Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Compositionality — how the structure and meaning of a whole come from its parts and how they are combined — the principle of compositionality and compositional semantics, the emergence of compositional structure in languages under transmission, and composition and binding in learned models (generalization to unseen combinations, attribute binding, composing distributions or concepts). A work on making a model generalize compositionally is ML practice and goes to the anthology (ADR-031).
14 of 547 CLAIM documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| CLAIM-001 | The conditional-diffusion literature already covers directed transport between observational scenarios, including changing covers | The owner at U34, quoting A93’s research opportunity: “I think conditional diffusion literature has us covered here.” Accepted for compositional conditioning and transport between distributions; never shown for changing covers, and limited by A97 itself. | Proposed |
| CLAIM-007 | In a bridge or chain of progressive reconstruction the path itself is observable: which constraints survive early, which are introduced late, and which relations emerge only through refinement | A97 §3 on Schrödinger bridges, recovered. The manuscript keeps the bridge only as “a complementary path-space viewpoint” (§7); the programme of observing the path was dropped without critique. | Proposed |
| CLAIM-013 v2 | Existing text-to-image benchmarks evaluate object presence, attributes and whether spatial or interactive relations are depicted, not the stance, social alignment or uptake an image conveys, and need extending to pragmatically consequential relations and social uptake | The manuscript’s §10, Case III, which calls T2I-CompBench and GenEval object-centered baselines. CLAIM-097 qualifies that for T2I-CompBench. | Proposed |
| CLAIM-048 | Communication transports structures of observational constraints rather than invariant semantic objects, and translation, language-model reconstruction and conditional image generation are realizations of one process | Proposal v6 (A100, under U35), Contextual Information Transport Across Language, Images, and Communicative Frames. It made conditional generation “part of the central theory, not merely an application at the end”. Revised at A102 with structuralism. | Superseded — by CLAIM-036 |
| CLAIM-066 v2 | Kirby, Cornish and Smith show structure emerging under transmission with an experimenter’s filter, not filtering by measured inductive biases, and Griffiths and Kalish’s convergence to the prior holds only for sampling learners | From the readings of Griffiths and Kalish (NOTE-599) and Kirby, Cornish and Smith (NOTE-596, Skimmed). The manuscript cites them as showing one thing; they show different things. | Superseded — by ARG-035 |
| CLAIM-071 | Conditional generation shares a mathematical vocabulary with translation, not a mechanism: denoising time is not telephone-game time, and a generative failure is not a proof of sheaf contextuality | A97’s own two limits, which the manuscript §7 keeps. They became the whole of the claim once generation was demoted from realization to analogue. | Active |
| CLAIM-076 v2 | Cross-modal chains expose reconstruction because media differ in expressive affordances, and each interpreter, such as a captioner, may add motives and relations that were not present | A110 §21, recovered. The manuscript’s Case III keeps the protocol and drops the reason for it; C7 Appendix D keeps a related control (captioners blind to the source). | Proposed |
| CLAIM-087 v3 | Cross-modal reconstruction often improves a receiver’s access to an implicit communicative relation beyond the text-only baseline, as when posture or expression conveys solidarity better than the source text’s literal content, though a rendering made without the situation is a garbling and cannot improve what an ideal receiver could do | A97 §6 and A100 §11.10, recovered. The manuscript treats a change of medium only as a risk to relational fidelity (§7), though §11 allows that an adapted receiver may recover an implicit relation. | Proposed |
| CLAIM-090 v3 | Serial transmission is reconstructive and filtered by the transmitters’ inductive biases, so a stable endpoint can be far from the ancestral utterance while remaining interpretable | The manuscript’s §8, citing Bartlett, Griffiths and Kalish, and Kirby, Cornish and Smith together as showing the same thing. CLAIM-066 separates what each shows. | Proposed |
| CLAIM-097 v2 | T2I-CompBench is not purely object-centered: its non-spatial relation category already probes interactions, but scores them with CLIPScore and near human ceiling | From the reading of T2I-CompBench (NOTE-592). The benchmark’s interaction category is the nearest existing measurement of what Case III proposes, and its weaknesses are the manuscript’s argument. | Active |
| CLAIM-120 | Conditional generative systems realize stochastic transports between constrained distributions, and their compositional fidelity depends on preserving the compatibility and relational structure of constraints across media | A97’s “strongest new proposition”, made central by proposal v6 (A100). The manuscript’s Abstract calls conditional diffusion and Schrödinger bridges “generative analogues”, and §7 claims only a shared vocabulary; no turn up to U37 argues for the demotion. | Superseded — by CLAIM-071 |
| CLAIM-123 v6 | In conditional generation, adding scores composes two conditions exactly when the ratio p(y₁,y₂|x)/(p(y₁|x)p(y₂|x)) does not vary with the noisy state x, which conditional independence of the conditions at that state is sufficient for but does not require | The manuscript’s §7, citing Composable Diffusion, T2I-CompBench and Schrödinger bridges. The readings support it, and the manuscript is more careful than the paper it cites. | Active |
| CLAIM-321 v4 | In text-to-image diffusion models each condition of a prompt can be met while their conjunction or relational binding fails: benchmarks find single objects rendered almost always and spatial relations and attribute binding failing most, and composing by adding scores does worst on relations | Split from CLAIM-123 under ADR-043: the empirical part of that claim, that each condition can be met while their conjunction or relational binding fails, documented by GenEval, T2I-CompBench and T2I-CompBench++, by Attend-and-Excite’s “catastrophic neglect”, and by Composable Diffusion losing to an energy-based baseline on relations. It corroborates CLAIM-123’s derivation without carrying it, and does not say what causes the failures. | Active |
| CLAIM-480 v2 | Composing interactions can turn an interior into a constraint, and sometimes into a constraint of a new kind: in the chain E = ½k₁(x₁ − x₂)² + ½k₂(x₂ − x₃)², with x₂ internal and equilibrated, the boundary sees E_eff = ½·k₁k₂/(k₁ + k₂)·(x₁ − x₃)², a coupling between x₁ and x₃ that neither part has; if the interior relaxes slowly, eliminating it leaves a history-dependent constraint, a memory kernel, so interconnection fixes an effective constraint on the boundary’s trajectories and separation of time scales decides only whether that constraint is approximately memoryless; and for cellular sheaves the reduced constraint can relate three boundary values where every relation among the parts is pairwise | From A22 §4 and A23 §IV of the exchange of 2026-10-27, which called composition “the most important missing mechanism” after the owner’s U20. Eliminating an equilibrated interior leaves an effective boundary relation that no part has; an outside system interacting only with the boundary meets it as a constraint, and “The interior has effectively become apparatus.” Baez and Fong’s black-boxing functor is the general theorem for passive linear networks. The memory case ties it to Mori–Zwanzig and to CLAIM-478: interconnection yields an effective constraint on the boundary’s trajectories in every case, and separation of time scales decides only whether it is approximately memoryless. For cellular sheaves the reduction can need a richer relational vocabulary than the parts used (THEORY-1090; QUESTION-283). It is independent evidence for CLAIM-477, that constraints are realized by organization. | Active |