⚠ Correction (2026-07-02). This post asserts as its central correction-of-record that “planning is a directed-exploration operator.” That thesis was subsequently refuted three ways by the controlled, noise-matched plan-vs-π ablations (Part 7: no coverage gain, discovery earlier without planning), and the exploration advantage was relocated first to the world model and finally — after the SAC control — to an on-policy pathology of PPO with the world model as a sample-efficiency lever (Part 8). The five bets below mostly resolved null (Part 7/8 scorecards). Also: the “R²≈0.999 value-probe” cited below was retracted as an instrument by the R²-criterion postmortem — the redundancy conclusion rests on the direct nulls, not that probe. Kept unedited as the planning document of record.

A planning document. The Part 5 method map mapped where four families (state/behavioral/planning/PPO) win across Panda + DMControl. A supervisor review then reframed why, and pointed at what to do next. This post is the honest reframe + five proposals we’re now running in parallel.

Two corrections that reshape the program

1. The breakthrough is exploration, not learning speed. Our sharpest result — TD-MPC2 beats PPO on HopperHop 367 vs 33 — is not a ceiling story and not even mainly a “speed” story. It’s that planning is a directed-exploration operator: a learned model + lookahead reaches gaits (and, generally, hard-to-reach states) that PPO’s on-policy sampling never finds. On dense/explorable tasks PPO’s throughput wins; on exploration-bottlenecked tasks, planning explores its way to a policy PPO can’t reach at any budget. That is the interesting claim, and it subsumes the “speed-not-ceiling” one.

2. Anti-collapse is a JEPA property, and we tested it on the wrong object. A pure JEPA has no reward/value loss — anti-collapse is the only thing preventing latent collapse, so it’s load-bearing. TD-MPC2 is not a pure JEPA: its value/reward loss already anchors the latent (value-sufficient, linear value-probe (R^2\approx0.999)). So our DMControl “anti-collapse” study actually measured redundancy (a relational regularizer fighting an already-anchored latent), not JEPA anti-collapse. Only the nav H-JEPA result — where the latent is genuinely self-predictive and collapses to effective-rank ≈ 0 — is a real JEPA test. To carry the JEPA question onto DMControl honestly, you strip the value loss (Proposal D).

The five bets

A — Planning as a directed-exploration operator ★ flagship

Hypothesis. On exploration-bottlenecked tasks, model-based planning explores more effectively than model-free RL, and that exploration is the win. Make it explicit and steer it with abstraction.

  • A1 — Isolate. On HopperHop and the actuation-weakened escape frontier, measure state-space coverage (occupancy entropy; number of distinct structural-entropy communities visited) for TD-MPC2 planning vs PPO vs a no-plan ((\pi)-only) ablation. Show planning → coverage → success.
  • A2 — Amplify. Add an intrinsic-novelty bonus to the MPPI objective (model-disagreement / latent count) — Plan2Explore inside TD-MPC2. Does it push the escape frontier to even weaker actuation?
  • A3 — Direct (SE). Run min-2D structural entropy on the replay-buffer latent graph → communities are abstract states, boundaries are bottleneck subgoals → a high level plans toward under-visited communities. This uses SE for what SE is actually for — structure discovery — not as a latent regularizer.

White space. Plan2Explore / LEXA / MAX / Go-Explore explore via curiosity or disagreement, but nobody uses a learned abstraction to define the exploration targets a planner pursues. We already own SE, a planner, and a controlled exploration-difficulty axis. That intersection is the paper.

B — “When does a behavioral prior help RL?” (taxonomy paper) — finalize

The 2-axis taxonomy (prior-fit × exploration-difficulty) with the escape-difficulty frontier is drafted and honest: a fitting prior is a speed lever; a mismatched one is dead weight (an anchor); on exploration-hard tasks it’s an escape. Finish the figures and submit — it’s a clean empirical/analysis contribution.

C — Abstraction as variance-reduction

On the 16-task benchmark, the SE “glass” latent beats TD-MPC2 on 6/16 tasks — a wash on the mean but with visibly lower seed-variance on several (ReacherHard 976±3 vs 883±151). Possible real signal: SE structure doesn’t raise the mean, it reduces collapse-seed variance / improves worst-case. Cheap per-seed check first; report honestly either way (it’s a section, not a headline).

D — JEPA anti-collapse, done right

  • D1 (running) — the SE arm on TD-MPC2 DMControl: does SE also fight the value structure like uniformity, and does SE+uniformity combine? (A redundancy datapoint, correctly scoped.)
  • D2 — the real test: a pure self-predictive JEPA latent on DMControl (no reward/value loss), probed on a geometric readout and a value/return readout. Does the downstream-dependent taxonomy hold on a genuine JEPA?
  • D3 — pixels, where the information term is load-bearing: non-generative JEPA vs generative Dreamer world model, matched budget. On low-dim state, anti-collapse is redundant (we showed that); on pixels it should matter.

E — Structural entropy for structure discovery, not regularization

We used SE to regularize the latent — redundant, and the wrong bias for continuous geometry. SE’s real strength is community/hierarchy detection → subgoal / option discovery and hierarchical planning (this merges into A3). That’s where “glass” belongs after SE-as-representation proved a dead end.

Running them in parallel (this week)

box thread
b3060 (DMControl, 4×3060) SE arm finishing (D1) → A1 coverage → A2 intrinsic-novelty MPPI → A3 SE-subgoals
b3060b (Panda/pixels, 4×3060) beat-PPO tail → D3 pixel JEPA-vs-Dreamer → D2 pure-JEPA-on-DMControl
no-GPU / writing C per-seed robustness · B finalize taxonomy paper · planning-exploration proposal

Priority if time-boxed: A1 (de-risk the flagship) > C > B > D2/D3 > A2/A3. Everything is deterministic, disk-backed, and reported with matched controls — including the nulls, of which this program has produced many on purpose. Results will land in the ledger as they complete; a Part 7 will report them.