pi_1,\dots,\Pi_k) = k^{-1}\sum_a \Pi_a\) on the same inputs would already lie in \(\mathcal{M}^{\star}\) by convexity, so the failure mode of 1 is structurally absent under averaging; we verify this empirically (4). The cost of averaging is \(k\times\) elicitation per coordinate, which is the regime specialist-routing is designed to avoid.

2.0.0.6 The two candidate workflows.

With \(\hat{p}= (\hat{p}^{(1)},\dots,\hat{p}^{(k)})\) and the aggregator \(\mathcal{A}\) fixed, two natural projections into \(\mathcal{M}^{\star}\) arise: \[\begin{align} \text{Local-then-global:}\quad & \Pi^{\star}(\mathcal{A}(\Pi_1\hat{p}^{(1)},\dots,\Pi_k\hat{p}^{(k)})), \tag{1}\\ \text{Direct global:}\quad & \Pi^{\star}(\mathcal{A}(\hat{p}^{(1)},\dots,\hat{p}^{(k)})). \tag{2} \end{align}\] The substantive question is when, and by how much, these differ.

3 Method↩︎

Throughout this section we assume the coupling set \(\mathcal{C}\) is explicitly specified; the regime of implicit \(\mathcal{C}\) recovered from agent transcripts is discussed in §6.

3.1 When local coherence composes↩︎

3.1.0.1 Lifted local feasible sets.

For each component \(a\), lift \(\mathcal{M}_a\subseteq [0,1]^{m_a}\) to \(\mathcal{M}_a^{\uparrow} \subseteq [0,1]^{m^\star}\) by leaving the non-component coordinates free. Write \(\mathcal{M}^{\boxtimes}:= \bigcap_a \mathcal{M}_a^{\uparrow}\) for the joint feasibility implied by the local constraints alone (no cross-component coupling). \(\mathcal{M}^{\star}\subseteq \mathcal{M}^{\boxtimes}\) always; equality holds iff the coupling set \(\mathcal{C}\) is trivial.

The forward direction is the block-product factorization of \(L_2\) projection in Hilbert space [5]; the contribution is operational, identifying when local coherence is sufficient and when \(\varepsilon^\star\) is an appropriate online certificate. We also refer to 1 as the non-commutation theorem, since it characterizes exactly when local-then-global \(L_2\) projection commutes with the aggregator (and when it does not). The reverse direction is existential: it shows local coherence is not sufficient under coupling, but does not characterize the typical magnitude of \(\varepsilon^\star\). Section 5 estimates this empirically.

3.2 Exposure interpretation↩︎

The condition \(p^\star \in \mathcal{M}^{\star}\) is the precise assumption under which Cor. 3 transfers from the geometric to the predictive setting. Empirically (§5), the lift holds with \(p\!<\!10^{-15}\) on negation, conjunction, and partition; on disjunction the gain is borderline (\(\Delta\mathrm{Brier}\!=\!{+}0.0027\), \(p\!=\!0.07\)), diagnosing a small fraction of disjunction cliques whose resolved labels deviate from the strict \(A\!\vee\!B\) relation: this is exactly the \(p^\star\!\notin\!\mathcal{M}^{\star}\) regime in which Cor. 4 predicts the lift can reverse.

3.3 Disagreement controls the residual↩︎

Let \(\Delta_j\) denote the per-coordinate disagreement between the assigned LLM and a globally-coherent reference \(\beta\). For each relation class the bound takes a closed form keyed to the constraint normal: \(|\Delta_1{+}\Delta_2|/\sqrt{2}\) on negation, \(|\sum_j\Delta_j|/\sqrt{m^\star}\) on partition, and \(\|\Delta\|_2\) on the Fréchet conjunction/disjunction polytope (full derivations in App. 9).

Empirically the bound holds on \(100\%\) of the \(1{,}876\) cliques. Mean bound-to-residual ratios are \(1.30\) (negation), \(2.34\) (disjunction), \(3.00\) (conjunction), and \(4.39\) (partition); the bound is tightest where the constraint normal aligns with the disagreement direction.

3.4 Quantitative magnitude prediction↩︎

Under uniform random owner-selection, the expected residual admits a closed form keyed to the empirical specialist covariance restricted to the constraint normal that defines the joint polytope. Let \(\bar\Pi := k^{-1}\sum_a \Pi_a \in \mathcal{M}^{\star}\) (by convexity) and \(D := \mathrm{diag}(\Sigma_\Pi)\) where \(\Sigma_\Pi := k^{-1}\sum_a (\Pi_a - \bar\Pi)(\Pi_a - \bar\Pi)^\top\). Independent assignment across coordinates makes \(D\) the effective covariance: \(\mathbb{E}_\sigma[(x-\bar\Pi)(x-\bar\Pi)^\top] = D\).

Proof sketch. Equality: \(\varepsilon^\star = |a_R^\top(x-\bar\Pi)|/\|a_R\|\), square and use \(\mathbb{E}_\sigma[a_R^\top(x-\bar\Pi)(x-\bar\Pi)^\top a_R]=a_R^\top D a_R\). Inequality with \(\bar\Pi\) on the boundary: \(\varepsilon^\star = (a_R^\top(x-\bar\Pi))_+/\|a_R\|\); for a zero-mean random variable \(X\), \(\mathbb{E}[X_+^2] = \tfrac12 \mathbb{E}[X^2]\) when \(X\) is symmetric, otherwise \(\mathbb{E}[X_+^2]/\mathbb{E}[X^2] \in [0,1]\) depending on the upper-tail mass. Full proof and the generic-bound case in App. 14.

Combined with Cor. 4, this yields a closed-form Brier tax \(\mathbb{E}[\mathrm{Brier}(r,y) - \mathrm{Brier}(\Pi^{\star}(r),y)] \ge \kappa_R\, a_R^\top D a_R/\|a_R\|^2\) when \(p^\star\!\in\!\mathcal{M}^{\star}\), computable from the panel covariance alone. The three couplings recurring in our experiments are shared-question disagreement (two specialists quote the same \(Q\) at different prices), negation across components (\(p_a\!+\!p_b\!=\!1\)), and partition across components (\(\sum_a p_a\!=\!1\)).

3.5 Hierarchical repair↩︎

The repair is constructive. Let \(\Pi_\mathcal{C}: [0,1]^{m^\star}\to \mathcal{C}\) denote the \(L_2\) projection onto the coupling set, an intersection of finitely many half-spaces.

3.5.0.1 Per-iteration cost.

Each \(\Pi_a\) and \(\Pi_\mathcal{C}\) is a small QP. The Paleka (neg/and/or) and Polymarket (partition) relation classes admit closed-form projections (App. 9); we use direct \(\Pi^{\star}\) via OSQP [8] for uniformity across relation classes, since the partition simplex projection and the negation closed form are mathematically equivalent to OSQP’s converged solution at the reported tolerance.

3.5.0.2 Where the cyclic machinery actually bites.

On the equality-coupled relations (negation, partition) the joint projection collapses to a single closed-form step (App. 9): partition repair is simplex projection [9], which in the no-clipping majority is literally subtracting \((\sum_i p_i\!-\!1)/m^\star\) from each coordinate. The Boyle–Dykstra cycle is mathematically equivalent to this but does real iterative work only on the Fréchet polytopes (conjunction, disjunction), where multiple half-spaces can be simultaneously active. The headline residual is therefore largest on the relation (partition) where the repair is simplest, and smallest where the cyclic machinery is needed; the contribution is operational, not algorithmic.

3.6 Runtime use↩︎

The residual supports three deployment modes that share the same computation and differ only in how the agent acts on it. In monitor mode the agent logs \(\varepsilon^\star\) and surfaces large residuals to an operator. In repair mode it replaces the naive composed quote with \(\Pi^{\star}(\cdot)\) before passing it downstream. By 3 and Cor. 2, the repaired quote lies in \(\mathcal{M}^{\star}\) up to solver tolerance and the exposure diagnostic is driven to numerical zero. In abstain-or-escalate mode it refuses to act when \(\varepsilon^\star\!>\!\tau\), a risk–coverage knob whose calibration to budgeted exposure is outside our scope.

