A, B, C or D? I trained all four

Keep the encoder, the 50-step history, the data and the training budget fixed, and change only the objective. A reconstructs the present sensor readings. B predicts the residual between the real state 10 policy steps (0.1 s) ahead and the nominal model’s rollout over the same goals. C separates runs contrastively. D only minimizes latent size. Each rewards a different kind of information, and only B is asked what happens next. The horizon matters: over a single 10 ms step, the residual is mostly encoder quantization.

Freeze each encoder and fit a linear probe from its latent z to the ratio of payload mass to torque constant, m/kₜ, on held-out arms. The target is a ratio because the motor currents reveal payload mass only relative to the torque constant. B passes when the probe’s R² is above 0.9, scored per history window as the median over the same three or more training seeds used for every objective. Then sort cups: three identical-looking cups hold 50, 100 and 150 g, the arm wiggles each for 0.5 s, the probe reads the latent, and the order must be right in 9 of 10 sim trials.

The table holds A, B, C and D and every seed, whatever they show. A high B score would support the course’s answer for this data; a strong A or C deserves an investigation, not a footnote. Keep the held-out split, the probe’s prediction for every window, and a check that the forward model uses z at all, such as shuffling latents between runs. This written companion defines the experiment and claims no training results.