2026-09-04
An attention mask answers the simplest question in a reasoning machine: who may look at whom? Every cycle of thought, every cell of a problem attends to some neighbors and ignores the rest — and that geometry of permission decides, before any weight fires, what the machine can possibly conclude. If deliberation is the engine, the mask is the steering wheel.
For most of this project the steering wheel was welded straight. The mask was set once, from a first reading of the problem, and held. Later we added a reflex — committed conclusions could crack open a few extra lanes of attention — but a reflex is not a driver: it had no learned parameters, no dedicated heads, no ability to weigh what it saw. The machine could think for seven cycles but could not meaningfully re-decide where to look as its understanding changed. An audit this week made it official: dynamic masking was the most under-resourced organ in the machine, billed as central, built as an afterthought. That is now being fixed — a dedicated, multi-head mask head whose whole job is steering.
But the deeper discovery came from excavating our own history.
Two years of hand-built puzzle solvers — Sudoku, map coloring, KenKen, logic circuits — each needed hand-crafted masks: for Sudoku, attention along rows for some heads, columns for others, boxes for others. We wrote those rules by hand, domain by domain. Digging through all of them this week, we found that every one is the same rule wearing different clothes:
Allow attention between two cells IF they share a constraint factor of type t.
That’s it. Same row? A shared row-constraint. Same cage? A shared cage-constraint. Connected by a wire? A shared gate. One line, five domains, every mask we ever built. The rules we thought we were inventing were instances of a single sentence about structure.
When one sentence covers everything, it deserves to become a language. So the mask head is getting a DSL — a tiny formal language of looking, whose entire vocabulary is that sentence and a few honest atoms: shares-a-group, committed-by-the-solver, self, all, union, intersection. The neural mask head proposes a program; deterministic machinery executes it into an exact mask. Neural proposes, symbolic disposes — the deepest law of this project, now applied to attention itself. And the language carries its safety in its grammar: there is no negation atom, so a program can open lanes of attention but can never close the baseline ones. The forbidden act isn’t checked for. It is unwritable.
The test we have pinned for this system is our favorite in the whole campaign. We will hand the machine a Sudoku puzzle described in plain prose, and ask its mask head to steer: parse the rules, propose the program, execute the mask. The pass bar is exact equality — bit for bit — with the mask we wrote by hand two years ago, before any of this existed. The machine, reading language, must re-derive its own history.
A steering wheel, a one-sentence language of looking, and a driving test graded by our younger selves. That is where dynamic attention is headed.