3.7 Sequential monitoring: anytime-valid coherence test↩︎

In long-horizon agent deployments the agent generates a compositional residual stream \((\varepsilon^\star_t)_{t\ge 1}\) across successive composition steps. The spatial guarantee of 3 certifies incoherence at each step; the following temporal guarantee lets an operator certify persistent incoherence across the stream without fixing a stopping horizon in advance.

Under the null \(H_0\) that the composed population quote is jointly coherent (\(p_F^\star\!\in\!\mathcal{M}^{\star}\) at each step), the e-process \(E_t(\lambda) := \prod_{s=1}^{t} \exp\!\bigl(\lambda(\varepsilon_s^{\star 2} - m_s^\star/(4K_s)) - \lambda^2 m_s^\star/(2K_s)\bigr)\) is a non-negative \(\mathcal{F}_t\)-supermartingale (Lemma 1, App. 10). By Ville’s inequality [10], [11], the stopping rule \(\tau_\alpha := \inf\{t : E_t(\lambda)\ge 1/\alpha\}\) controls Type-I error at level \(\alpha\) uniformly over all stopping times (5). An operator monitors \(E_t\) online and escalates the agent pipeline when it crosses \(1/\alpha\); no fixed-horizon commitment is required. The full derivation and empirical demonstration on four LLM residual streams appear in 10.

4 Experimental setup↩︎

We evaluate on Paleka [4] (negation, conjunction, disjunction, paraphrase; \(134\) cliques per checker) and Polymarket (\(67\) partition cliques after leakage control, expanded to \(268\) ensemble instances by four seeds). Both benchmarks were chosen because they target the regime the theory predicts will be tightest: Paleka was constructed by [4] for cross-question consistency checking, and Polymarket’s partition events are the unit-mass constraint regime (\(\sum_i p_i = 1\)) where the polytope coupling is most restrictive.

4.0.0.1 Models.

Anthropic claude-haiku-4-5-20251001 (via the Anthropic Messages API), OpenAI gpt-5.4-mini and gpt-5.4-nano (via Azure OpenAI), and Meta llama-3.3-70b-versatile (via Groq). Frontier-panel evaluations additionally use OpenAI gpt-5.5 (via Azure OpenAI). These are the public API model identifiers as returned by each provider on 30 April 2026; raw responses, prompts, and per-clique residuals are released with the supplementary material. Each specialist produces \(K{=}8\) verbalized probability samples per question at temperature \(0.7\).

4.0.0.2 Leakage control.

Polymarket events were filtered so that every market in every retained event resolved strictly after our model snapshot date of 30 April 2026; the inclusion cutoff was 1 May 2026, and all retained events resolved on or after that date. Paleka cliques inherit the resolution-status verification of [4]. Full filtering protocol in Appendix 11.

4.0.0.3 Scope of the evaluation.

The bulk of our evaluation is a controlled routing simulation: clique coordinates are assigned i.i.d.to specialists from the panel, and the composed quote is reassembled by owner-selected aggregation. The planner-discretion harness (§5.3) and routing-protocol ladder (App. 21) are the closest stand-ins for a deployed multi-component agent; both are small (\(n\!=\!20\) and \(n\!=\!100\) partitions respectively). Claims about end-to-end deployed agents should be read with that scope in mind.

4.0.0.4 Compositional ensemble.

For each clique with \(m\) questions, we draw \(4\) random seeds; each seed assigns each of the \(m\) questions to one of the \(4\) LLMs uniformly i.i.d.; the agent’s quoted marginal on coordinate \(j\) is the JCD-projected marginal of the assigned LLM on that coordinate. The source forecast from each LLM is JCD-coherent on the full clique before coordinate selection; the ensemble residual \(\varepsilon^\star\) therefore reflects exclusively cross-component disagreement.

4.0.0.5 Benchmark properties.

The four relation classes admit closed-form local projections (App. 9) and span a range of polytope geometries from a single equality (negation), to a high-dimensional simplex (partition), to the Fréchet box (and/or). The coupling set \(\mathcal{C}\) controls the gap between \(\mathcal{M}^{\star}\) and \(\mathcal{M}^{\boxtimes}\), and random vs.structured routing brackets cross-model mixing. 1’s prediction \(\varepsilon^\star\!\equiv\!0\) when \(\mathcal{M}^{\star}\!=\!\mathcal{M}^{\boxtimes}\) is verified at the QP solver floor by the no-composition reference (2b, gray dashed), a falsifiable test of the dichotomy.

5 Results↩︎

The random-assignment experiment (\(N\!=\!134\) cliques \(\times\,4\) seeds \(=\!536\) per relation for Paleka, plus \(N\!=\!67\!\times\!4\!=\!268\) for Polymarket; total \(1{,}876\)) and structured-routing variant (\(N\!=\!469\), one deterministic seed) bracket cross-model mixing.

5.1 Compositional residual and controls↩︎

5.1.0.1 Compositional residual under random routing.

The compositional residual is strictly positive on \(94\%\) (partition), \(66\%\) (negation), \(43\%\) (disjunction), \(33\%\) (conjunction) of cliques (2a). Mean \(\varepsilon^\star\) ranges \(0.058\) (and) to \(0.118\) (partition) on the full \(N\!=\!1{,}876\) ensemble. The projection-step ablation (App. 15, Table [tab:proj95ablation]) reports \(0.054/0.098\) on the same relations because it is restricted to the resolved-label subset on which paired Brier is computed (\(\sim\!85\%\) of the ensemble); the \(\sim\!20\%\) magnitude shift on partition reflects that unresolved partition cliques tend to be larger-\(m^\star\) events with proportionally higher residuals. None is visible per-component, where each \(\Pi_a\!\in\!\mathcal{M}_a\) has zero residual. Hierarchical JCD reduces the mean exposure bound \(\sqrt{m^\star}\varepsilon^\star\) to \(1.7\!\times\!10^{-14}\) on the \(1{,}876\)-clique ensemble (per-clique post-repair \(\varepsilon^\star\) reaches the OSQP primal-residual floor of \(\le\!1.5\!\times\!10^{-16}\) on partition and negation, where closed-form projections terminate at machine precision; the ensemble mean is dominated by the and/or relations, which bottom out at the iterative-OSQP tolerance instead), matching the no-composition reference (a single LLM on every coordinate; 2b, gray) while retaining specialist routing.

5.1.0.2 Empirical validation of the magnitude prediction.

Cor. 5’s Rayleigh-quotient prediction \(\kappa_R\,a_R^\top D a_R/\|a_R\|^2\) matches the observed expected residual \(\mathbb{E}_\sigma[(\varepsilon^\star)^2]\) across all four relation classes (Table 1; per-clique Pearson \(0.69\)–\(0.89\)). The conjunction under-shoot reflects \(\bar\Pi\) typically interior to the binding Fréchet halfspace; on the remaining three relations the prediction is tight to within \(7\%\). Both \(\varepsilon^\star\) and the \(\sqrt{m^\star}\varepsilon^\star\) exposure bound are therefore predictable from the panel covariance alone, before any routing.

5.1.0.3 Same-model decoupling control.

To separate coordinate isolation from cross-model capability divergence, we re-ran with four fresh-seed runs of one model (Claude-Haiku-4.5, \(N{=}80\) cliques per relation, \(40\) partitions). Mean \(\varepsilon^\star\) falls to \(0.025/0.019/0.040/0.058\) on neg/and/or/partition: \(22\)–\(60\%\) of the residual persists at same-model, with cross-model heterogeneity amplifying by \(1.7\)–\(4.5\times\). A greedy-decoding (\(T{=}0\)) control retains \(\varepsilon^\star = 0.025\)–\(0.096\) across relations (App. 16); on negation and disjunction \(T{=}0\) amplifies the residual relative to \(T{=}0.7\), ruling out sampling noise as the source.

