ANT Model and Graph Digestion: Bottleneck-Guided Whole-Object Compression, Reconstruction, and Fortification

This record separates model and graph digestion into a standalone ANT research track. It treats a neural model or graph as a typed causal multigraph rather than an unordered collection of tensors or edges. Fast structural bottleneck detection proposes candidate regions; Ariadne threads combine message flow, calibration ablation, rare-case Sentinel interventions, structural geometry, uncertainty, reuse, and execution cost; atomic multi-site bites preserve connected or hypergraph functional motifs; motif and Queen codebooks support reconstruction; and graph fortification is formulated as a named-shock inverse problem. The package includes a standalone preprint, source code, the supplied measured graph-bottleneck library, five-seed raw and aggregate results, result matrices, figures, publication prompts, and the four prior graph notes that motivate the design constraints. The new controlled experiment uses six repeated graph modules, three common useful routes, two rare Sentinel routes, and one deliberately misleading high-message structural dead end. The structural cascade flags all 11 cross-module edges, demonstrating that detection alone does not solve selection. Ariadne whole-route digestion matches the planted functional oracle in all five seeds: zero median test and Sentinel MSE, all five useful route groups complete, the rare route preserved, and no decoy selected. Attention-only, calibration-only, curvature-only, and local edgewise selectors select the decoy and lose a rare route. Lossless motif-codebook reconstruction recovers the exact 155-edge graph and degree sequence at an explicitly accounted 6.04× structural description reduction. After attacking the rare route, named-shock Ariadne repair achieves median test MSE 4.57e-7, compared with 1.93e-5 for a curvature proxy and 1.47e-4 for Ghosh–Boyd. Ghosh–Boyd nevertheless produces the largest algebraic connectivity, showing that topology-only robustness and named functional recovery are different objectives. The evidence is synthetic and deliberately scoped. The record does not claim superiority on arbitrary real graphs, a directed-curvature guarantee, successful compression of a medium or large language model, hardware acceleration, or automatic Taleb-style antifragility. The included staged programme proceeds from planted graph truth to small graph/model co-digestion, a medium open model with teacher-unloaded `.antcolony` execution, and only then a renewed 3B-model experiment. «Не сжимать отдельные числа. Обнаруживать, поедать, восстанавливать и укреплять целостные функциональные организмы» «Не награждай модель за ускорение, достигнутое ценой ошибки.» «Не награждай представление за малый размер, если оно перестало быть teacher.» S.V.E. Meta-License v4.0 [CC BY-NC-SA 4.0 + Symmetric Addendum]