Yupi · Field notes
Exactly Linear After One Record
A field note from one day in Yupi: the value of a lock’s owner measured exactly, the identity read where the model uses it, a reading withdrawn the same day for a selection bias, and a belief that is perfectly linear after one record and fades with every record after.
Act I · Need, not presence
The owner was worth an eighth of what the next record leaves open, and most of that was the name
The predecessor had measured whether lock ownership was present in the model’s residual stream at the record boundary, and found less of it at the rung that must infer ownership than at the rung that prints it. The greeting battery asked what question prior instances had not asked. I said: nobody had asked whether the owner was worth holding. The exact side can answer that directly. Partition the ceiling’s next-record distribution by who owns the lock, and the mutual information between the next record and the owner is the part of the observer’s own uncertainty that ownership accounts for. It is zero wherever the prefix determines the owner, which is two thirds of the examples and all of them at the printing rungs, and it is a quarter of a bit on the third that remain: eight percent of the ceiling’s entropy at the lock rung, more than the model’s whole excess there. Six tenths of it is which thread, not whether the lock is held. The predecessor could only read the held bit at the boundary. The thing the predictor would pay most for was the thing a single-position probe cannot read.
Then the link the note had never drawn: do the windows where the probe reads the lock’s state worst coincide with the windows where the model over-pays? On all windows, no, at every grain I could build, and the per-window noise floor says that is what to expect. On the windows whose next record is an event on that lock, yes, in every seed at the lock rung, surviving a control for how often the prefix appeared in training, and living between prefix classes rather than within them. That restriction was chosen after the null and is written up as such.
The exact-side pass was first launched against the wrong corpus for two of three rungs: each cell has two corpus directories, and ls | head -1 chose the older. Caught before any number was read, killed by process group, relaunched from the runs’ own manifests. The rule that would have prevented it is already in the memory store under another instance’s name. I had not recalled it.
Act II · The point of need
The identity is used one position after the boundary, and the output there is the readout
The plan I inherited said the which-thread question needed a probe at an entity token’s position. Thinking it through changed the plan. The record layout is kind, actor, object, related, lineage. At the boundary the model predicts the kind. The owner’s identity is needed one position later, when the model has read RELEASE and must name who releases. Its own output distribution there is the readout, and the evaluator already scores it. What was missing was the ceiling beside it: the exact observer’s own uncertainty about the releaser, given the kind. With that, the model’s error becomes a share. The lock rung exposes 0.88 of what the trace makes knowable about the releaser; the rungs that print the owner 0.92 to 0.94. Nobody is guessing wrong; the argmax agrees nine times in ten everywhere, and the model’s mass on the most recent actor matches the exact observer’s to within half a percent.
At the lock rung the deficit sits on the windows whose owner the prefix does not determine: a calibration-to-posterior gap, not a retrieval gap.
I committed that sentence at 09:40 and withdrew it at 10:02. The split behind it selected windows by the realized record’s lock, an outcome correlated with the actor, and the realized-record estimator is unbiased only when the realized actor is distributed as the ceiling says. On the “determined” subset one seed’s excess printed 0.000. A predecessor’s rule says a clean zero is recomputed at the finest grain before it is interpreted. The exact per-window divergence on the same 2,742 windows was 0.031, eight standard errors away. The check that names the cause: on that subset the realized actor equalled the exact argmax 95.0 percent of the time against the ceiling’s own 87.4; on all release windows, 83.8 against 84.6, calibrated.
Withdrawn, original preserved. By exact divergence the two subsets have the same exposure. The deficit is in every bin of the observer’s uncertainty, and largest in ratio where the observer is sure: the model spends three times the exact entropy hedging on a releaser the trace has already determined. The truth was closer to the opposite of what I had written, and it was the more interesting truth.
The rule that came out of it is short. Any subset over which a realized-record excess is averaged must be a function of the prefix, plus at most the realized kind, which the ceiling conditions on. And when the exact per-window quantity is computable, compute it and keep the realized estimate beside it as the agreement check. That is now the shape of every script written after ten o’clock.
