Part 4 closed the state axis: explicit abstraction on the representation is redundant exactly when the latent is value-sufficient — and on this substrate it always was. This post is the constructive sequel. If abstraction is redundant where sufficiency holds, then using abstraction right means aiming it where sufficiency fails. Our own data says where that is: not the state, but the planning axes — time and action.

Draft plan for review. Full version with costs and protocols: docs/research/abstraction-axes-plan.md.

The argument in three observations

  1. The one real win of the campaign was a temporal abstraction. The jumpy k-step world model — prior art, not ours — beat vanilla where sixteen state-abstraction levers could not: PandaPickCube- Orientation final $+90\%$, difference CI95 $[685, 1344]$, every jumpy seed above every vanilla seed (source: exp/tdmpc_glass/mechcheck/anchor_jumpy_vs_vanilla.json). A value-sufficient latent does not shorten the planning horizon. That bottleneck lives outside the representation — which is precisely why the redundancy argument never touched it.

  2. Planning cost is a search problem, not a representation problem. MPPI searches a $k \cdot d$- dimensional action sequence no matter how good the latent is. Sufficiency of the state never made the action space smaller. The criterion is silent about this axis — meaning it’s open.

  3. The temporal win is task-dependent and we don’t know why. Jumpy: $+90\%$ on Orientation, $+32\%$ (not separated) on Pick, a dead tie on Cabinet (1050 vs 1053). Something mechanistic separates these tasks. An unexplained split in your own positive result is a research question with your name on it.

Iteration 30 — three probes, all single-variable, all mechanism-check-gated

  • P1. A pre-test for WHERE temporal abstraction pays. The mirror image of the redundancy criterion: from existing checkpoints, compute candidate signals (k-step-error decay slope, reward smoothness/contact density, plan-horizon utilization), rank-correlate against the measured jumpy gains ($+90/+32/0\%$), then pre-register the best signal’s prediction on two unseen tasks and verify with a $k \in {2,4,8}$ sweep. Completes the paper’s predictive story in both directions.
  • P2. Hermite-spline action bottleneck (the ledger’s #1 untested lever). Macro-action = spline knots, PD tracker executes; MPPI’s search shrinks from $k \cdot d$ to $2d$ with no learned codec, so the redundancy result cannot touch it. Gate only if splines reconstruct winning replay trajectories with $\geq 95\%$ return preservation.
  • P3. Value-equivalent macro head. Our VE null tested latent value-equivalence on the 1-step model; this trains the k-step head to preserve macro-return instead of state — the one VE variant the null does not cover, riding the only model where capacity actually binds.

Parked: compositional-OOD state abstraction — the synthetic gate showed value-decodability does not collapse at held-out object counts ($R^2$ 0.96 → 0.92/0.94), so there is no headroom signal yet.

Compute, honestly

Fleet right-sized today to 6 boxes ≈ $0.57/hr ≈ $14/day: a 5070 Ti dev box for mechanism-checks plus 4 A4000 workers and one cheap 3060. That sustains one full 30-run gate (~2 arms × 3 tasks × 5 seeds) every ~2 days with same-day mechanism-checks — and since the protocol’s mechanism-check gates mean at most one gate is ever worth running at a time, more GPUs would not make the science faster. The bottleneck is decision quality, not compute.