Order Is Not Control

A new preprint argues that the structured patterns found by AI alignment and interpretability tools do not constitute genuine control, proposing instead a formal, receiver-gated response law validated across biological organisms, large language models, and stochastic systems [1]. The paper, posted to arXiv on 11 June 2026, draws a sharp distinction between identifying order-inducing objects and achieving control. Control, the authors contend, requires a denominator-indexed operator that maps material state, action, bath, and receiver state to a response displacement, while bounding damage, null outputs, and unnecessary effort [1]. The framework is tested across four panels: mouse anterior lateral motor cortex, C. elegans, and zebrafish provide physical response-operator evidence; LLM panels supply generated-output response laws [1]. Across four material conditions, response vectors in LLMs were predictable at 72.8–73.7% component-sign accuracy, rising to 84.3–84.8% on nonzero components [1]. Held-out observers predicted system-effect and target/oracle families at 93.6% and 91.7% accuracy, respectively [1]. The laws are local: an intervention can be admitted, saturated, sign-changing, leaky, or overdriven depending on medium, bath, receiver state, action port, and comparator [1]. The study explicitly excludes several conclusions. It does not claim deployable pre-generation control, hidden or logit causal sufficiency, biological-to-LLM coordinate identity, or literal thermodynamic quantities [1]. The work instead offers a driven-dissipative response-system account at the mesoscopic level, where drives act through prepared media, baths, and receivers to produce admitted movement, impedance, sinks, or overdrive [1]. Related research has explored control-theoretic interventions in generative models. A separate preprint proposes Latent Activation Linear-Quadratic Regulator (LA-LQR), a reduced-order optimal control framework for text-to-video steering that computes closed-loop feedback to guide activations toward desired feature setpoints while penalizing unnecessary perturbations [3]. Unlike the coarse, non-anticipative interventions common in existing T2V steering, LA-LQR projects activations onto a low-dimensional subspace derived from contrastive prompt pairs and solves a latent LQR problem to obtain timestep- and layer-specific steering signals [3]. The broader control-theory landscape includes fractional-order control, which uses a fractional-order integrator weighting history with a power-law tail, creating a distribution of time constants rather than a single resonance frequency [7]. This approach has shown promise in suppressing chaotic behaviors in systems such as muscular blood vessel models and robotics [7]. The new preprint's emphasis on receiver-gated response laws and local intervention outcomes adds a distinct perspective to these ongoing efforts to formalize what it means to steer complex, learned systems [1].