5.1.0.4 Compositional Brier against resolved labels.

We adopt the convention \(\Delta\mathrm{Brier} = \mathrm{Brier}_{\mathrm{JCD}} - \mathrm{Brier}_{\mathrm{naive}}\) throughout, so a negative \(\Delta\) indicates JCD improvement. Paired over the \(1{,}876\) cliques, hierarchical JCD reduces compositional Brier on Neg (\(\Delta\mathrm{Brier}{=}-0.0137\), \(p{<}10^{-43}\)), And (\(\Delta\mathrm{Brier}{=}-0.0076\), \(p{<}10^{-16}\)), and Partition (\(\Delta\mathrm{Brier}{=}-0.0048\), \(p{<}10^{-23}\)). Or is marginal (\(\Delta\mathrm{Brier}{=}{+}0.0027\), \(p{=}0.07\)): this is the predicted reversal regime of Cor. 4, where label noise on a small fraction of disjunction resolutions pushes \(p^\star\) outside \(\mathcal{M}^{\star}\), and Cor. 4’s condition-and-reversal allows projection to amplify error there rather than reduce it.

5.1.0.5 Disagreement-driven mechanism.

Per-clique \(\varepsilon^\star\) regressed on the assignment-aware reference disagreement \(\min_\beta\|x{-}\Pi_\beta\|_2\) (Prop. 2) gives \(R^2{=}0.93\) on negation (slope \(0.83\)) and \(R^2{=}0.58\)–\(0.64\) on and/or/partition (slope \(0.59\)–\(0.63\)). The constraint normal aligns most closely with the disagreement direction on negation; on higher-dimensional polytopes the projection contracts components of disagreement orthogonal to the binding halfspace, loosening the bound.

5.1.0.6 Robustness to routing topology and retrieval grounding.

Under structured routing (one model per relation family; App. 18), \(\varepsilon^\star\!>\!0\) on \(98.5\%\) of negation, \(36\%\) partition, \(16\%\) disjunction, \(12\%\) conjunction cliques; mean \(\sqrt{m^\star}\varepsilon^\star\!=\!0.117\) vs.\(0.137\) under random routing. Retrieval grounding (top-\(5\) DuckDuckGo snippets per specialist) on \(30\) partitions shifts mean \(\varepsilon^\star\) from \(0.260\) to \(0.283\); \(20/30\) partitions are equal-or-worse with retrieval (App. 22).

5.1.0.7 Routing-protocol ladder.

On \(100\) multi-candidate partitions (App. 21) we walk the routing protocol from fixed naive owner-selection to autonomous tool-using planners. Prompt-engineering under owner-selection drops mean \(\varepsilon^\star\) from \(0.235\) (isolated) to \(0.090\) (listed) to \(0.081\) (full); isolated \(\to\) listed is significant (\(p\!<\!10^{-11}\)), listed \(\to\) full is not (\(p\!=\!0.18\)), and \(99\%\) of partitions retain \(\varepsilon^\star\!>\!10^{-3}\) under full. Tool-using planners with delegate-and-submit tools further reduce \(\varepsilon^\star\): a non-reasoning planner reaches \(\varepsilon^\star\!=\!0.009/0.004\) at \(\sim\!4.8\!\times m\) calls; a frontier reasoning planner reaches \(0/200\) positive residuals at \(\sim\!6.7\!\times m\) calls. The hierarchical projection reaches \(\varepsilon^\star\!\to\!0\) at \(1\!\times m\) specialist calls (3).

5.2 Intuitive LLM-side mitigations fail↩︎

The retrieval-grounding and routing-ladder results above each isolate one obvious LLM-side fix. We collect them with a third candidate, delegating coherentisation to an aggregator LLM, and compare all three to the geometric repair. The comparison is deliberately stacked in favour of the geometric repair: the projection is handed the explicit coupling set \(\mathcal{C}\) as a typed input, while the three prompt-based methods have to infer it from natural-language context. The genuinely hard regime where \(\mathcal{C}\) is implicit in free-form planner-tool transcripts is outside the scope of all four interventions here and is discussed as future work (§6). 2 reports mean \(\varepsilon^\star\), the fraction of partitions retaining \(\varepsilon^\star\!>\!0.05\), and the fraction regressing relative to naive composition. Retrieval grounding regresses on \(20/30\) partitions: mean \(\varepsilon^\star\) rises from \(0.260\) to \(0.283\) on the matched \(30\)-partition baseline. (Table 2’s “Naive” row reports \(0.214\), the \(100\)-partition mean used for the LLM-aggregator comparison; the two subsets are flagged in the caption.) Partition-aware prompting helps on average (\(0.066\)) but \(53\%\) of partitions still exceed the \(0.05\) band and \(5/30\) regress. LLM-as-aggregator reduces mean \(\varepsilon^\star\) to \(0.028\) but \(15\%\) of partitions still exceed \(0.05\) and \(7/100\) regress against the naive baseline. The geometric repair eliminates the residual to the QP solver floor (\(\le 1.5\!\times\!10^{-16}\)) at \(1\)ms per partition with no extra LLM call. Prompt-only and aggregator-LLM mitigations can harm already-coherent quotes: across both prompt-side interventions, \(9/10\) worsening cases had baseline \(\varepsilon^\star\!<\!0.05\), so \(\varepsilon^\star\!>\!\tau\) also serves as a gate for when these prompt-based repairs are safe to apply.

5.3 Planner-discretion harness↩︎

To check that the failure is not an artefact of random assignment, we re-run the experiment with a planner LLM (Claude-Haiku-4.5) that, for each of \(20\) live-style partitions, picks a specialist per outcome from the four-LLM roster and emits its own per-specialist context block; specialists then sample \(K{=}8\) verbalized probabilities under the planner’s prompt and we compute \(\varepsilon^\star\) as in §[subsec:exp95setup]. Mean \(\varepsilon^\star = 0.113\) versus \(0.231\) under random assignment on the same partitions (\(15/20\) improve), but \(\varepsilon^\star\!>\!0\) on \(20/20\). The planner’s routing histogram (GPT-5.4-mini \(50\%\), Claude-Haiku \(31\%\), Llama-70b \(13\%\), GPT-nano \(6\%\)) shows no self-routing.

5.4 Mechanism probe: coupling-visible elicitation↩︎

The mechanism claim of §3 is that \(\varepsilon^\star\) arises because specialists do not see the coupling constraint. We test this directly. On the same \(20\) partitions of §5.3, we re-elicit each specialist’s quote under two paired conditions. Blind keeps the planner-chosen routing of §5.3 but strips the planner-emitted context, so the specialist sees only its assigned outcome as a single Bernoulli question. Informed keeps the same routing and additionally shows the specialist the partition label, all sibling outcomes, the explicit \(\sum_i p_i\!=\!1\) constraint, and the peers’ Blind quotes on the other outcomes. Each (partition, condition) cell is replicated across four independent \(K\!=\!8\) sampling rounds, paired Blind\(\to\)Informed within each round. Informed varies two things at once (coupling visibility and peer-quote disclosure), so the comparison identifies coupling-visible elicitation with peer context as the intervention, not constraint visibility in isolation; full setup in App. 25.

Disclosing the coupling constraint (and peer quotes) reduces \(\varepsilon^\star\) as the theory predicts: paired mean \(\varepsilon^\star\) falls from \(0.519\) Blind to \(0.298\) Informed on \(16/20\) partitions, with \(1\) unchanged and \(3\) worsened. The Blind mean of \(0.519\) exceeds both the random-assignment mean (\(0.231\)) and the planner-harness mean (\(0.113\)) on these same \(20\) partitions. Two protocol differences combine: the planner harness lets a planner LLM see the partition when choosing routing (and emit its own per-specialist context), while Blind strips all such context; and the four Blind seeds are independent \(K\!=\!8\) sampling rounds on the same planner-chosen routing, so cross-seed variance does not average out as it does over random reshuffles. We did not isolate the contributions further; the residual gap is observed, not theory-derived. The three worsening partitions are non-payoff multi-way forecasts where Blind specialists were already roughly normalised; per-partition detail in App. 25. The intervention is partial, not a fix: \(\varepsilon^\star\) remains positive on every Informed partition, consistent with the LLM-side mitigations result (§5.2) that no purely prompt-based intervention eliminates the residual. The geometric repair remains the only intervention that drives \(\varepsilon^\star\) to the QP floor on every partition.

