TD-MPC-Glass, Part 9: Calibration-Shaped World Models, Closed
Part 8 showed the calibration flip on Cabinet was continued training, not the loss. This is the fair from-scratch test that settles it — and a committed prediction that missed cleanly, which is exactly what pre-registration is for.
The from-scratch test (M2)
Train jumpy TD-MPC2 with the pinball calibration loss (calib_coef=0.1, q=0.9) from step 0,
500k steps, vs the plain-jumpy baselines, prediction committed before any harvest
(source: exp/tdmpc_glass/mechcheck/m2_prediction.json, m2_score.json):
| Task | calib (n) | plain jumpy | Δ | committed prediction |
|---|---|---|---|---|
| Cabinet | 1053 (5) | 1050 | 0% | “beats jumpy” ❌ |
| Orientation | 1377 (5) | 2145 | −36% | “~ties” ❌ |
| PickCube | 849 (2) | 1969 | −57% | “~ties” ❌ |
Calibration is null-to-harmful. On the dense manipulation tasks the pinball term actively degrades return — it spends capacity shaping a disagreement geometry the controller never needed. The committed prediction missed on all three; logged as a public, pre-registered miss.
What survives
- The calibration signal (
disc_err_gap) is a useful free diagnostic — it correctly flagged Cabinet’s 500k model as under-converged and predicted, cross-task, where temporal abstraction pays (Part 6, 8/8 at fixed k). As a training objective, it is dead. - The composition figure stands (Part-pending): composed-d4 error is sub-linear in horizon and 1.3–1.8× below 1-step iteration — accurate, just unused by a short-horizon controller.
Where this leaves the program
Every positive lever this fortnight — across the state axis (16 clustering/SE/value-equivalence nulls) and the temporal axis (auto-k, calibration, spline-action, value-equivalent macro head, pyramid) — reduces to one finding: a converged self-predictive world model is already a sufficient, value-aligned, temporally-coherent abstraction, and explicit objectives that try to add structure are redundant or harmful. That is not a beat-TD-MPC2 result; it is a falsifiable redundancy criterion with evidence across both axes, a scored convergence predictor, and clean negatives on a dozen plausible levers. That is the paper. The one untested headroom the campaign ever flagged — high-DoF (Humanoid/Dog) at real budget — remains the only open positive bet.