CDIP P2 White-Box Symplectic Verification on an 8B LLM Perturbation–Response Reconstruction of an Antisymmetric Form and Audit-Grade Null Tests

This paper reports an audit-grade, white-box perturbation–response experiment that operationalizes the CDIP P2 module on a commodity 8B language model: reconstructan antisymmetric response form Ω from controlled internal perturbations and establish that the observed structure is not a baseline-mismatch artifact under a dedicated null. Using Qwen3-8B (HF: Qwen/Qwen3-8B), we inject small residual-stream perturbations at a fixed transformer block and read out a frozen d-dimensional subspace response. We (i)reconstruct Ab(α) and Ωb (α) =1 (Ab − Ab⊤ ); (ii) verify non-zero2Ω and paired singular values; and (iii) perform a surrogate test in which baseline readouts are mismatched across prompts to destroy prompt-consistent coupling. For d = 8, block index 28 (0-based), and αtest = 0.005, we obtain Ω= 8.503 × 10−2 with four paired singular-value pairs and a direction-consistency statistic that significantly exceeds the mismatch-baseline null (rank p = 0.001, N = 999 surrogates). This provides a reproducible, low-cost empirical anchor for the CDIP P2 “perturbation–response” interface.