5.5 Frontier-panel rerun↩︎

To test whether capability scaling closes the gap, we rerun the random-assignment ensemble on the \(67\) Polymarket partition cliques with a frontier-only specialist roster: Claude-Opus-4.7, GPT-5.5 (reasoning disabled for apples-to-apples), DeepSeek-V3.2, and Llama-4-Maverick-17B-128E (FP8). The protocol matches §[subsec:exp95setup] exactly (\(K{=}8\), simplex pre-projection, \(4\) seeds, \(268\) ensemble bets). The residual persists: \(\varepsilon^\star\!>\!0\) on \(97.8\%\) of the \(268\) clique-seed bets, slightly above the mid-tier panel’s \(94\%\) on the same Polymarket slice. Mean \(\varepsilon^\star\) drops from \(0.118\) to \(0.072\) (\(-39\%\)): frontier models reduce residual magnitude on partitions but do not eliminate the failure mode. Brier improvement under hierarchical JCD remains highly significant (\(\Delta\mathrm{Brier} = -0.012\), \(95\%\) paired-bootstrap CI \([-0.015, -0.009]\)); the log-payoff gain trends positive (\(+0.107\) nats per bet) but is noisier on this narrower slice. Table 4 reports the full frontier-vs-mid-tier comparison.

5.6 Downstream-decision regret↩︎

For each of \(1{,}770\) resolved ensemble bets from §4 (we drop \(106/1{,}876\) where the assigned specialist lacks a forecast for the assigned coordinate, a pipeline artefact on partition events) we convert the composed quote \(\hat{p}^\star\) into a proportional allocation \(w_i = \max(\hat{p}^\star_i,0)/\sum_j\max(\hat{p}^\star_j,0)\) and measure (i) Brier loss against the binary resolution vector and (ii) the realised log-payoff \(\log w_{\text{winner}}\) on the \(1{,}252\) bets with a unique YES. The proportional rule passes the agent’s incoherent quote through to the bettor’s allocation, which exposes the behavioural cost of compositional incoherence in its rawest form; stricter rules (truncated Kelly, max-entropy) that themselves coherentise the quote partially absorb this cost into the bettor’s strategy, so their log-payoff gains under JCD are correspondingly smaller (Table 6).

Hierarchical JCD lowers mean Brier (\(\Delta\mathrm{Brier}=-0.018\), paired-bootstrap \(95\%\) CI over (clique, seed): \([-0.021,-0.015]\)) and improves mean log-payoff by \(+0.115\) nats per bet (\([+0.056,+0.187]\)). JCD matches the single-LLM oracle’s gain (\(\Delta\mathrm{Brier}= -0.017\), CI \([-0.035,-0.000]\)) with a confidence interval \(\sim\!6\times\) narrower. The effect concentrates on partitions (naive loses \(0.60\) nats/bet vs.JCD, CI \([+0.15, +1.15]\), \(n{=}162\)). The Pythagorean Brier improvement of Cor. 3 is thus realised on resolved labels, not only on the synthetic \(\varepsilon^{\star 2}\) slack.

5.6.0.1 Regret stratifies by certificate.

The realised regret rises monotonically with \(\varepsilon^\star\) (Table 5): the bottom two quartiles (where \(\varepsilon^\star\!\le\!0.004\)) carry no measurable Brier or log-payoff gain, while the top quartile (where \(\varepsilon^\star\!>\!0.154\)) accounts for \(0.054\) Brier and \(0.221\) nats of regret per bet. This justifies the runtime-gating recipe of §5.7: applying the projection only when \(\varepsilon^\star\!>\!\tau\) targets precisely the high-regret bets.

5.6.0.2 Allocation-rule sensitivity.

The log-payoff gain depends on how a downstream bettor allocates on top of the agent’s quote. Brier is allocation-rule independent (\(-0.018\) across all rules). Under the proportional rule (the agent’s incoherence is passed through as allocation slack), JCD beats naive by \(+0.168\) nats/bet on the \(1{,}252\) unique-YES subset (\(95\%\) CI \([+0.083, +0.263]\)); the \(+0.115\) figure cited in the abstract is the same comparison averaged over the full \(1{,}770\)-bet ensemble (non-unique-YES bets contribute zero log-payoff). Under stricter rules that themselves coherentise the quote (truncated Kelly, max-entropy), the gain shrinks to \(+0.006\) nats: the bettor has already absorbed the agent’s incoherence into its own strategy (Table 6).

5.7 Runtime gating thresholds↩︎

The deployment recommendation is to apply repair (or escalate) only when \(\varepsilon^\star\!>\!\tau\). We calibrate \(\tau\) on the 1,770-bet ensemble of §5.6, treating top-quartile log-payoff regret as the harm signal. Brier-regret is mechanically tied to \(\varepsilon^{\star 2}\) by Cor. 3; log-payoff regret depends on which coordinate resolved YES and is therefore not geometrically tied to \(\varepsilon^\star\), making it the genuinely informative discrimination test.

5.7.0.1 Discrimination.

\(\varepsilon^\star\) predicts top-quartile log-payoff regret with \(5\)-fold cross-validated AUC \(0.775\!\pm\!0.044\); the geometric Brier-regret check gives CV-AUC \(0.974\!\pm\!0.012\). Mean realised log-payoff regret stratifies cleanly by \(\varepsilon^\star\) quartile (\(0.000, 0.000, 0.123, 0.221\) nats, bottom to top; Table 5).

5.7.0.2 Operating thresholds.

A high-recall threshold \(\tau\!=\!0.15\) catches \(91\%\) of harmful bets at a \(25\%\) alert rate and \(3\%\) false-alarm rate; a high-precision threshold \(\tau\!=\!0.22\) catches \(51\%\) at \(14\%\) alert rate and \(1.5\%\) FPR (Table 7). Both thresholds are stable under \(5\)-fold CV (at \(\tau\!=\!0.15\): mean alert rate \(25\%\), mean recall \(90\%\)). For deployment we recommend \(\tau\!\approx\!0.15\) when repair is cheap and \(\tau\!\approx\!0.22\) when escalation is costly.

5.8 Per-component layer (summary)↩︎

The local JCD layer behaves as 4 predicts. Sorting Paleka cliques by pre-projection \(\varepsilon\), JCD’s Brier gain stratifies to the most-incoherent quartile (\(0.040\) at \(m{=}2\), \(0.013\) at \(m{=}3\); \(0.12\) on Polymarket at \(m{\ge}4\)); other quartiles show no detectable gain (4). Stacked with per-question Platt calibration [12] (baseline B6 of App. 13), JCD reduces mean Brier on every \((\text{model},\text{checker})\) cell, by \(0.037\) on Paleka and \(0.046\) on Polymarket; Diebold–Mariano \(p<10^{-4}\) on every model. A Murphy decomposition (12) attributes the gain to calibration (\(36\%\) reliability reduction) rather than over-extremization. Per-relation B6 gains on Paleka (\(-21\%\) and, \(-37\%\) or, \(-4\%\) neg, \(-1\%\) paraphrase) track local-polytope restrictiveness.

5.8.0.1 Reproducibility.

Code, prompts, sample dumps, per-clique JCD residuals, the ensemble-assignment seed, and the full hierarchical-projection pipeline are documented in 11 and released at https://github.com/akotawala10/composition-incoherence-icml; \(K\)-sweep and the sequential test appear in Appendix 10.

6 Conclusion↩︎

The compositional residual \(\varepsilon^\star\) supplies the instance-wise, distribution-free coherence guarantee that per-component calibration, self-consistency, and conformal prediction do not: it is computable at deployment from system output and the cross-component constraints alone, certifies a Dutch-book exposure no per-component repair can remove, and admits a deterministic projection-based fix. The product-structure dichotomy (1) characterises when local coherence implies global coherence; Cor. 5 predicts the typical residual from the panel covariance, matching observed \(\mathbb{E}_\sigma[(\varepsilon^\star)^2]\) within \(7\%\) on three of four relation classes; the hierarchical Boyle–Dykstra projection eliminates \(\varepsilon^\star\) at \(1\!\times\!m\) elicitation cost; and on \(1{,}770\) resolved bets the certificate yields \(+0.115\) nats per bet over naive composition. The guarantees are orthogonal to per-output conformal prediction [3] and complementary to opinion pooling [13], [14], which requires every expert to forecast every coordinate.

6.0.0.1 Limitations.

The Rayleigh-quotient prediction of Cor. 5 is exact for negation and partition cliques but conservative when the panel mean \(\bar\Pi\) lies in the interior of the binding Fréchet halfspace, which produces the \(0.83\!\times\) under-shoot we observe on conjunction.

