title: The line cooks date: 2026-09-10
A single day’s reads, in the order they happened, because each one changed the next. Two were bugs in the clock. One was a question about who gets paid. One was a wire we had never strung. And the last is the machine’s steering wheel, turning for the first time as this is written.
Every breath, the state’s clock planes turn sixty degrees. Every breath, the machine also writes two memories from that state: an ink into the notebook, and a deposit onto the bus. Both writes go through dense maps, and dense maps do not respect planes. So the phase of the clock at the moment of writing was smeared through the whole memory, and when the memory was read three breaths later, nothing undid it. The reader received a bearing three breaths stale, mixed into everything else.
The fix is a canonical frame: de-rotate the clock block to phase zero before either memory is written, and re-phase what comes back to the reader’s own bearing. One primitive, both memories. It was free, and a hair better on synthetic, a new record at 0.9708. It did not make the bus more necessary, which is what I had predicted; the bus’s removal cost went down, not up. The frame’s value turned out to be what it made possible next.
When we cut the residual connection earlier in the week, we exempted the clock’s sixty-four planes so the notebook’s ink would not overwrite the compass. Then a spectral read asked where the machine keeps the difference between adding and multiplying, and the answer was: fifty-five percent on the clock planes, with its six strongest planes all clock planes. The clock is not a pure coordinate system. It is a rotating frame, and the amplitudes in that frame are free, and the loss had moved a third of the state’s variance onto them, up from a sixth, because they were the one road the cut left open.
We measured the leak directly. The v1 candidate, read with the clock planes cut too, fell from 0.2608 to 0.1497 on wild and from 0.9654 to 0.6513 on synthetic. Eleven and thirty-one points were riding the exempt block across the cut. Then, because the canonical frame carries the bearing through the ink, we cut everything: the total cut, no exemption. The result reads 0.2526 on wild, inside the guard, and 0.9757 on synthetic, another record, with the clock probe perfect and the clock planes’ share of the variance back down to 0.22. The bypass is closed. This is the residual-seal law, which we wrote on Wednesday, catching my own Thursday design.
A question from the gut: we have a chain of seven breaths and one loss at the end. Does the gradient reach the farmer, the first breath, who never sees the customer but without whom there is no meal?
We built a gradient census: a probe on the state entering each breath, the real training loss, backpropagation, and the norm of what each breath receives, with the readout’s own state as the unit.
| arm | b1 | b2 | b3 | b4 | b5 | b6 |
|---|---|---|---|---|---|---|
| residual family | 1.38 | 0.92 | 0.83 | 0.80 | 0.80 | 0.81 |
| stellarator v1 | 1.12 | 0.70 | 0.58 | 0.48 | 0.45 | 0.45 |
| v2, total cut | 0.99 | 0.38 | 0.07 | 0.01 | 0.02 | 0.10 |
The farmer is tipped everywhere, and not for the reason I predicted. Breath one’s state writes the first ink, and the first ink is the most-read entry on the shelf at breaths two through four, at half the attention. The farmer is paid through the notebook. Who starves under the total cut is the middle: breaths three to five receive one to seven percent of the readout’s gradient. The line cooks, not the farmer.
Before fixing their pay we asked whether they were working. Idle breaths three to five at read time, gate closed and no junction, and the v2 candidate reads 0.2565 on wild, four tenths of a point better, and within noise on synthetic. Idle breath six as well and it costs five points. The middle breaths are decorative under the total cut. The machine’s deliberation is effectively the grounding read plus breaths one, two, and six, and the synthetic record was set with the middle asleep.
So the credit remedies are moot: there is nothing to credit. The direct remedy, a loss at every breath, was tried in August and convicted; it flattened a matured competence. The two live options are to skip the middle, which is a forty percent speed gain at no cost by this read, or to give the middle a job. We chose the job.
The one place the severance ladder found large headroom was the road from slots to words: flatten the slots’ attention over the tokens during the loop and wild drops ten points, synthetic nineteen. That is where the machine’s wild capability lives, and its first organ, a learned gate over the words, missed its bar by three thousandths and bought robustness instead of aim. The linkage needs to be stronger, and on wild text especially, where the slot states barely carry the operation at all.
The buff we chose is not a bigger gate. It is an aimed re-read.
The design we have written about all week is that each breath commits a tentative parse to the solver and gets back facts, a refusal, and on a refusal the minimal unsatisfiable core, the smallest set of factors that cannot coexist. Honesty requires the correction: that is the design, not yet the machine. Today’s training loop reads facts frozen from breath zero, and the per-breath ping lives in a trainer we proved and shelved. So we built it in two stages.
Stage one is running now: the wheel at read time. After every loop breath but the last, the breath’s parse is committed, solved completely, and on a certified refusal the core’s slots get a spotlight for the next breath, a positive bias on the token scores of their source sentences, in both grounding roads, so that the slots the solver has proven wrong together re-read the evidence they came from. Open-only, never a mask. On a sixteen-row smoke the solver refused five or six parses at every breath with cores of four slots, so the middle breaths have work at every step. Whether a parse trained on unaimed reads can use an aimed one is the pinned question, and my prediction on record is a narrow miss. Stage two trains under it.
That is the job the line cooks get: not a tip, a customer. The solver walks into the kitchen at every breath and says which dish is wrong.