bench(worldmodel): nuisance jitter fails structurally — reverted, recorded

[aurascoper]

Node 19's preferred generator repair. Acceptance criteria registered before implementing; depth fixed at 0.10 — the constant _render_state already uses — chosen by symmetry, explicitly not tuned. Registered rule: fail any gate = revert, do not adjust the constant.

All four gates passed

gate	result
G1a vel = 10·(c1/betaPower − 1) exact	2.33e-15	✅ jitter cancels in the ratio
G1b raw probe recovers vel	0.3605 (≥0.30)	✅
G2 oracle ladder	100.0 / 91.5 / 73.5 / 37.0 / 34.5	✅ bit-identical
G3 alphaPower post-from-pre	1.000 → 0.9709 (<0.99)	✅ subsidy reduced
G4 chi recoverable	0.6653 (>0.50)	✅

And the outcome regressed

Paired, 3 seeds, identical apart from the constant:

metric	jitter=0	jitter=0.10	delta
pos	+0.6065	+0.3118	−0.2947
vel	+0.0227	−0.0280	−0.0508
chi	+0.9566	+0.9306	−0.0260
peak_amp	+0.9947	+0.9814	−0.0133
ax	−0.0467	−0.3264	−0.2797
ay	−0.2305	−0.4947	−0.2642

It degraded the task factors 10–20× more than the nuisance factors — the exact inverse of its intent.

Mechanism

The task factors are multiplicative modulations on the nuisance carriers: c0 = alpha·(1 + 0.10·pos), alpha = f(chi, peak_amp, offset).

Carrier sd 0.220 against a total pos contribution of 0.0270 — leverage 8.2 : 1. A 10% carrier perturbation injects 0.0259, i.e. 96% of the entire task signal. The intervention put noise straight into the channel carrying the signal.

No usable depth exists — structural, not mistuned

depth	G3 R²	passes	injected noise vs entire task signal
0.01	0.9997	no	0.10×
0.05	0.9932	no	0.48×
0.10	0.9731	yes	0.96×
0.20	0.8953	yes	1.97×

The minimum depth that removes the subsidy already injects noise the size of the whole task signal.

Node 19's stated advantage of repair (2) over repair (1) does not survive: raising the velocity depth improves SNR directly, whereas jittering the carrier degrades it.

Reverted per the registered rule

synthetic_1f.py is unchanged on main. This node is the record.

The transferable part: the gate design was insufficient

G1–G4 verified the information is present in the raw observation and the control task is unchanged. Neither is the same as "a model can still learn it". A repair can pass every raw-signal check and still halve what the encoder recovers.

Any future generator change must add G5: factor recovery for pos and vel under the incumbent objective must not degrade. Training the incumbent is permitted — node 14 excludes candidate arms, not the existing baseline — and G5 would have caught this before the outcome measurement.

Next

The fault is the generator's structure, not its constants. Candidate: give pos and vel their own additive observation channels, independent of alpha/beta, so the task signal is not a rounding error on a nuisance carrier and the two can be perturbed independently. Larger than either node-19 candidate; needs its own pre-registration, including G5.

Ledger-only. Verified: smoke 2 runs, ledger and ADR gates green.