liftOS

It began with only the legs. Before attempting a whole body or a convincing movement, I built a deliberately narrow lower-body model: quads, glutes, hamstrings, pelvis, femurs, knees, tibias and fibulas. The first question was not whether it could look finished. It was whether a small, real anatomical system could be extracted, rendered, inspected and improved without pretending that a full-body engine already existed. The work ran through a manual builder-and-judge loop: Claude produced a bounded candidate; Codex independently tried to disprove it against the fixed visual and technical criteria. That separation mattered. It made the first work legible as an experiment, not a polished animation dressed up as a system.

The first motion spike asked a narrow question: could the reconstructed body perform a recognizable squat without using a stock animation? We built a 3.2-second form study from the model’s measured skeletal landmarks. An analytic solver calculated the hips, knees and ankles frame by frame, kept the feet planted, and calibrated the movement against published joint-angle ranges. Making the skeleton move was only half the problem: muscles and tendons still had to deform convincingly around it.

The squat, four iterations. The same movement went through four judged versions, each exposing a different problem: a first working prototype, a first skinning repair, a bone-surface and Achilles correction, and a final historical baseline. The progression is useful precisely because it is not a victory lap. The early versions could meet selected mechanical checks and still look wrong once the whole body moved. The last version is the strongest of this first set, not a claim of clinical or anatomical certification.

The proportion explorer makes one relationship visible: femur length versus lower-leg length. These are not the same squat animation stretched onto three bodies. Change the hip-to-knee and knee-to-ankle segments, and LiftOS re-solves the movement to keep the body’s centre of mass over the midfoot. In this setup, a relatively long lower leg lets the torso remain more upright; a relatively long femur calls for more forward lean. The body’s geometry is producing the movement, not decorating it afterward.

The Movement Combine was a one-night experiment in making my AI tools compete. Claude, Codex, and Cursor (running Grok) each got the same brief: teach the 3D anatomical figure in LiftOS five classic lifts — bench press, deadlift, overhead press, chin-up, and bent-over row — working alone in parallel copies of the project, judged against the app's existing squat. It was, essentially, a failed experiment. Overnight I burned through more than half my weekly limit on all three tools, and each one confidently handed in its assignment, certain it had satisfied the requirements. The truth couldn't have been further from that: skeletons hunched over barbells they never actually grip, bars floating through shins or missing altogether. The revealing part wasn't the janky animation, though — it was the confidence. Three state-of-the-art models graded their own work as done, and not one could see what any human sees within two seconds of pressing play.