Part 6 ended with one committed prediction outstanding: CheetahRun, a different domain — locomotion, not manipulation. The signal said 1.017: barely calibrated-conservative, the ambiguous zone. We wrote “weak-positive” and walked away. The runs finished today.

The result

Jumpy ($k{=}4$) vs vanilla TD-MPC2 on CheetahRun, 3 seeds per arm, 500k steps, finals ($\geq$400k mean): jumpy 620 vs vanilla 558 — $+11\%$, difference CI95 $[-120, 248]$, not CI-separated (source: exp/tdmpc_glass/mechcheck/p1_cheetah_oos_score.json). A weak positive trend from a weak positive signal — the predicted shape, in a domain the signal had never seen.

The predictor’s closing scoreboard

  • 8/8 ordering-consistent facts at fixed k (3-task screen, 4-fact k2 block, cheetah domain transfer) — every one committed before the corresponding harvest.
  • 0/3 on upward cross-k extrapolation (the k8 block) — the documented iteration-drift confound.
  • Plus the unimodal dose-response: $k{=}4$ optimal on all three Franka tasks.

The honest claim: the calibration ratio of a trained jumpy model predicts, across tasks and at least one domain shift, where temporal abstraction will pay — but only when comparing like-for-like k. Small ns everywhere; this is a validated hypothesis, not a theorem.

What’s next: from predictor to method

A predictor you can compute from a checkpoint suggests a method: calibration-selected temporal grain — train short-budget probes at $k \in {2,4,8}$, compute the signal, commit to the best k, train fully. If short-budget signals predict full-budget ordering, TD-MPC2 + auto-k beats vanilla TD-MPC2 with the k hyperparameter chosen by the model itself — an abstraction lever aimed where the campaign showed the headroom actually lives. The mechanism-check (does the 100k-step signal predict the known 500k ordering?) is queued on the fleet as this posts.