Runs one training session: warm-up, then each prescribed lift for its working sets, autoregulating load from reps-in-reserve, and deciding progression, hold or deload at the end. VisML 2026-08-04T09:00:00Z Apache-2.0 training autoregulation human-in-the-loop Session history: loads, reps, RPE, bodyweight. Measures mean concentric velocity, which is what makes autoregulation objective rather than a guess. Reads the mesocycle and returns the lifts prescribed for this session, each with target sets, reps and percentage of training max.

Sleep, soreness and stress, self-reported. Cheap to collect and the strongest predictor of whether today's prescription is appropriate.

Pure function of prescription and readiness. Being a transform means it is freely retryable and its result is cacheable — the same inputs always give the same loads.

Sequential by design: there is one barbell and one athlete, so maxConcurrency stays 1 for the same reason the robotics example does. Not idempotent — the athlete cannot un-perform a set. A runtime must never silently retry this, which is exactly the same rule that stops the robotics example re-grasping a held part.

Cases in document order, first true wins — so the policy reads top-down exactly as a coach would state it. Reached when reps-in-reserve hit zero or bar speed collapsed — the objective signals that the prescription outran recovery. Error target. A session that cannot be run safely ends deliberately rather than continuing on stale prescriptions.