This preprint reports reproducible architectural patterns in the activation manifold of a self-organizing neural network (GCIS / ZEUS) that localize SHA-256 password preimages within prime-bit-size layers. Across eight independent runs with fresh random initialization, three methodologically distinct diagnostics converge on the same structural reading: the network organizes preimage information geometrically, and the geometry is reproducible across initializations. Key findings: - Architectural template across 8 runs: 174 layer pairs satisfy three structural conditions without exception — prime bit-sizes, layer-index gap of exactly 2 or 10, and password byte distribution symmetric within each pair to within fractions of a percent. - Partial blind byte recovery: A count-based candidate filter with frozen rules recovers 50–53% of true preimage bytes across three runs against the printable-ASCII universe (Fisher combined p = 1.23 × 10⁻⁴, against a random expectation of 33.7%). - Geometric mirror/gauge signal: A position-template correlation under symmetry transforms survives in 8/8 runs against random printable controls, under both permissive and strict layer-selection variants (Fisher combined p ≈ 2.1 × 10⁻⁹). - Bypass rather than break: The findings do not algebraically invert SHA-256, MD5, or ECC, but empirically challenge the operational assumption that cryptographic outputs are structure-blind to neural geometric reading. This threat model is not addressed by the post-quantum cryptography programme.
Stefan Trauth (Tue,) studied this question.