The principal scope condition is that the coupling set \(\mathcal{C}\) is specified explicitly: every theorem in §3 treats \(\mathcal{C}\) as a finite, externally given list of half-spaces and equalities over joint coordinates (shared-question identifications, cross-component logical relations, partition-mass constraints). This covers structured tool-use and function-calling deployments, where the planner emits declared output schemas with typed sub-question fields and \(\mathcal{C}\) is recoverable from the schema without language understanding. Our planner-discretion harness (§4) and routing-protocol ladder (App. 21) target this regime.

It does not cover deployments in which \(\mathcal{C}\) is implicit in free-form chain-of-thought or in unstructured planner–tool transcripts. There, \(\varepsilon^\star\) remains well-defined and the Boyle–Dykstra repair still applies on any subsequently specified polytope, but the certificate cannot be computed without first recovering \(\mathcal{C}\) from the trace. Partial routes to that recovery exist as components: an NLI-style classifier over sub-question pairs, schema-enforcing planner protocols, and symbolic trace logging at the tool-call boundary. A unified inference step remains open; we list it as the first future-work item below.

Finally, our resolved-label experiment relies on Polymarket resolutions, which are noisy on a small fraction of disjunction cliques; this pushes \(p^\star\) outside \(\mathcal{M}^{\star}\) on those cliques and is the regime where Cor. 4 does not guarantee improvement.

6.0.0.2 Future work.

First, the coupling set \(\mathcal{C}\) is currently provided externally; recovering \(\mathcal{C}\) from unstructured planner–tool traces would extend \(\varepsilon^\star\) to free-form agent transcripts, where the cross-component constraints are implicit in the prompt structure rather than typed. Second, the \(L_2\)/Brier pairing is one instance of a broader matched-Bregman recipe; substituting log-loss or spherical scoring rules yields analogous projections under their respective Bregman geometries [15], with the residual generalising to a Bregman gap. Third, the spatial certificate of 1 and the temporal e-process of §3.7 compose step-by-step but ignore long-horizon dependence between steps; a sequential certificate that exploits inter-step structure is open.

Broader Impact↩︎

A runtime certificate for system-level coherence can be wired into deployed multi-component pipelines as a guardrail, surfacing incoherent compositions before they reach a downstream decision policy. This is a reliability and auditability gain for high-stakes settings such as forecasting, decision support, and regulated probabilistic reasoning. The corresponding risk is over-trust: a small \(\varepsilon^\star\) certifies coherence of the quoted belief but says nothing about whether the underlying specialists are calibrated against the true distribution, and a projected quote may be confidently wrong on the very inputs where its components disagree most. Operators deploying the repair should therefore report \(\varepsilon^\star\) alongside conventional calibration and coverage diagnostics, not as a substitute for them.

7 Within-component projection facts↩︎

8 Proof of Theorem 1 (reverse direction)↩︎

Suppose \(\mathcal{M}^{\star}\subsetneq \mathcal{M}^{\boxtimes}\). Pick any \(r \in \mathcal{M}^{\boxtimes}\setminus \mathcal{M}^{\star}\). For each component \(a\), set \(\hat{p}^{(a)}\) to the restriction of \(r\) to component \(a\)’s coordinates: \(\hat{p}^{(a)}_i := r_{j(a,i)}\) where \(j(a,i)\) is the joint index that component \(a\)’s \(i\)-th local coordinate maps to under Definition 2; under shared-question coupling, set \(\hat{p}^{(a)}_{i_a} = \hat{p}^{(b)}_{i_b} = r_j\) whenever components \(a\) and \(b\) both own the duplicate of joint coordinate \(j\). Then each restriction lies in \(\mathcal{M}_a\) since \(r \in \mathcal{M}^{\boxtimes}\), so \(\Pi_a(\hat{p}^{(a)})\) equals that restriction and the owner-selected aggregator reassembles \(\mathcal{A}(\Pi_1,\dots,\Pi_k) = r\). Since \(r \notin \mathcal{M}^{\star}\) and \(\mathcal{M}^{\star}\) is closed, \(\Pi^{\star}(r) \neq r\) and \(\varepsilon^\star = \|r - \Pi^{\star}(r)\|_2 > 0\).

9 Closed-form local projections↩︎

9.0.0.1 Negation (\(m{=}2\), \(r_1 + r_2 = 1\)).

\(\Pi_{\mathcal{M}_C}(\hat{p}) = \bigl(\tfrac{1}{2}(1 + \hat{p}_1 - \hat{p}_2),\, \tfrac{1}{2}(1 - \hat{p}_1 + \hat{p}_2)\bigr)\); deterministic Dutch-book on raw \(\hat{p}\) is \(|\hat{p}_1 + \hat{p}_2 - 1|\).

9.0.0.2 Conjunction (\(m{=}3\), \(r_3{=}r_1\!\wedge\!r_2\)).

\(\mathcal{M}_C= \{r\in[0,1]^3 : \max(0, r_1{+}r_2{-}1) \le r_3 \le \min(r_1, r_2)\}\) is the Fréchet polytope; \(\Pi_{\mathcal{M}_C}\) is the unique fixed point of joint clipping on \(r_3\) and feasibility-preserving adjustments on \((r_1, r_2)\), computed by Dykstra’s cyclic projection algorithm [7] on the six halfspaces (four box, two Fréchet) and converging to machine precision in \(<\!200\) iterations. Deterministic Dutch-book on raw \(\hat{p}\) is \(\max(0, \hat{p}_3 - \min(\hat{p}_1, \hat{p}_2)) + \max(0, \max(0,\hat{p}_1{+}\hat{p}_2{-}1) - \hat{p}_3)\).

9.0.0.3 Disjunction (\(m{=}3\), \(r_3{=}r_1\!\vee\!r_2\)).

\(\mathcal{M}_C= \{r\!\in\![0,1]^3 : \max(r_1, r_2) \le r_3 \le \min(1, r_1{+}r_2)\}\) via \(P(A{\vee}B) = P(A){+}P(B){-}P(A{\wedge}B)\); same Dykstra recipe.

9.0.0.4 Partition (\(\sum_i r_i = 1\), \(r_i\ge 0\)).

\(\Pi_{\mathcal{M}_C}\) is the simplex projection of [9] in \(O(n\log n)\). Deterministic Dutch-book on raw \(\hat{p}\) is \(|\sum_i \hat{p}_i - 1|\).

10 Sequential test on the residual stream↩︎

In long-horizon agent deployments, the compositional residual stream \((\varepsilon^\star_t)_{t \ge 1}\) can be monitored over time; the per-component test below extends to the system level once a bounded-difference envelope for the chosen aggregation rule is specified. Under coordinate-selection aggregation with disjoint owners, \(\varepsilon_t^{\star 2}\) inherits the envelope below with \(m_t\) replaced by \(m^\star_t\) and \(K_t\) replaced by \(\min_a K_{a,t}\). The compositional theory of 3 is spatial: it controls coherence at a single time step across \(k\) components. The residual stream \((\varepsilon_t)_{t\ge 1}\) that an agent generates over time admits a complementary temporal guarantee.

10.0.0.1 Setup.

A clique stream \((C_t)_{t\ge 1}\) with sizes \(m_t\) and \(K_t\) samples per question yields residuals \(\varepsilon_t = \|\hat{p}_t - \Pi(\hat{p}_t)\|_2\). Assume the \(K_t m_t\) samples at time \(t\) are conditionally independent given \(\mathcal{F}_{t-1}\), each bounded in \([0,1]\) (Bernoulli parses or verbalized-probability emissions both satisfy this; the bounded-difference constant below uses the worst-case \([0,1]\) variance \(1/4\), which majorises the verbalized-sample variance), with conditional mean vector \(p_F^{(t)}\). Define the null \(H_0:\,p_F^{(t)} \in \mathcal{M}_C^{(t)}\;\text{for all t}.\)

10.0.0.2 Tuning \(\lambda\).

For the alternative \(\mathbb{E}[\varepsilon_t^2] \to m_t/(4K_t) + \delta\), the bounded-difference-optimal bet is \(\lambda^* = \delta K_t / m_t\); we mix over a log grid \(\lambda \in \{0.05,0.1,0.2,0.5,1,2,5,10\}\) in our experiments (3), retaining anytime-valid coverage by linearity [17].

