bench(worldmodel): channel ablation — ax inert, ay harmful, generator at fault

[aurascoper]

Pre-registered before running. H1 (velocity is the missing variable): zeroing ax costs nothing, ax alone ≈ random. H2 (horizon binding): both ablations degrade together.

Result

arm	log(d_f/d_0)	vs FULL	success
MPC full	+0.0886	—	35.0%
MPC ay_only (ax=0)	+0.0971	+0.0086	36.5%
MPC ax_only (ay=0)	+0.0374	−0.0512	33.5%
random	+0.0402	−0.0483	34.5%
zero	+0.0341	−0.0545	37.0%
ORACLE full	−11.11		100.0%
ORACLE ay_only	−8.72		91.5%
ORACLE ax_only	−4.52		73.5%

H1 confirmed, H2 refuted — |A−FULL| = 0.0086 against |B−FULL| = 0.0512, a six-fold separation.

The pre-registered discriminator resolves. ax alone is worth 73.5% to an exact controller and 33.5% to MPC — below the 37.0% of doing nothing. The authority exists; the planner cannot use it. The deficit is representational, not reachability.

New: all of MPC's harm is in ay

ay_only reproduces full MPC; ax_only collapses to the zero policy. So MPC only ever acts through ay, and those actions are uncorrelated with need (node 17, r = +0.04) — strictly worse than not acting.

Both channels fail, differently: ax is unrepresentable (vel R² = 0.033), ay is representable but at SNR 0.35, so its choices are noise.

The generator explains both

Contributed by review, verified here independently. _render_state does carry velocity — c1 = beta * (1 + 0.10·vel) — but as a 10% modulation on beta = 10^offset · 21^(−chi):

channel	total var	nuisance	task
c0 (pos)	4.82e-02	99.95%	0.845%
c1 (vel)	1.24e-02	100.13%	0.0106%

pos gets ~80× the share. (Review reported 1.15% / 0.05% from a different freeze-point — magnitudes differ, conclusion identical; both are recorded.)

And chi/peak_amp/offset are held per trajectory, so predicting them at t+1 is exact and free. The loss is minimised by encoding the nuisance, and VICReg's variance term rewards precisely those directions.

So node 18's vel R² = 0.033 is not a training failure. It is a working objective behaving correctly on a generator that made the task variable 0.01% of the signal.

Velocity is not destroyed: vel = 10·(c1/betaPower − 1) exactly, max error 2e-15. A ratio of two observed channels — nonlinear, hence invisible to a linear probe, trivially within an MLP's reach. Nothing prevents learning it; nothing rewards it.

Closes the representation-objective line

Nodes 6–18 were measuring the wrong thing correctly. No representation objective can encode a variable made 0.01% of the observed variance while paying a variance penalty for ignoring the other 99.99%.

Next: generator repair, under node-14 blind discipline

Two candidates, not equivalent:

1. Raise the velocity modulation depth — cheapest, but it is a knob, and tuning it until an arm wins is unfalsifiable.
2. Let chi/offset vary within-trajectory — removes the free-prediction subsidy without touching velocity's coefficient. Changes what the objective is rewarded for rather than what it is shown. Preferred.

Either way: specify using only oracle/random/zero, never a candidate arm; require the repaired generator to reproduce the ladder ordering (oracle > ay-oracle > ax-oracle > zero ≈ random ≈ MPC) before any arm is rerun; and add an acceptance assertion that a probe on the raw observation recovers vel above a floor, so a future generator edit cannot silently reintroduce this.

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