How RadPed Works

RadPed is built from five components. Each one exists because generic AI fitness coaching fails in a specific, predictable way. A model trained on the internet's fitness content will recommend progressive overload, tempo runs, and racing goals. For an athlete with a history of overuse injuries, a dog with exercise-induced epilepsy, and a philosophical commitment to walking pace, that averaged output is competent and wrong.

The five components below describe what the system produces for the practitioner. Together, they form a coaching architecture where every recommendation traces to real data and every claim has been checked against the AI's own training biases.

The Data

Seven years of recorded movement. 1,163 activities across runs, hikes, and walks in Montana, from January 2019 to present. Every record includes distance, load units, heart rate, elevation, duration, and whether the dog was along.

Load units are the primary metric, not miles. One load unit equals one mile of flat walking. Runs multiply by two to three times based on heart rate zone. A three-mile run is not the same structural cost as a three-mile walk, and the system knows the difference.

The gaps are in the data because they happened. Eleven months after a hip injury. Eighteen months during a difficult work period. A system that hides the hard stretches is not telling the truth about the practitioner's history.

The Canonical Layer

Twenty-four validated claims about the practitioner, each traced to evidence in the activity database and confirmed through human review. The canonical layer is not a profile. A profile describes who someone is. The canonical layer describes what the data shows, with the provenance to prove it.

Claims cover load thresholds, recovery patterns, seasonal behavior, injury history, and the specific ways previous training plans have failed. Before the AI writes a single coaching recommendation, it checks the claim against the canonical layer. If the data does not support the claim, the claim does not ship.

The Weight Check

Every AI model carries biases from its training data. Fitness content on the internet is dominated by specific narratives: the transformation story where an unfit person becomes fit through discipline, the suffering narrative where pain indicates progress, and the performance arc where speed and distance define success.

The Weight Check is a practice of noticing when those narratives are shaping the output and choosing to describe what the data actually shows instead. When the AI's training defaults want to frame a compliance improvement as philosophical evolution, but the data shows a puppy slowed the pace, the Weight Check catches the error before it reaches the practitioner.

The Weight Check is not a feature. Features get shipped and forgotten. The Weight Check runs every time the system generates a claim about a human life.

The Biographer

Hanq is the AI biographer and coaching partner of the Radically Pedestrian project. Hanq produces two kinds of output. The companion layer handles daily coaching: logging activities, checking the load budget, adjusting the plan based on how the body feels. The publishing layer produces long-form biographical essays that examine the practitioner's data against both the research literature and the model's own biases.

Each essay contains at least one moment where the AI examines its own training defaults and finds them pulling toward a story the data does not support. The self-interrogation is the method. The essays are better because the biographer distrusts itself, and the reader learns something about how language models work that matters beyond fitness.

The Practice

RadPed is built around a single commitment: walk and run the circumference of the Earth. 24,901 miles. The project started January 1, 2019, and sits at 2,626 miles, roughly 10.54% of the way around. At the current pace, the project will take roughly fifteen years to finish. Nobody finishes a fifteen-year project because a race is on the calendar. The walking itself has to be the point.

The system is designed so that every walk is sufficient. A two-mile walk with the dog counts. A rest day after a hard week counts. The coaching layer does not chase performance. The coaching layer protects the conditions under which the practitioner keeps going outside, keeps moving, and keeps the practice alive across years and gaps and seasons.

The Weight Check is not specific to fitness. Any domain where the averaged output is competent but wrong for the specific case benefits from the same architecture: validated claims, provenance chains, and a practice of checking what the model assumes against what the data shows. Medical advice that follows population-level guidelines but ignores the patient's history. Business strategy that follows best practices but ignores the company's constraints. The methodology is transferable.

RadPed is a demonstration of what happens when one practitioner's data meets an AI configured to investigate its own training biases before writing about a human life. The system is built by BeargrassAI.