Methods

An explicit framework treating accuracy, latency, offline reliability, and recovery UX as the four non-negotiables for an AI assistant operating in a safety-sensitive vehicle environment. Most AI product evaluation focuses on capability ceiling: what the system can do at its best. The Must-Not-Break Contract focuses on capability floor: what the system must never fail to do, regardless of conditions. The two are different product disciplines, and the floor is the one that determines whether users trust the system over time.

A related principle, “offline UX is designed, not generated,” treats the offline state as a deliberately authored experience rather than a fallback. Both principles informed the platform strategy work for the AI experience platform Lucid publicly announced as Lucid Intelligence.

See Lucid Intelligence for context.

A capability model spanning Level 0 (fixed command recognition), Level 1 (natural-language vehicle control), Level 2 (multi-turn conversational and contextual interaction), and Level 3 (proactive multimodal companion behavior). Each level is defined by what the assistant can do, what it can reason about, and what kind of interaction loop it supports, rather than by which model powers it.

The model has been used to map competitor offerings, structure internal strategic conversations, and frame partnership conversations. Originally developed as part of the business case for an in-house in-car AI experience.

See Lucid Intelligence for context.

A benchmarking framework organized around five pillars: functional excellence (intent recognition, task completion, hallucination avoidance), cognitive and emotional design (recovery, transparency, calm under pressure), personalization and personality (consistency, adaptability, style range), connectivity and edge model behavior (offline behavior, latency, regional variation), and ecosystem integration and monetization (third-party services, business model fit).

Three cross-cutting capabilities: intent understanding across multi-sentence queries, voice personalization and adaptability, and session and profile context retention, are evaluated separately across all five pillars. Applied to ten production in-car voice systems plus an internal prototype, yielding a structured comparative view that informed platform strategy and feature prioritization.

See AI voice prototyping for Lucid Assistant for context.

Most concept-evaluation methods produce a verdict at the concept level: this idea worked, that one didn’t. Useful features often get lost when their parent concept underperforms. The feature-performance taxonomy categorizes features along three lenses (use likelihood, standout appeal, irrelevance) to surface specific patterns: breakout features, hidden potential features inside underperforming concepts, high-attention features with low projected use, features with high practical value but low emotional appeal, and features users actively avoid.

The taxonomy lets concept evaluations produce two outputs rather than one: a verdict on the concept, and a separable list of features worth carrying forward independently.

See Navigation strategy (Lucid Maps) for context.

A standardized method for measuring how interface elements on central screens and instrument clusters get obscured by physical components of the cabin, the steering wheel, hand positions on the wheel, and natural variation in driver posture and seating geometry. The method combines structured measurement of obscuration patterns across a representative range of driver geometries with visualization output that designers and engineers can read directly during design review.

Developed for early-program design review on the Midsize platform. Applied across product reviews by a colleague on the research team. Recognized internally as a Breakthrough Innovation.

A protocol for measuring driver distraction during in-cabin secondary tasks, adapted from the Detection Response Test for use on a static lab rig rather than a driving simulator. Participants face a large display showing a pre-recorded first-person driving video, urban traffic, while attempting a secondary task on the vehicle’s screens. A briefly-displayed visual stimulus appears at random intervals at a fixed central location; participants respond by pressing a steering-wheel-mounted button. Reaction time and miss rate are captured from synchronized stimulus and response timestamps. Sessions run a fixed length within a within-subjects design comparing multiple interaction modes against a touchscreen baseline.

The adaptation preserves the rigor of the standard DRT while making it practical for early-stage hardware evaluation. Developed and applied to evaluate Lucid Gravity’s steering wheel capacitive controller across two capacitive button interaction styles. The protocol surfaced a perceived-versus-measured-safety mismatch that informed the final interaction model.

See Gravity multimodal HMI for context.