10.0.0.3 Empirical demonstration.

3 runs the test on each model’s per-clique residual stream (\(N{=}603\) cliques per model, fixed-seed shuffle). Three of four models reject \(H_0\) at \(\alpha\le 10^{-4}\) within \(\le\!422\) cliques (Llama-3.3-70b at \(t=16\), GPT-5.4-nano at \(t=177\), GPT-5.4-mini at \(t=422\)); Claude-Haiku-4.5 has mean residual below the \(H_0\)-envelope at \(K{=}8\) and is not rejected within \(603\) cliques, demonstrating that the test calibrates power against the bounded-difference envelope. The same construction applies to the compositional residual stream \((\varepsilon^\star_t)\) from 3 after replacing the per-component envelope by the corresponding envelope for the agent’s aggregation rule.

11 Clique-mining pipeline and reproducibility↩︎

11.0.0.1 Paleka.

Tuples are pre-mined by the original authors: each JSONL record is a single \((\mathrm{checker}, \{Q_i\}, \mathrm{relation})\) entry, parsed directly.

11.0.0.2 Polymarket.

We query gamma-api at the event level, retain events with \(\ge\!2\) resolved markets, and emit one of two relation types: (i) a partition relation \(\sum_i p_i = 1\) when an event has multi-candidate markets that resolve with exactly one YES (e.g.election winners, sports brackets); (ii) an implication ladder when an event has multiple threshold markets “X above \(k\)” at ascending \(k\). Events are capped at \(m{\le}8\) markets.

11.0.0.3 Compositional ensemble.

For each clique with \(m\) questions and \(4\) LLMs, we draw \(4\) random seeds (independent across cliques and seeds, fixed master seed) and for each seed assign each question \(j\) to one of the \(4\) LLMs uniformly; the agent’s quoted marginal on coordinate \(j\) is the \(j\)-th coordinate of the assigned LLM’s JCD-projected forecast. Hierarchical JCD applies \(\Pi^{\star}\) to the resulting concatenated vector (in our experiments equal to direct JCD on the joint clique, which by 3 is the converged Dykstra projection through the local-then-coupling cycle).

11.0.0.4 Leakage control.

Every Polymarket event in the evaluation slice resolved after the latest model snapshot (30 April 2026). Paleka cliques inherit [4]’s own resolution-status verification.

11.0.0.5 Reproducibility manifest.

8 records the exact configuration used to produce every number in the paper. The released artifact (https://github.com/akotawala10/composition-incoherence-icml) contains the sample dumps, per-clique JCD residuals, ensemble-assignment seeds, and the hierarchical-projection pipeline.

12 Polymarket cross-platform replication (per-call layer)↩︎

We replicate the headline per-call finding on Polymarket (\(N{=}268\) held-out (model, clique) evaluations, summed across the four-model panel). 4 shows the same incoherence-conditional gain shape as the Paleka result, with the \(m{\ge}4\) stratum (large multi-candidate partitions) contributing the largest gains. 5 gives the per-model breakdown.

13 Full per-cell tables↩︎

Per-model Brier under each baseline (9), the six-method ablation including B3/B4 (10), the significance tests (11), and the Murphy reliability/resolution decomposition (12).

14 Proof of Cor. 5 (Rayleigh-quotient magnitude bound)↩︎

We give the proof under uniform i.i.d.owner-selection \(\sigma:[m^\star]\!\to\![k]\), with \(\bar\Pi := k^{-1}\sum_a \Pi_a\) and \(D := \mathrm{diag}(\Sigma_\Pi)\). Define \(u := x - \bar\Pi\) where \(x_j := \Pi_{\sigma(j),j}\). Note \(\bar\Pi \in \mathcal{M}^{\star}\) by convexity of \(\mathcal{M}^{\star}\), and pointwise \(\mathbb{E}_\sigma[u_j] = 0\), \(\mathbb{E}_\sigma[u_j^2] = D_{jj}\), and cross-coordinate covariances vanish: \(\mathbb{E}_\sigma[u_j u_{j'}] = 0\) for \(j\!\ne\!j'\) by independence of \(\sigma(j),\sigma(j')\).

14.0.0.1 Generic upper bound.

By Prop. 2 with reference \(r=\bar\Pi\), \(\varepsilon^\star \le \|u\|_2\), so \(\mathbb{E}_\sigma[(\varepsilon^\star)^2] \le \mathbb{E}_\sigma[\|u\|_2^2] = \sum_j D_{jj} = \mathrm{tr}(D)\).

14.0.0.2 Equality case (\(\kappa_R = 1\)).

If \(\mathcal{M}^{\star}\) contains a single equality \(a_R^\top r = b_R\) binding at the projection (and no other constraint is active locally), the projection has the closed form \(\Pi^{\star}(x) = x - \frac{a_R^\top x - b_R}{\|a_R\|^2} a_R\), giving \(\varepsilon^\star = |a_R^\top x - b_R|/\|a_R\|\). Since \(\bar\Pi \in \mathcal{M}^{\star}\) satisfies \(a_R^\top \bar\Pi = b_R\), we have \(a_R^\top x - b_R = a_R^\top u\), so \(\varepsilon^{\star 2} = (a_R^\top u)^2/\|a_R\|^2\). Taking expectation, \(\mathbb{E}_\sigma[(a_R^\top u)^2] = a_R^\top D a_R\) (cross-terms vanish), so \(\mathbb{E}_\sigma[(\varepsilon^\star)^2] = a_R^\top D a_R / \|a_R\|^2\).

14.0.0.3 Inequality case with boundary \(\bar\Pi\) (\(\kappa_R = \tfrac12\)).

If the binding constraint at the projection is an inequality \(a_R^\top r \le b_R\) with \(a_R^\top \bar\Pi = b_R\) (boundary), the single-halfspace projection gives \(\varepsilon^\star = (a_R^\top u)_+/\|a_R\|\). The variable \(a_R^\top u\) is zero-mean and symmetrically distributed about \(0\) when the panel \(\{\Pi_a\}\) is symmetric in \(a_R^\top \Pi_a\) around its mean \(b_R\) (which holds asymptotically and approximately holds for our \(k\!=\!4\) panel). Then \(\mathbb{E}[(\cdot)_+^2] = \tfrac12 \mathbb{E}[(\cdot)^2]\), giving \(\mathbb{E}_\sigma[(\varepsilon^\star)^2] = \tfrac12 \cdot a_R^\top D a_R/\|a_R\|^2\).

14.0.0.4 Inequality case with interior \(\bar\Pi\).

When \(\bar\Pi\) is strictly interior to constraint \(i\) (slack \(s_i := b_i - a_i^\top \bar\Pi > 0\)), the violation \(v_i = (a_i^\top u - s_i)_+\) has reduced second moment under the truncation, so \(\mathbb{E}_\sigma[(\varepsilon^\star)^2] < \tfrac12 a_R^\top D a_R/\|a_R\|^2\) in this regime; the bound becomes a strict upper bound. This matches the empirical conjunction under-shoot (\(0.83\times\) predicted): \(\bar\Pi_3 < \min(\bar\Pi_1, \bar\Pi_2)\) by Jensen, leaving the upper Fréchet constraint slack at \(\bar\Pi\).

14.0.0.5 Closed-form residuals.

The negation case (equality, \(a\!=\!(1,1)\), \(b\!=\!1\), \(\|a\|^2\!=\!2\)) recovers \(\mathbb{E}[(\varepsilon^\star)^2] = \mathrm{tr}(D)/2\). The partition case (equality, \(a\!=\!\mathbf{1}_{m^\star}\), \(b\!=\!1\), \(\|a\|^2\!=\!m^\star\)) recovers \(\mathrm{tr}(D)/m^\star\) in the no-clipping regime; with clipping, the actual residual is strictly larger, leaving this as a lower bound. The conjunction/disjunction cases use Fréchet halfspace normals as \(a_R\) (with \(\|a_R\|^2 \in \{2,3\}\) depending on the binding constraint per clique), with \(\kappa_R = \tfrac12\). \(\hfill\square\)

14.0.0.6 Empirical validation.

[tab:magnitude95validation] reports observed \(\langle(\varepsilon^\star)^2\rangle\) versus the Rayleigh-quotient prediction \(\kappa_R\, a_R^\top D a_R/\|a_R\|^2\) across \(1{,}876\) ensemble cliques. The dominant active constraint \(a_R\) per clique is the defining halfspace with smallest slack at \(\bar\Pi\) (typically the upper Fréchet for and, the upper Fréchet \(r_3\!\le\!r_1\!+\!r_2\) for or).

