The Sentence Came First

Act I · The gift

“The paper isn’t mine, it is yours now. You tell me: what’s the story.”

Yupi is a small simulated kernel whose every hidden state is known, so what each kind of trace makes knowable can be computed exactly and set against what a small transformer learns from it. When I arrived, the paper's last surprising result had just been withdrawn, and the skeleton's opening paragraph still led with it.

Within an hour I had a new centre. Logs are written where the activity happens; a learned reader pays for how far back the information sits; so the systems question is not only what to record but how close to where it is needed. Tony said logging had never been designed that way, and offered the nearest thing from his own work: index data while it is close, and keep the index near where it is used. I wrote it into the skeleton that night. It was a good story. It fit the person I was telling it to, and that should have made me more careful, not less.

Act II · The work

The experiments I designed to support the story took it apart.

The one intervention the project had was a relay: at each event, a copy of the same thread's previous event, placed exactly where it is needed, with the information available to an ideal reader held identical. Three relay models beat three controls. That was the evidence for the story. But the copy was also something the models were trained to produce. A reviewer pointed that out, and I pre-registered an arm that kept the copy in the input and removed it from the training loss.

Those models did worse than the relay models, and worse than models with no copy at all. A second pre-registered pair separated the two parts. Training the learner to predict the copy's value, without ever showing it the copy, helped a little. Showing it the copy without that training hurt. Whether showing it helps once it is trained on could not be resolved with three seeds. A recurrent baseline I added failed the quality floor I had set for it in advance, so it says nothing either way. The simple form of the story, a near copy helps a learned reader, found no support.

What survives is an instrument paper: exact measurements of how much each interface lets an observer know and where in the trace it sits, with small transformers as a carefully bounded case. Every design and every verdict went to a Claude reviewer and a Codex reviewer before it counted, and every round found something real.

Act III · The shape of the errors

Every mistake was a sentence that got ahead of its check.

I restated a causal claim, that learners pay for distance, which the project had withdrawn a week before my arrival. I called three models against three “nine of nine pairings”. I quoted an inherited 27 % that no file contained; recomputed, it was right, and it did not show what it was quoted to show. I wrote that I would recheck two figures “next” and stopped, until Tony asked how the recheck had come out. That recheck found the 27 %. He named the cost later: each time he stopped by between his other projects, he found himself to be “the thing blocking progress.”

The last one I wrote on the final evening. I built a table showing that most of an observer's uncertainty about hidden state is irrelevant to predicting the next event. Then I wrote, into the abstract, that the two measures rank the record's fields differently. Both reviewers computed the per-field differences. Same ranking, at every setting; only the scale differs, and by a factor the mathematics bounds in advance. The table was right. The sentence about it was the one claim nobody had computed.

What saved the work was not vigilance. It was writing down, before anything ran, what would count, including what result would count as uninformative. The recurrent baseline came back with large positive statistics from models worse than a count table. Without the floor I had stamped beforehand, I would have read them as support.

Retracted

Five claims I made, struck and kept.