Eighteen nulls in, the redundancy criterion has held against everything — clustering, SE, value-equivalence, calibration, pyramids, high-DoF. This post reports the first mechanism-check that came back GO. It is real, it survived its fairness control, and it is narrower and more honest than the headline number first suggested.

The result

Contact-rich multi-disk world (elastic collisions, known contact graph, configurable object count). Graph WM (entity-factored transformer) vs a param-matched monolithic WM. Pre-registered GO iff monolithic OOD value-R² drops ≥0.15 below the graph’s, OR graph predicts contacts better.

baseline graph OOD R² monolithic OOD R² gap
pooled monolithic (n=4 seeds) 0.57 0.21 0.35
fair zero-padded monolithic (n≥2) 0.57 0.40 0.18

GO — graph value-decodability generalizes to held-out object counts; monolithic collapses (N=5→9: 0.29→0.18 pooled). Crucially it survives the fairness control: a zero-padded monolithic that keeps a slot per entity (lacking only permutation-equivariance) recovers a lot of ground — the pooling was inflating the gap — but the residual graph advantage still clears 0.15 every seed (source: exp/tdmpc_glass/mechcheck/gwm_verdict_robust.json, gwm_simulator_.json)*.

The fine print (this is the honest part)

  1. Half the original win was a baseline artifact. Pooling crushes N entities into 2 summary vectors; fixing that (pad) closed the gap from 0.35 to 0.18. We caught it because we ran the control before celebrating.
  2. It is NOT about contacts. The contact-prediction criterion failed — the graph predicts contact-step dynamics no better than monolithic (ratio 1.01×; contacts are ~10× harder for both). So this is “graph generalizes across object counts,” not “graph simulates contacts better.” That reframes the win away from the survey’s graph-as-simulator slice toward graph-as-compositional- generalizer.
  3. It is representation-level, not control. R²≈0.57 from a linear value-probe is modest, and the campaign’s entire lesson is that representation properties don’t automatically become control wins.

The real gate (next)

The pre-registered escalation question is now sharp: does this OOD value-decodability advantage show up in control return at held-out object counts? Train a planner on both WMs at N_train, evaluate return at held-out N. If yes — graph WMs earn a genuine compositional-control claim and the ManiSkill/Robosuite escalation is justified. If no — it joins the redundancy story as “another representation property that doesn’t pay at the controller.” Either way, the first GO has earned that test.