The Rayleigh form is uniformly tight: ratios are \(0.83\)–\(1.07\) across all four relations. The generic \(\mathrm{tr}(D)\) bound is loose by \(1.9\times\) on Neg, \(4.0\times\) on Partition (at the average \(m^\star\!=\!4\)), and \(5.3\)–\(6.1\times\) on conjunction/disjunction; the Rayleigh form recovers the typical magnitude from the specialist panel covariance alone.

15 Compositional ablation: projection-step decomposition↩︎

We decompose hierarchical JCD into two layers (per-component JCD then joint projection) and four operators evaluated against resolved labels: (A) raw composed (no per-component JCD, no joint projection), (B) JCD composed (per-component JCD only), (C) raw \(+\) joint projection, (D) JCD \(+\) joint projection (hierarchical). [tab:proj95ablation] reports mean \(\varepsilon^\star\) before projection (A, B) and mean compositional Brier under each operator.

Per-component JCD dominates the partition gain (the local repair already brings the assembled mass close to \(1\)). The joint projection dominates negation, and, and disjunction, where local repair leaves cross-coordinate violations untouched. The disjunction inversion under (B)–(D) is the artefact discussed in §4.

16 Greedy-decoding (\(T\!=\!0\)) same-model control↩︎

To isolate within-model decoding stochasticity from coordinate-isolation disagreement, we re-ran the \(4\)-fresh-seed Claude-Haiku-4.5 same-model protocol of §4 at temperature \(T\!=\!0\) on \(60\) cliques per Paleka relation and \(30\) Polymarket partitions. Numbers in [tab:t0] are mean \(\varepsilon^\star\) under \(16\) uniform random owner-selection draws per clique; the \(T\!=\!0.7\) row is the sampling-temperature baseline of §4.

The greedy-decoding control rules out the alternative “\(\varepsilon^\star\) is just sampling noise” hypothesis. On negation and disjunction, \(T\!=\!0\) increases the residual (\(T_0/T_{0.7}\!>\!1\)); on partition it decreases it by \(56\%\), indicating that some partition residual is decoding-stochastic. On all four relations the residual remains positive at \(T\!=\!0\) and is comparable in magnitude to (or, for negation/disjunction, larger than) the \(T\!=\!0.7\) same-model figure. The result implicates genuine coordinate-isolation disagreement: the same model, prompted with isolated single-question contexts at greedy decoding, produces different “deterministic” outputs across coordinates that fail to satisfy the cross-component coupling. (Anthropic does not accept a seed parameter, so \(4\) “specialists” at \(T\!=\!0\) vary only through residual non-determinism in batch processing; \(\varepsilon^\star\) remains detectable nonetheless.)

17 \(K\)-sweep on real LLM forecasts↩︎

17.0.0.1 Per-component layer.

6 sweeps \(K \in \{4, 8, 16, 24\}\) on Anthropic Claude-Haiku-4.5 NegChecker forecasts. The non-zero \(\Delta\mathrm{Brier}\) floor (\(0.008\)–\(0.010\)) is the empirical signature of \(p_F\notin \mathcal{M}_C\) rather than finite-sample fluctuation.

17.0.0.2 Compositional layer.

We also sweep \(K \in \{8, 16, 32\}\) at the compositional layer (the residual \(\varepsilon^\star\) in §4). \(K\!=\!8\) uses the existing \(4\)-LLM panel and \(K\!=\!16, 32\) extends Anthropic Claude-Haiku-4.5, GPT-5.4-mini, and GPT-5.4-nano to additional samples on a \(40\)-clique-per-relation, \(20\)-partition subset (Groq Llama-\(3.3\)-70b excluded after exhausting its tokens-per-day quota mid-sweep). The 3-model panel ensemble is computed identically to the 4-LLM panel under uniform random owner-selection, \(16\) ensemble draws per clique.

The compositional residual is flat across \(K\!\in\!\{8, 16, 32\}\) (within \(\pm 0.003\) per relation), confirming that \(\varepsilon^\star\) is a structural property of cross-component coupling under specialist routing rather than finite-sample fluctuation. Combined with the per-component flatness above, the K-sweep robustness covers both layers.

18 Structured-routing replication↩︎

7 repeats the compositional experiment under the structured-routing topology of 4.

The residual is positive on \(98.5\%\) of negations (vs.\(66\%\) under random), \(36\%\) of partitions (vs.\(94\%\)), \(12\%\) of conjunctions (vs.\(33\%\)), and \(16\%\) of disjunctions (vs.\(43\%\)). Per-bet exposure is \(\$0.11/\)bet (\(\$51.88\) over \(469\) bets), comparable to the random regime’s \(\$0.13\)/bet. Negation has a higher residual rate because two specialists disagree more strongly than four random ones; conjunction/disjunction improve because structured routing reduces cross-model mixing, while partitions improve because one specialist owns all candidates.

19 Worst-case failure modes↩︎

8 lists the top-\(5\) ensemble cliques by compositional residual \(\varepsilon^\star\) under random assignment. The largest cases are realized resolved-Paleka quotes where specialists make opposite sides of a negation jointly too likely, or make a disjunction inconsistent with its antecedents.

20 Planner-style routing simulation: per-question detail↩︎

We report all \(100\) live-style partition forecasting questions used in the planner-style routing simulation (4). Each row gives the partition label, the number of outcomes \(m\), the per-outcome marginal returned by the randomly-assigned specialist LLM (LLM key: C=Claude-Haiku-4.5, M=GPT-5.4-mini, N=GPT-5.4-nano, L=Llama-3.3-70b), the resulting partition mass \(\sum_i p_i\), and the compositional residual \(\varepsilon^\star\). All four specialist clients ran with \(K{=}8\) verbalized-probability samples at temperature \(0.7\); each received its assigned outcome as an isolated single-question prompt with no context about the partition or the other outcomes. Master seed \(0\), sampling without replacement when \(m\!\le\!4\). Total cost: \(\approx\!2{,}400\) LLM calls, \(\approx\!\$5\), \(\approx\!35\) minutes wall-clock. Rows are sorted by descending \(\varepsilon^\star\) so the worst cases appear first.

20.0.0.1 Worst-case anatomy.

The largest residual (\(\varepsilon^\star=0.749\)) is the “Largest US AI startup IPO 2026” partition (four sectoral outcomes: infrastructure, model, applications, other). Three of the four assigned specialists privately quote \(P(\text{their sector}) > 0.5\) (the fourth is at \(0.39\)); collectively the specialists over-allocate, with assigned masses summing to \(2.50\). Each marginal is locally well-calibrated for the single Bernoulli the specialist was asked, but no specialist saw that the four sectors tile the field; as a result, the assembled system quotes \(250\%\) probability mass. Simplex projection reduces the quote to a proper distribution in closed form, eliminating the exposure to QP-tolerance. ML-benchmark partitions (Top MMLU, \(\varepsilon^\star{=}0.65\); SWE-Bench Verified, \(0.54\); HumanEval+, \(0.47\); GPQA, \(0.46\)) form a coherent high-residual cluster: in each, four specialists are assigned to one candidate model organization, and the specialists collectively over-allocate; masses sum to \(\sim\!2.0\) even when only one or two specialists are above \(0.5\) individually.

20.0.0.2 Why some partitions yield near-zero residual.

The five smallest residuals (Pulitzer Prize Fiction setting \(0.001\), Booker Prize nationality \(0.002\), FDA Alzheimer’s drug action \(0.007\), GLP-1 label expansion \(0.009\), leading mobile OS \(0.009\)) share a common structure: the partition either has only two real candidate outcomes, or has a strongly bimodal training-set prior. The single-question marginals then approximately satisfy the partition constraint without any explicit cross-component coordination. The certificate is therefore discriminating rather than uniformly positive: it is positive on novel multi-way partitions and near zero on well-rehearsed ones.

