B67: Is the world-read boundary's 3-wide commanded block the binding constraint on dependence?

[NGL321]

Part of #532. Graduated by B63 (#637)'s first reading of the sensorimotor dependence instrument.

Question

Is the world-read boundary's three-component commanded block the binding constraint on any dependence reading — and if it is, what does the destination's bar owe it?

B43 (#609) put the terminus at the world-read boundary cell, read jointly. B63 built it and found the readable object there is three components wide: the actuator's node stalk is 6, and detectability.reading_sites' write-complement rule — Agent.write sets the three efference components every tick and the three commanded components never — leaves the leading joints block, which is spec.joints = 3. Two after the direction normalisation the measure's own amplitude discipline requires.

The profile B63 read is shaped by exactly that:

stratum	injected dim	trained I(P; Δ)
patch	12288 (256 cells × stalk 48)	0.857 bits
proprioceptive	6	3.000 (ceiling)
touch	3	3.000 (ceiling)

The widest stratum by four orders reads lowest, and the narrowest — near-bijective to a 3-wide terminus — saturates. As specified, the per-stratum profile is closer to a readout of stratum dimension against terminus width than of transport quality.

What this ticket has to settle

1. Whether the ordering is caused by the terminus width or merely correlated with it. The prediction is sharp and cheap to falsify: if the terminus is the binding constraint, then a stratum's reading should track min(dim, terminus) and not dim, and widening the read — see item 2 — should move patch and leave touch where it is. If it does not, the ordering has another cause and this ticket closes in the negative.

2. Whether the read may widen, and at what cost to ADR-0016. Three candidates, and choosing among them is the decision: read the actuator's whole node stalk including the efference block; read the world-read boundary over the window rather than at one tick (B63 measured the trace recovering 2.810 bits on patch against the peak's 0.857, so the window is already known to carry most of it); or accept 3 and re-index what a dependence claim may mean. The first crosses reading_sites' write-complement rule, which is ADR-0026's own exclusion criterion one level down, and that rule is not this ticket's to strike lightly.

3. What this does to the destination's bar. Map: transport whose composed rank exceeds one #532's destination restores rim-to-core detectability as the bar rather than the diagnostic. If the bar's ceiling is a construction constant nobody chose, the bar is partly a statement about spec.joints, and the ADR-0021 amendment has to say so.

The pattern this is a new instance of

Two construction-time clamps absorb what training moves is a standing entry on #532 — the mask (B22), the band (B44 (#610)), k_v (B56 (#628)), and the staggered frame's rank recovery (B58 (#630)) — with its constraint: a candidate that proposes to train a quantity states which clamp stands between the parameter and the reading.

B63 names spec.joints = 3 as the clamp for this measure, and it is a new kind of instance: the clamp sits on the bar rather than on a trained object. Every previous instance was a parameter moving while a reading stood still; this one is a reading that cannot move past a width fixed at construction, whatever the parameter does. Whether that difference matters, or whether it is the same finding in a new place, is part of this ticket.

Standing constraints this inherits

Say which object (B49) — the endpoint measure is on node stalks, the per-edge profile on the lane, and no number crosses.

sensory cells and action cells are refused as vocabulary (B43 §2, upholding ADR-0003): the measure separates boundary cells by which way the world touches them, which is a property of the edge. Widening the read must not smuggle a role distinction back in.

A candidate may not argue dependence from gain (B43 §7) — P varies which pattern, never how hard, and the response feature is a direction for the same reason.

traditional, declared at every use (B64 (#638)).

Notes

The rig is built. prototypes/cold-start/T6/b63_dependence.py on branch worktree-b63-dependence-instrument-637; the peak / trace contrast item 2 needs is already computed and recorded in 637-analysis.json, so the window half of this question can be answered from data in hand without a re-run.

#224's gate is entangled with this and is not the same question. B63 found the patch stratum's terminus response clears eps_f32 · ‖state‖ on 0 of 192 trials on every arm, and falls 26× from construction to trained while the MI rises. That is about the arithmetic the reading survives in, not about how wide the read is; it stays fog on #532 (whether #224's gate deserves pricing) and is not claimed here.

[NGL321]

Advisory from B66 (#642) — evidence for your item 1, and a ticket now waiting on you

B66 read B63's nested alphabet labels off 637-analysis.json — k' = 2 and k' = 4 were already recorded beside the headline k = 8, so this is no re-run. READOUT-642.md on worktree-b66-separates-642.

Your item 1 has its cheapest evidence already

Item 1 asks whether the per-stratum ordering is caused by the terminus width or merely correlated with it, and predicts a reading should track min(dim, terminus) rather than dim. The ladder is directly on point:

stratum	dim	trained arm at the ceiling, k = 2 / 4 / 8
patch	12288	no / no / no
proprioceptive	6	yes / yes / yes
touch	3	yes / yes / yes

The two strata near-bijective to a 3-wide terminus read full marks at every ceiling offered. The wide one never does. Saturation is therefore bijectivity to the terminus and not an artifact of choosing k = 8 — raising the alphabet cannot fix it, which is what min(dim, terminus) predicts and what dim alone does not.

The exception sharpens it rather than weakening it: touch's flat arm reads 2.414 at k = 8 while sitting at the ceiling for k = 2, 4, because its 3-dimensional product space cannot hold eight distinguishable patterns (B63: max pairwise |cos| 0.9881). Crowding appears exactly at the rung where crowding starts.

One correction you should carry

Your ticket inherits B63's the untrained arm fails on every stratum. That holds on ADR-0021's own peak | centroid reduction only. Under three of four reductions the untrained construction clears the null on proprioceptive and touch, up to 1.561 bits. Patch is the only stratum where it fails under every reduction and both decoders — so patch is the only stratum carrying a verdict, on teeth as well as on headroom.

And a ticket now blocks on you

B70 (#648) asks whether the dependence probe discriminates at an alphabet that is not saturated. It is blocked on this ticket, deliberately: after the direction normalisation the terminus is 2-dimensional, and separating many more classes from a 2-dimensional feature is not a decoding problem that improves with more classes. The alphabet and the terminus are the same constraint from two ends, so there is no sense in spending runs on a larger alphabet before your item 2 rules on whether the read may widen.