RESEARCH ═════════════════════════════════════════════════ interface-level alignment THE PROBLEM ───────────────────────────────────────────────── AI alignment research focuses on model behavior — making models that do what we intend. But there's a layer above that gets less attention: the interface where human intent meets AI understanding. Current interfaces lose information at every step: • context dies at session boundaries • corrections don't persist • models assume typical users, miss individual patterns • intent gets flattened to text If scalable oversight requires humans to reliably communicate intent, the interface is part of the alignment problem. THE THESIS ───────────────────────────────────────────────── Interface-level alignment is a distinct research surface. Model alignment asks: does the AI do what we want? Interface alignment asks: can we even express what we want? The gap between intent and expression is where misalignment begins. Fixing models won't fix lossy communication. APPROACH ───────────────────────────────────────────────── I'm building systems that treat the interface as a protocol design problem: • persistent context that survives session boundaries • personal dialect modeling — how you communicate, not just what you say • bidirectional adaptation — both human and AI learn each other's patterns • intent classification at capture time WHY ADHD ───────────────────────────────────────────────── Neurodivergent users show interface failures in high relief. Executive dysfunction isn't lack of ability — it's a lossy interface between intent and action. Building for edge cases often reveals what everyone needs. ACTIVE PROJECTS ───────────────────────────────────────────────── ARX knowledge graph for persistent context PID intent classification daemon MAAP protocol for portable dialect OPEN QUESTIONS ───────────────────────────────────────────────── • how do we measure intent preservation? • what's the minimal representation of dialect? • can context be portable across agents/vendors? • what would mutual adaptation look like at scale? ───────────────────────────────────────────────── looking for research collaborators robb[at]arclabs.ai