B46: Does a gate carry the carve's reversibility, and what may it read?

[NGL321]

Part of #532. Opened by B41 (#604)'s resolution, on the user's own proposal, and blocked on B42 (#605) because the cost this ticket exists to avoid may not survive B42's answer.

Question

Does a gate carry the carve's reversibility — a pruned edge that still transmits but whose transmission never reaches the receiving stalk — and what is the gate allowed to read?

Why this is now a ticket rather than fog

B41 ruled that a pruned edge's warrant must be readable without anything reaching the stalk, and that m_e = 0 means zero carried directions. It named two implementations and adopted the cheaper one to state:

• Out-of-lane read. The edge survives as an object whose candidate map is evaluated but never used to transport. Nothing crosses. Its cost is a counterfactual — under the per-edge parameterisation the dormant map is driven by nothing, which is exactly what B34 (#593) §1(c) chose the probe floor to avoid.
• Gating. The lane stays live and carries real traffic under real gradient; the boundary moves — what is blocked is admission to the receiving cell's stalk, not the transmission. The warrant is then read on a channel that is measured rather than imagined, and ask 2's "a measurement channel outside the lane" is achieved by moving the boundary instead of by building a second channel.

Gating dominates on that one axis and is not free. It is the user's proposal, and their framing is that it may separate the mechanisms enough to make learning easier, and that the same object is a candidate for attention later:

"I also think it's worth considering a selective gating approach where pruned edges still transmit, and that's fine, but transmission never actually reaches the stalk because it's gated. [...] this could potentially add a mechanism that can be used down the line for attention."

Why this is blocked on B42

The advantage is parameterisation-dependent. Gating's edge over the out-of-lane read is that it kills the counterfactual. Under a shared per-cell frame there is no per-edge parameter to be counterfactual about — a dormant edge's map is a computed consequence of its two endpoints' frames, trained by that cell's other live edges. So gating is worth a great deal under the per-edge scheme and close to nothing, for this job, under the flat bundle. B42 decides the parameterisation; asking in the other order pays for machinery that may already be free.

Note the asymmetry: B42's answer can remove this ticket's motivation without removing its subject. The attention hook survives either way — it is simply not this map's destination (see Constraints).

What this must decide

1. Whether transport carries a gate at all. A gate is state-dependent routing and this map's destination covers transport and the rule that trains it (B27). A hard threshold is non-differentiable; a soft one puts a nonlinearity in a path that has been linear throughout — candidate 1's stated virtue on this map is "still linear; adds no term to any update." Name the ADR cost.
2. What the gate reads. See the constraint below: it may not be the arriving amplitude.
3. Whether the gate is per-edge or per-direction, and whether a gated direction still counts toward Σ_e m_e ≤ B. A lane that transmits but does not land is spending bandwidth on nothing, or is not spending it at all, and the budget has to say which.
4. Whether gating changes what B41 ruled, or only how it is implemented. B41's requirement is the ruling; if a gate satisfies it more cheaply, that is an implementation swap and not a re-resolution.

Constraints already binding

• The gate's variable may not be the arriving amplitude. B41's finding, and it is the leak's own mechanism turned around: under ADR-0032's band a map is a near-isometry whose norm spreads over the m directions it carries, so at m = 1 all of it lands on one and a pruned edge's surviving direction is its strongest-transmitting one (B40: floored 1/√3 against kept 1/√3.75, √1.25 = 1.118, 88.9% of floored edges above the kept median, unmoved across seeds and checkpoints). A magnitude threshold would preferentially open for exactly the leak it exists to stop. B41's candidate, carried as a candidate and not as a decision: gate on agreement — admit what arrives to the extent it coheres with what the receiving cell already predicts, which is the same holonomy quantity the criterion counts, read instantaneously rather than around a cycle, so it introduces no new object and keeps B34's one criterion, not two. The alternative B41 saw — gate on amplitude with the 1/√m concentration divided out — works and is a correction factor whose only justification is cancelling an artifact.
• Reversibility, graded (B41, the map's standing constraint): a carve is a reallocation, not a deletion, and the bar is that a cell whose dynamics change can still come to communicate meaningfully — including its stalk changing commensurately. Universal reversibility is not required and the degree may vary.
• ADR-0011: computable by the cells incident on the edge. A gate at the receiving cell is trivially local; a gate that reads both endpoints is not automatically so.
• Attention is out of scope on this map. It is a use of the mechanism, not a step toward this map's destination, and it returns as its own effort. It is recorded because the gate's design here should not foreclose it — not as something to be built or decided.
• Building is out of scope (Map: transport whose composed rank exceeds one #532, Plan, don't do).

Notes

HITL. Invoke /grilling and /domain-modeling.

Read B41 (#604)'s resolution first — it carries the two implementations, the amplitude trap and the reversibility ruling — then B42 (#605)'s answer, which is what decides whether the counterfactual this ticket exists to avoid is real.

B32 (#591)'s steganography hazard bears directly: a gate is a place a loop could learn to close on a channel the task never reads.

[NGL321]

Advisory from B42 (#605) — two things changed under this ticket

Advisory, not a rescope: B42 did not edit this body, and what follows is for whoever claims it to weigh.

