A glass-box creativity engine, the Concept Collider, rests on a single primitive: a concept is not merely a point but a structure held under tension, and pressing it along one of seven fracture types breaks it in a characteristic way. That taxonomy is load-bearing—every profile, every homology score, every negative-space detector is computed from it—yet it had never been audited. We give it the first empirical audit and find it imperfect: redundant, non-orthogonal, incomplete, and in one case a consequence miscast as a mechanism. From the audit we reconstruct rather than discard: we demote that consequence, promote a recurring self-defeating tension, and propose an empirically grounded tree of fractures. But the deeper finding is that no taxonomy is canonical: standard decomposition criteria each return a different basis for the same data. Computational creativity is a relativistic discipline—novelty is relative to a base, value to an observer, and the tension taxonomy to a method—so the reportable object is not a basis but the invariants: a small core of fractures (circularity, contradiction, self-defeating means) that survives every change of decomposition method and, across five model families spanning both Western and Chinese training ecosystems (Anthropic, Meta, OpenAI, Alibaba, Moonshot), every change of vendor—though the agreement weakens once the tensions are judged in Chinese, so the invariance holds within a language more than across it. That partial cross-ecosystem agreement weakens—without eliminating—the worry that such consensus merely records shared training text; we therefore report the invariants as robust descriptors under our measurement, not as proof that concepts possess a mind-independent structure. The contribution is a transferable criterion—an invariant is what stays put when both method and model family vary. --- Changes in this version (v2): This second version incorporates the mid-2026 literature on inter-model agreement: Ding (2026) audits agreement as a confidence signal and finds it a positive but weak predictor of correctness, while Liu (2026) supplies the mechanism — error decorrelation across independently trained models — and names its ceiling as a shared-error floor. Both are used to state the size of the observer-independence problem rather than only its direction. The approach is also situated in the psychometric lineage of van der Wal et al. (JAIR 79, 2024), who bring construct validity and reliability to bear on bias measures in NLP. Reproducibility artefacts are deposited with this version: the anonymised 380×7 fracture-profile matrix with its provenance metadata, the per-judge classifications from five model families spanning Western and Chinese training ecosystems, and the analysis and figure scripts. Concept names, tension texts and prompts are withheld; the matrix carries opaque identifiers, which changes no published value — verified by recomputation — while keeping the knowledge base of the audited system out of the release. One correction: the adjusted Rand index of the emergence test has been recomputed from the source corpus and revised from 0.02 to 0.065, and the silhouette is reported as flat across every number of clusters rather than monotonically rising. The conclusion is unchanged — the taxonomy does not emerge from the tension descriptions — and a language control (ARI = −0.001) is now reported alongside it.
Sebastian Wahl· Zenodo (CERN European Organi...· 0 citations
A glass-box creativity engine, the Concept Collider, rests on a single primitive: a concept is not merely a point but a structure held under tension, and pressing it along one of seven fracture types breaks it in a characteristic way. That taxonomy is load-bearing—every profile, every homology score, every negative-space detector is computed from it—yet it had never been audited. We give it the first empirical audit and find it imperfect: redundant, non-orthogonal, incomplete, and in one case a consequence miscast as a mechanism. From the audit we reconstruct rather than discard: we demote that consequence, promote a recurring self-defeating tension, and propose an empirically grounded tree of fractures. But the deeper finding is that no taxonomy is canonical: standard decomposition criteria each return a different basis for the same data. Computational creativity is a relativistic discipline—novelty is relative to a base, value to an observer, and the tension taxonomy to a method—so the reportable object is not a basis but the invariants: a small core of fractures (circularity, contradiction, self-defeating means) that survives every change of decomposition method and, across five model families spanning both Western and Chinese training ecosystems (Anthropic, Meta, OpenAI, Alibaba, Moonshot), every change of vendor—though the agreement weakens once the tensions are judged in Chinese, so the invariance holds within a language more than across it. That partial cross-ecosystem agreement weakens—without eliminating—the worry that such consensus merely records shared training text; we therefore report the invariants as robust descriptors under our measurement, not as proof that concepts possess a mind-independent structure. The contribution is a transferable criterion—an invariant is what stays put when both method and model family vary. --- Changes in this version (v2): This second version incorporates the mid-2026 literature on inter-model agreement: Ding (2026) audits agreement as a confidence signal and finds it a positive but weak predictor of correctness, while Liu (2026) supplies the mechanism — error decorrelation across independently trained models — and names its ceiling as a shared-error floor. Both are used to state the size of the observer-independence problem rather than only its direction. The approach is also situated in the psychometric lineage of van der Wal et al. (JAIR 79, 2024), who bring construct validity and reliability to bear on bias measures in NLP. Reproducibility artefacts are deposited with this version: the anonymised 380×7 fracture-profile matrix with its provenance metadata, the per-judge classifications from five model families spanning Western and Chinese training ecosystems, and the analysis and figure scripts. Concept names, tension texts and prompts are withheld; the matrix carries opaque identifiers, which changes no published value — verified by recomputation — while keeping the knowledge base of the audited system out of the release. One correction: the adjusted Rand index of the emergence test has been recomputed from the source corpus and revised from 0.02 to 0.065, and the silhouette is reported as flat across every number of clusters rather than monotonically rising. The conclusion is unchanged — the taxonomy does not emerge from the tension descriptions — and a language control (ARI = −0.001) is now reported alongside it.
Sebastian Wahl· Zenodo (CERN European Organi...· 0 citations