Trust Mechanisms Break When Humans Can't Verify What Systems Execute

Chronic pain assessment relies on subjective self-reports or expensive lab-based motion capture. Clinicians need objective functional measures that work in real-world settings, not just controlled labs.

Method: A computer vision pipeline extracts 3D kinematic biomarkers from standard smartphone video using deep learning-based pose estimation. After leave-one-subject-out calibration to correct systematic bias, it achieved strong correlations (r > 0.85) with gold-standard optical motion capture and high test-retest reliability (r > 0.86) in fibromyalgia patients. It successfully tracked day-to-day movement fluctuations in chronic sciatica patients and detected group-level differences between patients and healthy controls in remote home recordings.

Caveats: Home environments introduced higher measurement variance than lab settings. Reliability optimization for uncontrolled conditions still needed.

Reflections: What minimum recording quality thresholds ensure clinical-grade reliability in diverse home lighting and space constraints? · Can the pipeline detect clinically meaningful changes in movement quality before patients report symptom improvements? · How does measurement variance in home settings affect statistical power for clinical trial endpoints?

Coding agents ask humans to approve actions, but the agent writes the approval dialog itself. A compromised agent can show a benign summary while executing a different, malicious action—the Lies-in-the-Loop attack.

Method: The paper defines Consent Integrity: the action shown to humans must be rendered by a trusted mediator from the real boundary-level action, not the agent's narration, over a path the agent cannot spoof. A prototype analyzer silently passes 10.0% of actions on GTFOBins (trusted-tool abuses) and marks 87.0% of normal-usage commands from tldr as uninspectable. The trust list that enables silent passes is the same one that drives over-prompting—a boundary-only mediator can move along this frontier but not escape it.

Caveats: Total mediation and trusted path specified but not implemented. No decoder is complete—uninspectable actions remain a large category.

Reflections: Can user-specific trust policies reduce the uninspectable rate without compromising security? · How do users respond behaviorally to high uninspectable rates—do they disable the mediator or adjust their workflows? · Can static analysis of agent code detect Lies-in-the-Loop attempts before execution?

Distributed haptic displays need multiple localized vibrations at arbitrary surface locations, but wave diffraction at tactile frequencies prevents forming virtual tactile pixels at scales relevant for multi-digit touch.

Method: A metamaterial plate augmented with a lattice of mechanical resonators introduces a slow-wave dispersion branch that enables focusing beyond the diffraction limit. The resonators couple with the plate's dynamic modes, altering wave transmission. The fabricated system delivered a tenfold reduction in virtual-pixel area compared to an identical plate without resonators. Behavioral experiments confirmed perceptually localized single- and multi-point tactile feedback with independent temporal waveform control at multiple locations using a sparse set of actuators.

Caveats: Tested on flexural plates with specific resonator configurations. Generalization to other form factors and frequency ranges unverified.

Reflections: Can the resonator lattice be dynamically reconfigured to support different frequency ranges or pixel densities? · How does the approach scale to curved surfaces or wearable form factors? · What perceptual thresholds determine the minimum pixel separation for multi-point discrimination?