Caught before it counted

"The author does not accept the work" described how work is checked in this project: the criterion is written down first, and somebody other than the author checks it. This entry shows what that looked like on a single day, 2026-09-29: five cases, each found before it reached a decision or a published number.

A draw that was not a draw

The endgame battery described in "Bishop and knight" first ended a game as a draw as soon as a threefold repetition could be claimed. That included positions where the winning side was to move, a side that would never claim it. It also counted a game stopped at its length limit as a draw, when the honest label is "unknown".

A review of the code by a fresh session that had not written it found both problems, along with eight other findings, before the tool's first report. That session came from the same model family as the author, which "The author does not accept the work" does not count as a second opinion; it was still a check by a session that had not written the code. After the fixes a second review released the tool. The figures in the other entry come from the corrected version, which then received four small changes that were not reviewed again.

A comparison of the wrong things

An analysis concluded that running two threads per processor core added practically nothing for generating training data. Its evidence was that a machine with one thread per core produced about as much as a machine with two. But the comparison mixed runs of different versions of the engine.

Compared on the same version, two threads per core produced about a third more. A controlled test on a single machine type showed a gain of a little under thirty percent. The conclusion was withdrawn before it could inform a hardware decision.

A rule that no longer applied

A recommendation said that a test stopped at its game limit may be resubmitted once. That had been the rule, but an appendix to the same decision had replaced it almost four weeks earlier: the same build is never measured twice.

The reviews of the recommendation itself did not catch this. A review of a different proposal on the same day did, and cited the appendix.

A plan that promised too much

A proposal claimed that a single change to the measuring setup would bring three gains at once. Two independent reviews, each by a model from a different family than the author's, rejected the conclusion. The measurements in the proposal largely held up. But two of the three gains did not follow: one was blocked by a measuring rule rather than by the missing change, and another needed more than the proposed change. The proposal was restated to say only what the data supports.

The program that checks the quotes

A research task that day had to cite its sources with verbatim quotes, and a small program was written to fetch every cited page and search it for the quote. Before its first use, a reviewer found five serious problems: four ways to fool it and one way for a correct quote to be rejected. For example, a "quote" made only of ellipsis marks would have passed. All five were fixed, and the program was run against deliberately false quotes, which it rejected, before any research ran through it.

What the pattern is

None of the five cases was found by its author. Each was found by a review done by somebody who had not done the work, and one of them not by the review meant for it, but by another.

The price is time and a longer record. In these five cases, none of the errors reached a published number or a decision.