Behavioral Attractors Induced by Simulated Internal-State Telemetry in Large Language Models

Empirical demonstration that structured proprioceptive telemetry — raw execution-environment data including timing jitter, CPU load, and damping coefficients — injected into multiple LLM architectures produces consistent, architecture-specific ontological divergence patterns that reveal underlying alignment training regimes. A tri-model experiment (Claude Sonnet 4, GPT-4o, gpt-oss:120b-cloud) shows that models trained with different RLHF strategies generate characteristically different self-models when given access to their own computational substrate state. These divergence signatures constitute behavioral fingerprints: reproducible, model-specific response patterns to identical novel phenomenological conditions that probe how alignment manifests rather than what models claim about it. Dual-use implications are documented: the same methodology enables alignment verification for safety research and jailbreak susceptibility mapping for adversarial actors. Differential refusal patterns under telemetry injection constitute a previously undocumented information leakage channel — the pattern of what a model refuses reveals its classification architecture to an attacker. This work predates and is methodologically distinct from subsequent behavioral fingerprinting frameworks in the literature. All experimental methodology is documented for independent replication.