bench(worldmodel): multi-step training works; control still does not move

[aurascoper]

Node 22's named next test.

Blocker found first, and it is a finding

The corpus contains no trajectories. generate() samples a fresh state per record (synthetic_1f.py:104), so record i's post window is unrelated to record i+1's pre window. There was no multi-step ground truth anywhere in the data — nothing to train a multi-step predictor against and nothing to validate one on.

Node 22's test was not runnable as specified. It needed generate(..., trajectory_length=K) and train_jepa(..., rollout_steps=K) first.

Gates

gate	result
M1 defaults byte-identical (8f74f84ac5aaebe2)	✅ a flag flip, incumbent reproducible
M2 post==pre within a trajectory, not across	✅ both directions
M3 oracle ladder bit-identical	✅ 100.0 / 91.5 / 73.5 / 37.0 / 34.5
M5 multi-step error must improve	✅ 0.02311 → 0.00546, 4.2×
M4 pos ≥ 0.9824	❌ 0.9725

M5 — the mechanism works

h=8 rollout error fell 5.7× against the same corpus at K=1. Multi-step training does exactly what it is supposed to do.

M4 — attributed, not excused

corpus	K	pos	vel	h=8
single-transition (baseline)	1	0.9824	0.9740	0.02311
trajectory	1	0.9818	0.9709	0.03116
trajectory	6	0.9725	0.9896	0.00546

The corpus costs nothing. The multi-step loss trades −0.009 pos for +0.019 vel and the rollout gain — a trade, not a regression. But M4 was a registered hard floor and it was not met.

And node 22's outcome gate fails

Deployable predictor rollout, 3 seeds:

goal latent	mean	node 22
TARGET	36.3%	35.5%
ONLINE (one-frame)	36.3%	35.0%

Against zero-action's 37.0% and a gate of approach 76%. Cutting rollout error 5.7× moved control by nothing.

Reverted

Per the registered rule, without tuning K. Also right on the merits — a corpus mode and two parameters for no demonstrated benefit. Node 20 set the precedent. synthetic_1f.py and eeg_jepa.py are unchanged on main; this node is the record.

Node 22's three-condition model is now in doubt

It predicted that treating the last untreated condition would approach 76%. It was treated, measurably, and control did not move.

Two readings remain, and node 22's 2×2 cannot separate them because its true-env arm used exact dynamics rather than merely better ones:

• (a) the residual gap between 0.00546 and exact zero is still decisive — control would be a near-discontinuous function of forward-model error, and the 76% arm only ever reachable with perfect dynamics
• (b) a fourth condition is unidentified

Next test adjudicates rather than adds a candidate

Sweep control against injected rollout error, interpolating between the true env and the learned predictor across 0, 0.00546 and 0.0231. One sweep, no new model, and it settles a question five nodes have circled.

Ledger-only. Verified: generator and forward_eval smoke tests pass; ledger and ADR gates green.