SkillPaste turns a construction worker's demonstrated task into a robot skill, with evidence attached to every claim.
We tested it on cement mixing in simulation. A human-reviewed demonstration defines the task; a PyTorch policy learns to choose 8 of 32 mixing strokes; a simulated Unitree G1 humanoid executes them in MuJoCo physics with 2.6 mm mean tracking error. Every step, from dataset to checkpoint to robot run, is hash-linked in a Skill Passport, so a result can't silently come from the wrong model.
We also fine-tuned NVIDIA's GR00T N1.7 (4000 steps) on simulated G1 tile-placement data. On 20 held-out seeds it scored 0/20 (PARTIAL); our scripted teacher scored 15/20. We report that failure instead of hiding it.
Our mixing policy reaches 72% of the teacher's improvement, below our own 80% target. Simulation only: no real cement, force sensing or hardware.