Living log. Context: Part 5 method map, Part 6 five bets.

The bet

On the full 16-task DMControl benchmark, the SE-structured “glass” latent beats plain TD-MPC2 on 6/16 tasks. The means are small and both-ways — a wash. But on several tasks glass had visibly lower seed-variance (e.g. ReacherHard 976±3 vs 883±151). The hypothesis worth a cheap check: maybe SE structure doesn’t raise the mean, it reduces collapse-seed variance / improves worst-case — “abstraction as variance-reduction.” That would be a defensible section (not a headline).

Status board

item state verdict
Per-seed analysis of the 6 win-tasks + out-of-sample across all 16 ✅ done NULL

Progress log

2026-07-02 — VERDICT: NULL (analysis-only, n=4–5)

Tested whether glass’s 6/16 wins are a real worst-case/variance-reduction effect. They are not.

  1. Within the 6 win-tasks, glass has lower seed-sd and higher worst-seed 6/6 — but this is circular: the tasks were selected because glass’s mean is higher, which (with equal ceilings) is definitionally “TD-MPC2 owned the unlucky seed there.” Dropping the worst seed erases the win on 4/6 (only FingerSpin, ReacherHard survive).
  2. Out-of-sample across all 16 tasks, the effect reverses: mean seed-sd is glass 123.2 > TD-MPC2 114.5 (glass has higher variance overall); glass is lower-sd on only 8/16 and better-worst-seed on 9/16 (coin-flips). Glass has its own collapses that TD-MPC2 avoids: HopperHop 0 vs 179, HopperStand seed=15 vs 858, ReacherEasy ±89 vs ±3, WalkerWalk ±81 vs ±12.
  3. Cross-algo eff_rank correlation is unverifiable — no TD-MPC2 latent diagnostics were saved to disk.

Conclusion: glass ≈ TD-MPC2 is a genuine wash; the 6 wins are small-mean, both-ways scatter from unlucky seeds. No publishable variance-reduction section. This reinforces the broader redundancy result — an extra relational bias on an already-value-sufficient latent buys nothing on continuous control. Recorded honestly and closed. b3060:exp/proposal_C_variance/{VERDICT.md,per_seed_data.json}.