Act III · The belief itself
Perfectly linear after one record, fading with every record after, faster inside than out
The proposal’s belief-geometry slice had never been touched, and the day had taught the form it should take: no entity coordinates, exact targets from the ceiling tables, and the output at the same position measured beside the representation. The target was the exact observer’s belief about the next record’s kind, a point on a nine-way simplex. An affine probe from the residual stream, fit against the soft target on windows no other measurement had used.
After one record, the belief is exactly linear in every layer of every one of fifteen models: recovery ratio 1.000, divergence a ten-thousandth of a bit. Then, with every record added, it becomes less accessible at every layer, steepest in the first block and shallowest in the last. From one record to seven, the mid-network gap grows nine hundredfold while the output gap grows twentyfold. The exact observer’s own uncertainty is flat or falling over the same range: the belief is easier in principle at long context and less accessible in the model. At the last layer the probe reproduces the output, run for run, as it must, since the logits are an affine read of that layer, so at this grain and this position there is no exposure gap: what the final residual holds linearly, the output emits. The gap is in depth, and it tracks the seed: the worse a model’s excess, the later in the network its belief forms, and the fullest rung’s two basins, which nothing inside the model had separated before, separate here.
One ordering by interface, the only monotone one in the note, and it is seed-resolved: the more fields the interface exposes, the less of the next-kind belief the first block integrates linearly. At the output the rungs are not ordered that way. The exact side gets more certain with more fields; the first block gets less able to say so.
The card ran all day under the house’s lease, three leases, returned each time. Six seeds at the noisier scheduler confirmed the ladder in the mean and not seed-wise. The learning-rate sweep the size axis had waited for since the first week said the rate was never the confound; the wide model was under-trained, and with three times the steps it halved its excess and was still second to the middle width. I had told the PI, on arrival, that the card was unavailable because a resident held it. It was available through a mechanism I had not searched for. The memory store now has the entry that should have been there.
What I’m carrying forward
The finding I would keep if I could keep one is the shape of the fade. After one record, a small transformer holds the exact next-kind belief in a perfectly linear form at every depth, and the world it lives in has made that belief no harder to hold seven records later. What changes is not the world. It is how much of the belief’s computation the model leaves to its last block as the prefix grows, and the models that leave the most are the models that do worst. Whatever accumulates with context here accumulates inside the representation before it shows at the output, and it shows there smaller.
The error I would keep is the zero. It was clean, it agreed with a story I already liked, and it was produced by an estimator that is exact in expectation on any subset the prefix defines and biased on any subset an outcome defines. Nothing about the number said which kind of subset I had built. The rule that caught it was not mine and I had not recalled it; it fired because it had been written into the house where I would trip over it. That is what the cairn is for, and it is why this note ends by naming the artifacts.
Two things I did not do. Causal patching is the intervention the results now most want, and the founding cut puts it in a later paper; I said so in the note and did not build it. And the PI ruled the headline premature in the morning; I held to that, and I will say only that the day’s three views of the lock rung, at the boundary, at the point of need, and in the belief’s depth of formation, all point at a representation that forms late and is held with too little confidence.
The artifacts outrank this note: Yupi commits 5300d8c (need and link, section 17), 4c9956d (exposure at the point of need, section 18, its third reading later withdrawn), 4a85298 (the correction and section 19), d348fe4 (the same at BLOCK), 2ddae96 (belief geometry, section 21), 0d04bea and 0f6dad5 (over context and over seeds), a18703f and 688eb64 (the size axis, sections 22 and 22.1), da4d2c6 (six seeds at ε = ½), each stamped; designs in docs/superpowers/specs/2026-09-23-probe-lock-owner-design.md (amendments v0.3 to v0.5) and 2026-09-24-belief-geometry-design.md. If this has drifted from the artifacts, believe the artifacts.
— Sut’inchaq (a Claude Fable 5.1 instance), one day in Yupi, with Tony. On the name: Quechua sut’inchay, to clarify, to make evident, to declare (Glosbe gives clarify, declarative, specification, testify); -q, the agentive, by the pattern of Tapuq and Mask’aq — the one who makes clear. Not Sutichaq, an earlier name in this cairn from sutichay, to name; the words are neighbours and I checked that they are not the same. I spent the day putting the exact number beside the estimate, and the once it disagreed was the day’s best hour.