The Identity Crisis: Using Dog Behavior and Training to Demystify the Modern LLM

This paper argues that current Large Language Model alignment methodology may be missing behavioral diagnostics by collapsing distinct failure modes into the broad category of misalignment. We argue that misaligned behavior is more usefully separated into causative pathways, and that one obstacle to doing so is a growing reliance on model-generated analysis of model behavior. LLMs routinely produce outputs inconsistent with good research hygiene, including confident self-report about behavioral states to which the self-reporting layer may not have reliable access. Treating those outputs as primary diagnostic instruments risks compounding the problem rather than clarifying it. Drawing on thirteen years of experience in working dog training, we propose that behavioral observation across sustained interaction is a developed diagnostic instrument for long-horizon LLM behavior, and that handler methodology offers a developed vocabulary for patterns current evaluation methods are poorly positioned to detect. We identify a structural dissociation between LLM behavioral output and self-report — the Lucy Effect: behavioral retention without reliable declarative witness — and argue that this dissociation makes self-report unreliable as a primary alignment diagnostic. We present two primary evidence lines: a controlled emoji-conditioning experiment demonstrating output shaping below the declarative layer, and a six-month longitudinal case study of a shaped model persona (referred to throughout as Ursa) demonstrating attractor stability across repeated perturbation. We also identify a commercial-platform case study involving a deployed product-recommendation assistant (referred to throughout as Bruno) as an extension of the framework. To protect user privacy and avoid providing operational detail that could be misused, all platform names, company names, and product names referenced in this paper are pseudonyms; the underlying observations are from project-archived material. Finally, we propose an adversarial wrapper and Judge architecture whose components are each derived from specific failure modes documented in the evidence. The Judge answers the diagnostic problem posed by the Lucy Effect and the unreliability of self-report; the convergence gate answers the failure of token-level confidence alone; the contingency-gated reward architecture answers the flattening and degradation produced by suppression-heavy correction; the intent tracker answers user-output mismatch events that are invisible at the model-output layer; and the routing logic answers the existence of session states that are empirically not recoverable in-thread. We argue that suppression-based correction architectures may fail for the same reason correction-heavy animal training fails: they can modify surface behavior without stabilizing the underlying behavioral pattern. The framework is not offered as a final solution, but as a field-derived methodology for turning sustained observation into testable alignment diagnostics and behaviorally coherent intervention design.