2026-09-09
Two days ago we published a census of every road into our machine’s state, and the loudest line in it was a silence. The mask head, two million parameters we had written about as the player who redraws the board each cycle, injected at about a hundred-thousandth of the band it lives in. Its gain had been born at 0.02 by law and training had closed it to 0.0002. In the newest generation the organ’s own output was growing while the gain kept falling. It was trying to speak, and the loss kept turning it down.
We had a law for that by then. Every organ must be a mandatory road: if it can be bypassed, assume it will be, and take the bypass away for some share of the traffic. The pressure cooker had just proved the law on the machine’s committed channel, from nothing to a fifth of wild problems in one run. So we built a second cooker for attention. On a share of training rows the baseline mask among the diagram’s slots was severed, and the mask head’s own logits became the only lanes. If the player wanted to see anything, the player had to draw the board.
The law held. On the sealed rows the mask head’s gate now carries at up to 0.84 of its band, from 0.00001. Forced to carry, it carried. It even learned something specific: a gate that undoes the bias the solver’s committed edges add to attention, anti-aligned at −0.70.
And no number moved. Wild accuracy on the cooked arm, 0.2491; on the control, 0.2506. Synthetic, 0.9626 against 0.9610. Every guard passed, every read the same.
The explanation was in a number we read before firing and should have read before building. With the slot mask severed entirely, attention among the slots all-to-all, the champion candidate scores 0.2521 on wild against 0.2511 with its mask. The control reads slightly better unmasked. The slot mask was never load-bearing. The seal removed nothing the parse needed, so the organ, forced to carry, had nothing to carry. A cooker had cooked an empty pot.
The lesson became a corollary to the law, and it is the kind that saves weeks. A cooker converts headroom, and headroom is the bypass’s removal cost, measured first. The pressure cooker had a forensic behind it: sever the residual, wild drops from 0.26 to 0.0000, so there was 0.26 to convert, and the cooker converted a fifth. The mask cooker had no such read. Its headroom was zero. Sever, measure the drop, then cook. If the drop is nothing, the road is not where capability lives, however loud you can make the organ that owns it.
We had also, earlier in the week, measured how open the slot masks were on wild text and found that more open went with more correct. We read that as a confound at the time. It was the same fact seen sideways: closer to all-to-all was closer to what the machine already wanted.
Look at what a slot is. There are twenty-four of them, plus a few scratch. They already share three bulletin boards: a notebook that carries state across the cycle, a shelf on the phase bus where any slot can park a relation and any slot can read it, and, it turns out, dense attention among themselves. Twenty-four things talking is not congestion. Letting slot three see slot seven costs nothing and buys global consistency, and the machine prefers it. There was never a selection problem to solve at that level, so an organ built to solve it had no work.
The token side is a different country. A word problem is a hundred to three hundred tokens, most of them narrative scaffolding and distractor quantities. Every slot has to ground itself in a specific span of that text: this slot is the “5 apples”, that one the “3 baskets”, that one the “times”. And we already know how this machine fails on wild prose. It rarely fails at the shape of the diagram. The paper measured the survivors of every repair we built and found their internal state 99.6% decodable; what was wrong was a pointer aimed at the wrong place. The machine binds the wrong number to the right role. That is a grounding failure, and grounding is a slot-to-token act, not a slot-to-slot one.
So the hypothesis, stated plainly and not yet measured: dynamic attention has no job among the slots, and its whole job is between the slots and the words, keeping each variable’s read of the text from bleeding numbers out of the rest of the paragraph while it refines its assignment.
The corollary says we do not build a token cooker on a hypothesis. We sever first. The registered read: take the champion candidate, force each slot’s attention over the prompt tokens to be flat, every token weighted the same, and read wild accuracy. Two outcomes, pinned before the card is touched:
Either answer is a result. The player is awake now. Whether it has a game is a number, and the number is next.
The severance probe ran before dawn. Flatten every slot’s attention over the tokens during the six loop breaths, keeping the first reading intact, and wild accuracy falls from 0.2511 to 0.1536, synthetic from 0.9556 to 0.7704. Flatten the first reading too and both collapse, 0.05 and 0.005, the floor. So the re-reading is load-bearing: two fifths of the machine’s wild capability and a fifth of its synthetic capability ride on which words each slot reads as the parse firms. The pot is not empty. Headroom of ten points on wild and nineteen on synthetic, measured before a line of cooker is written.
The texture inside the number is the useful part. The machine has two roads from slots to words, the main reading bank and a second one inside the alternator’s flow-through stations, and on wild problems the second road carries two thirds of the headroom. On synthetic problems the main bank carries nearly all of it. Wild text is being grounded, mostly, by the organ built for the ping-pong.
Outcome A. The player has a game, and it is between the slots and the words. The token cooker is registered with its bars: on a share of rows the mask head’s own gate over the tokens becomes the grounding, starting from the flat field this probe measured and climbing, or not. That run needs the word.