Failure to Recover an Operational Linear-Response Regime in Small-Scale White-Box P2 Audits of a 1.5B LLM

We report a targeted negative-result study motivated by the P2/P2-Min branch of the Complex-Dynamic Intelligence Protocol (CDIP). The question is narrow: can a small open-source language model be placed into an operationally admissible local linear-response regime under a white-box perturbation protocol that reconstructs an antisymmetric response object $\hat\Omega$? We test Qwen2.5-1.5B-Instruct under a sequence of increasingly conservative protocols, progressing from matched pretrained versus random-initialized controls and null baselines to shorter propagation distance, central differences, fp32 injection/readout bookkeeping, and small-norm exemptions for linearity diagnostics. Across all tested configurations, the linearity gate fails for the main pretrained and random-initialized groups. The scale proxy $\alpha\lVert\hat\Omega\rVert_F$ is not approximately flat and instead increases systematically with $\alpha$, indicating absence of a recoverable operational linear-response window in the tested regime. Same-layer null controls collapse as expected, confirming that the pipeline is not a trivial artifact, but scrambled-token controls remain close to the base condition, implying that the measured quantities are not strongly sensitive to semantic organization under the present protocol. The strongest interpretation is therefore negative and regime-specific: for the tested 1.5B model, layer locations, token locations, and perturbation scales, we do not detect an admissible P2-style linear-response regime. This does not refute CDIP as a normative architecture, but it blocks any downstream promotion of these measurements to P2-style structural certificates in this setting. The audit protocol itself, including matched random-init controls, scrambled-token nulls, same-layer collapse checks, and scale-convergence diagnostics, constitutes a reusable methodological contribution independent of the negative empirical outcome.