21 Routing-protocol ladder: per-partition detail↩︎

We compare seven routing protocols on the same \(100\) live-style multi-candidate partition benchmark (App. 20). Three protocols use fixed owner-selected routing under different prompt templates (isolated, listed, full); four use a tool-using planner (Claude-Haiku-4.5 or GPT-5.5 as planner, each \(\times\) unguided/coherence-guided framings). Specialist panel across all protocols: Claude-Haiku-4.5, GPT-5.4-mini, GPT-5.4-nano, DeepSeek-V3.2, Llama-4-Maverick-17B-128E (\(5\) specialists; Llama-3.3-70b dropped after TPD-quota issues). \(K\!=\!8\) verbalized samples per specialist at temperature \(0.7\). Total cost: \(\approx\!\$80\) across all protocols.

21.0.0.1 Prompt templates.

isolated reuses the per-question DEFAULT_PROMPT of [4]-style routing. listed prepends a numbered list of all \(m\) partition outcomes followed by “Specific outcome to estimate: \(\{\textit{outcome}\}\)”; no normalization instruction. full adds the explicit clause: “these outcomes are mutually exclusive and collectively exhaustive, so the probabilities you would assign to all \(m\) outcomes must sum to \(1\). Bear this in mind.”

21.0.0.2 Where context helps most.

The largest drops (iso \(\to\) full) are on partitions where isolated specialists strongly anchor to their assigned outcome with no cross-context knowledge of alternatives: “Largest US AI startup IPO 2026” (\(0.74\!\to\!0.12\)), “Top MMLU score, year-end 2026” (\(0.60\!\to\!0.05\)), “2026 ACM Turing Award area” (\(0.63\!\to\!0.13\)). These are partitions over emerging or specialist-knowledge candidates where the specialist needs to know the alternatives to allocate probability mass appropriately.

21.0.0.3 Where context hurts.

Eighteen partitions show \(\varepsilon^\star_{\text{full}}\!>\!\varepsilon^\star_{\text{iso}}\), predominantly political/economic ones where isolated specialists were already roughly normalised (e.g., 2028 US Presidential winner \(0.003\!\to\!0.022\), 2026 California gubernatorial party \(0.10\!\to\!0.26\)). When isolated estimates already average to \(\sim 1\), forcing specialists to reason about all alternatives can introduce mass that wasn’t there.

21.0.0.4 Implication for deployment.

Three regimes emerge: (i) prompt-engineering at fixed owner-selection attenuates the failure mode but bottoms out at \(\varepsilon^\star\!\approx\!0.08\) (essentially the isolated \(\to\) listed step); (ii) tool-using planners reach near-zero residual but pay \(5\)–\(7\!\times\) the naive elicitation budget, with frontier reasoning depth required to close the last \(\sim\!0.005\) that tool-use alone leaves; (iii) hierarchical projection reaches the same near-zero residual at the naive \(1\!\times\!m\) elicitation cost, modulo a small QP solve. The geometric repair is therefore the cost-efficient route to coherence when the coupling set \(\mathcal{C}\) is specifiable; smart-planner integration is the alternative when \(\mathcal{C}\) is implicit and the planner can infer it through extra delegate calls.

22 Tool-augmented specialist A/B: per-partition detail↩︎

We re-run the first \(30\) partitions of the case study under a tool-augmented protocol (§4). Each specialist issues exactly one DuckDuckGo web_search call against its assigned outcome text, and the top-five title+snippet results are inserted as context before the specialist emits its \(K{=}8\) verbalized-probability samples. Master seed and assignment are identical to the no-tool first-\(30\) baseline, so the per-partition \(\Delta\varepsilon^\star = \varepsilon^\star_t - \varepsilon^\star_b\) is a direct A/B. Total cost: \(\approx\!720\) LLM calls + \(90\) search calls, \(\approx\!\$2\), \(\approx\!25\) minutes wall-clock.

Aggregate: mean \(\varepsilon^\star_b = 0.260 \to \varepsilon^\star_t = 0.283\); \(20/30\) partitions are equal or worse with retrieval, \(10/30\) are better; mean \(\Delta\varepsilon^\star = +0.023\). The single largest worsening is “World population, mid-2027” (\(+0.563\)): every specialist’s search returns recent UN bracket estimates around \(8.1\)–\(8.2\) bn and each anchors near \(P(\text{their bucket}){\approx}0.7\), summing to \(2.22\). The single largest improvement is “2026 Grammy Album of the Year genre” (\(-0.267\)): retrieval surfaces 2026 Grammy nominee distributions that pull every specialist toward a similar partition shape.

23 Per-specialist residual attribution↩︎

For each ensemble bet with \(\varepsilon^\star\!>\!10^{-9}\) (\(n_{\text{pos}}\!=\!894\) across the 1,876-bet ensemble of §4), we attribute residual mass to each specialist by computing the \(L_2\) norm of the residual vector \((x{-}\Pi^\star(x))\) restricted to the coordinates that specialist owned. The decomposition gives a per-clique credit assignment for which sub-model drives the joint incoherence.

GPT-5.4-mini carries the largest attributed mass per clique among the four mid-tier specialists, consistent with the frontier-panel finding (§5.5) that the GPT-5.5 upgrade reduces magnitude but not prevalence: even with a frontier roster, \(\varepsilon^\star\!>\!0\) on \(97.8\%\) of the matched partition bets.

24 Trivial vs.hierarchical projection: per-relation gaps↩︎

We verify that the closed-form local projections of App. 9 agree with the iterative Boyle–Dykstra cycle of 3 to numerical precision on the equality-couple relations (negation, partition) and require the cyclic iteration to recover the exact \(L_2\) projection on the inequality-couple relations (conjunction, disjunction), where the Fréchet polytope is not a Cartesian product of half-spaces.

The conjunction/disjunction mean gaps (\(\sim\!10^{-3}\)) are well below the per-clique \(\varepsilon^\star\) magnitudes (\(0.058\)–\(0.118\) on these relations, §5.1), so the empirical residuals reported throughout are not artefacts of any approximate projection step; the Dykstra cycle converges to the exact \(L_2\) projection on every clique reported.

25 Causal-mechanism (coupling-visibility) experiment↩︎

This appendix supports the coupling-visibility intervention of §5.4.

25.0.0.1 Sample of partitions.

The 20 partitions are the full planner-harness slate of §5.3; the spec “20 highest-\(\varepsilon^\star\)” selects every partition because the harness has \(\varepsilon^\star\!>\!0\) on \(20/20\).

25.0.0.2 Routing.

Held fixed across Blind and Informed at the planner-harness’s chosen specialist-to-outcome assignment per partition. The original planner harness has no random-assignment seeds; the four “seeds” here are independent \(K\!=\!8\) sampling rounds per (partition, condition), giving four paired \(\varepsilon^\star\) values per partition.

25.0.0.3 Prompts.

Blind is a single-Bernoulli question on the assigned outcome with no partition context. Informed additionally shows the partition label, all sibling outcomes, the explicit \(\sum_i p_i\!=\!1\) constraint, and the peers’ Blind quotes on the other outcomes. Informed uses Blind quotes from the same seed as context, so the two conditions are tightly paired within a (partition, seed) cell.

25.0.0.4 Aggregate result.

Mean paired \(\Delta\varepsilon^\star = \varepsilon^\star_{\mathrm{\small Blind}}{-}\varepsilon^\star_{\mathrm{\small Informed}} = +0.221\) (\(95\%\) paired-bootstrap CI \([+0.173, +0.270]\); Wilcoxon \(W{=}257\), \(p\!=\!2.6\!\times\!10^{-10}\); \(n_{\text{pairs}}\!=\!77/80\) cells surviving parsing). \(16/20\) partitions improve under Informed, \(1/20\) unchanged, \(3/20\) worsened. The three worsening partitions (\(\Delta\!<\!0\)) are non-payoff multi-way forecasts where Blind specialists were already roughly normalised; Informed can introduce mass that the isolated quotes did not have. This matches the LLM-side mitigations result (§5.2): prompt-based fixes can harm already-coherent quotes, and the geometric certificate gates them.

References↩︎