Living log. Context: Part 5 method map, Part 6 five bets, and its sibling Thread A (A3 is where this thread cashes out).

The bet

We spent a lot of the campaign using structural entropy (SE) as a latent regularizer — pushing the world model’s latent toward a min-2D-SE community structure. Two threads’ worth of results say that’s the wrong job for it:

  • Thread D showed anti-collapse regularizers (uniformity, VICReg, SE) are redundant-to-harmful on value-anchored control, and that even on a pure JEPA, SE-as-regularizer is at best neutral.
  • On continuous geometry specifically, a community bias is the wrong inductive bias — continuous control lives on smooth manifolds, not discrete communities.

But SE’s actual strength is what it was designed for: community and hierarchy detection. The unclaimed, untested uses are structural, not representational:

  • Subgoal / option discovery — run min-2D SE on the replay-buffer latent (k)-NN graph; communities are abstract states, boundaries are bottleneck subgoals.
  • Encoding-tree as planning hierarchy — the SE encoding tree gives a natural high-level / low-level split for hierarchical planning.
  • SE-guided exploration — plan toward under-visited communities (this is exactly Thread A’s A3).

In one line: SE belongs on the graph of states, not inside the latent vector. This thread is where “glass/SE” lives after SE-as-representation proved a dead end.

Status board

item state verdict
SE-as-latent-regularizer (retrospective) ✅ closed redundant / wrong bias (see Thread D)
SE-for-subgoal-discovery (A3) ❌ NULL (2026-07-02, n=2, 3 sparse tasks) SE-community subgoal shaping ties or hurts flat (Acrobot 460→382; CartpoleSwingupSparse 316→7 — hurts most exactly where it was meant to help; FingerTurnHard ~tie). See Part 8.

Progress log

2026-07-01 — repositioned; execution merges into A3

No standalone GPU run yet — this thread is a reframing plus a concrete plan that executes as Thread A’s A3 arm. The honest retrospective: every attempt to make SE earn its keep as a latent bias (Panda glass, DMControl glass, the D-thread SE arms) came back redundant or an over-weighting artifact. The forward bet is to use SE for structure discovery — replay-graph → SE communities → bottleneck subgoals → a high level that plans toward under-visited communities. That is A3, and it’s the one place the whole “glass” line can still contribute something novel: SE at the {planning × abstraction × exploration} intersection, not as a regularizer.

Dependency: A3 is gated on Thread A’s coverage machinery (the replay-graph + coverage logger already exist). It runs after the A1 flagship de-risk resolves.