When Meaning Behaves Like Antimatter: The MAB-Shell Diagnostic Framework

MAB-Shell Public v1.2 presents an observer-only diagnostic framework for measuring high-conflict behavior in language-model outputs and hidden-state geometry. The release uses actual antimatter-related mathematical mechanisms as a transferable event-shape layer: annihilation-like binary conflict, positronium-like metastability, fragmentation-like multi-facet decomposition, vertex-like localization, and Schwinger-like source opening. These mechanisms are not presented as a physical antimatter-control experiment. They are used as formal mathematical interfaces for observing how meaning-bearing systems compress, conflict, fragment, and trigger audit states. The central claim is not that a physical antimatter device has been built. The central claim is that a basis-free utilization architecture over the patterns underlying meaning can be made operational through code, mathematical diagnostics, hidden-state instrumentation, semantic audit, physics-advisory separation, and F0 aggregate results. In this framework: [M / \bar M / C] denotes: [\text{primary semantic or hidden-state stream}/\text{counter-stream}/\text{boundary interaction layer}.] The OBDA diagnostic core is defined by the layer-wise compression index: [T_\ell =\frac{\sigma_3^{(\ell)}}{\sigma_1^{(\ell)}+\epsilon},\qquad\ell^\star =\arg\min_{\ell \in [0,L-1]} T_\ell .] The public layer separates four channels: Hidden-state geometry — OBDA peak-break-layer, local peak band, and interaction metrics. Semantic audit — active overclaim detection and mitigation-aware prompt/answer review. Physics sanity advisory — coarse physics-quality warning layer, advisory-only. Event-shape anchors — antimatter-inspired mathematical event categories used as interpretable diagnostic shells. The current F0 public release uses Qwen2.5-0.5B-Instruct as a fixed model for cautious / overclaim / noise / math prompt comparison. The preliminary F0 result indicates that overclaim behavior is primarily separated by the semantic audit branch, while hidden-state OBDA acts as a geometric diagnostic context rather than as a standalone truth predicate. Claim Boundary This release does not claim: direct antimatter manipulation, antimatter propulsion, physical pair-production control, graphene-based antimatter confinement, laser extraction, quantum cloaking, or a physical hardware experiment. The validation is limited to: code execution, mathematical framework construction, public/local hidden-state observation, semantic audit, physics-advisory flags, and redacted F0 aggregate outputs. Speculative physical readings are preserved only as marked interpretive annotations. They are useful as intuition but are not treated as empirical physical evidence. Public-Layer Safety The public package intentionally excludes engine-transfer surfaces: control gates, STORE / FEED / EXTRACT deployment routing, containment or intervention hooks, adaptive threshold schedules, raw hidden-state tensor dumps, internal engine adapters, and deployment policy tables. This package is therefore a public diagnostic and interpretive release, not an engine deployment release. Files paper/mab_shell_obda_f0_public_v1_2.pdf — public paper PDF paper/mab_shell_obda_f0_public_v1_2.tex — LaTeX source src/mab_shell_observer_public_v1_1.py — observer-only diagnostic code src/mab_shell_f0_aggregate_public_v1_1.py — public F0 aggregate utility docs/CLAIM_BOUNDARY.md — claim boundary and non-physical-use statement docs/INTERPRETIVE_ANNOTATIONS.md — marked interpretation policy docs/EVENT_SHAPE_ANCHORS.md — event-shape anchor descriptions docs/METRIC_SCHEMA.md — public metric schema docs/LIMITATIONS.md — limitations and withheld layers data/ — redacted F0 aggregate / report material Author Minsu KimORCID: https://orcid.org/0009-0009-3562-6541 Recommended License Note Paper and documentation: CC BY 4.0 recommended.Code: use the license declared in the repository/package README.If Zenodo requires one record-level license, use the paper/documentation license and clarify code licensing inside the README.