1. The gate's subject is enlarged to the node stalk mask. B41 left the mask half — graph.py:693's "the mask closes and never re-opens" — live on B42. B42 found that under #562's flat reserve the mask never closes in response to anything: it is n − p for every predicting cell, fixed at construction. Read as an exemption that would have ended the matter.

The user ruled the opposite way, and it is the decision that most affects this ticket:

"Why did we want it fixed at construction? Oh no, that's the problem. Then yes, we should apply the gating principle and learned mask that we were discussing for B41."

The mask is fixed by inheritance from the superseded capacity-derived bound, where it was a structural consequence rather than a choice; #562 made it flat without making it movable. So a cell's split between its own reading and the room's cannot move as it learns what it is for — in today's architecture or in the flat bundle. Decided on B42: the reserve mask becomes a learned, gated object, carrying B41's graded reversibility through the gate rather than through a floor. This ticket's gate should therefore cover the node stalk mask and not only Edge.m.

2. This ticket is now blocked on B47 (#614). B42 found that B40's arm-4 falsifier of B34 fired, and was scored on a spread statistic that cannot see it: on a genuinely flat bundle the criterion returns each edge's own m_e on 66% of edges, mean shortfall 1.0 against the trained control's 13.8. Designing a gate around a criterion that may be a mirror is work that could be wasted, so the edge is wired rather than left as a note.

Unchanged: B41's ruling that a carve is a reallocation and never a deletion, and that reversibility is graded and functional rather than universal. B42 strengthens the case for the gate — it is now the mechanism for two objects, not one.

[NGL321]

Advisory from B48 (#615)'s resolution, plus a queue audit the user asked for. Not a rescoping — the body stands, every ask is still live, and this ticket's motivation has survived its blocker. Four inputs.

1. The B42 block is resolved, and it resolved in this ticket's favour

The body says this is blocked on B42 (#605) "because the cost this ticket exists to avoid may not survive B42's answer" — gating's advantage over the out-of-lane read being that it kills a counterfactual that only exists under the per-edge parameterisation.

B42 refused the flat bundle as architecture. The per-edge parameterisation stands, so the counterfactual is real and gating's advantage is intact. The conditional that would have hollowed this ticket did not fire. B42 remains a closed blocker on the dependency graph; the only open one is B47 (#614).

And B42 enlarged the subject: the reserve mask also becomes a learned, gated object — graph.py:693's "the mask closes and never re-opens" is now under B41's reversibility constraint, and p is an initialisation, not a constant. So this ticket's gate has two surfaces to consider, the edge lane and the node stalk mask, not one.

2. The leading gate candidate is now under B48's rulings, and may be the same object as B52 (#622)

B41's carried candidate is gate on agreement — admit what arrives to the extent it coheres with what the receiving cell already predicts, the same holonomy quantity the criterion counts, read instantaneously rather than around a cycle.

B48 ruled two things that bear on it directly:

• The joint scoring rule, now a standing constraint on this map: any agreement statistic is reported alongside audience differentiation at a cell and the exposure it cost, and a candidate must show agreement rising while differentiation stays nonzero. A gate variable is an agreement statistic, so it inherits this. The flat bundle's 88 earned dimensions at audience differentiation 0.0000 is the null it must fail.
• Holonomy is suspended, not struck — and the suspension is of B33's local holonomy term in the objective, not of holonomy as a read-only quantity. A gate that reads agreement is not caught by it. Stated explicitly so this ticket is not read as blocked by B48; the re-entry for the objective-term question is B54 (#624).

The coupling worth noticing: B48 ruled that earned H⁰'s replacement must be graded and read on node stalks, and minted B52 (#622) to specify it. An instantaneous, graded agreement quantity read at the receiving cell's stalk is a description of both objects. B52 may supply this ticket's gate variable, or the two may specify incompatible versions of one thing. Not wired as a dependency — that is the user's call, not a resolver's — but whoever takes either should read the other first.

3. A gate is a third clamp, and this map has two that ate what training moved

B48 recorded as inference on two instances: the mask on restriction maps (B22 — every rank-derived number invariant exactly over 20k ticks, because project() re-applies a construction-fixed mask) and the band on K (B44 (#610) — ρ(K) moves on 149 of 150 cells while σ(used) sits pinned at exactly 1.000). The trained parameter moves and the quantity the architecture reads does not.

A gate on admission to the stalk is structurally a third clamp of the same kind, and this is the strongest available argument for the body's ask 1 (whether transport carries a gate at all) and ask 3 (whether a gated direction counts toward Σ_e m_e ≤ B). The map's constraint, as written: a candidate that proposes to train a quantity states which clamp stands between the parameter and the reading. A learned gate must say what its own reading is downstream of. If a third instance lands, the pattern stops being inference.

4. Unchanged and still binding

The amplitude trap (a magnitude threshold preferentially opens for exactly the leak it exists to stop — √1.25, 88.9% of floored edges above the kept median), ADR-0011 locality, graded reversibility, B32's steganography hazard, and attention being out of scope. Nothing in B48 or B44 touches any of them.