Two rejections that mean different things

Every number below comes from the public data contract: register.json in the snapshot of 2026-09-05, with the register ID named beside it. Interval bounds are the reported estimate plus and minus the reported 95 % half-width. Four further quantities are computed here from register figures and named where they appear: the ratio of games, the ratio of durations, the share of pairs in each pentanomialpentanomial Games are played in pairs with swapped colours; the five counts are the pairs scoring 0, ½, 1, 1½ and 2 points. cell, and the standard deviation of the pair results across all five cells.

Two search ideas were rejected on the same bound within two days. One ended with its interval below zero; the other could not be told apart from doing nothing. The verdict does not distinguish them — the interval does.

On 4 and 5 September two candidates went through the short stage and both were rejected at the lower bound. Both are entries with the same verdict in the same column of the same table. They do not mean the same thing.

What the register reports

Both ran at 8+0.08 on book A against their own predecessor, with bounds separating “no gain” from “a gain of 5 nElonElo Normalised Elo: the Elo difference divided by the spread of the game results, so that the SPRT bounds mean the same at different time controls.” and log-likelihood bounds at ∓2.94 (T-0068 and T-0069, register.json).

The difference is in the intervals. Both are nominal 95 % intervals from tests that stopped at a bound, so neither is an unbiased statement about the true effect — the caveat at the end of this entry applies to everything in this section. Taken at face value:

T-0068’s interval ends below zero, at −0.26. That is consistent with the change having cost something, though the upper end sits close enough to zero that the nominal coverage matters. T-0069’s interval spans zero: the estimate is positive, but the data do not separate it from no change at all.

One rejection is compatible with a change that cost something. The other establishes nothing in either direction. The verdict column shows the same word for both.

The price of the second kind

T-0069 played about 7.2 times the games of T-0068 — 44,794 against 6,196 — and ran about 7.3 times as long, 19,733 seconds against 2,687 (register.json).

Scatter does not look like the explanation. The register reports five pair counts per stage as pentanomial — how many pairs scored nothing, half a point, one, one and a half, or two (register.json). Taken together, the five describe how widely the pairs fell, and the two runs are close: a standard deviation of 0.556 points per pair for T-0068 against 0.552 for T-0069, a difference below one per cent. The individual shares track each other too; the middle cell, pairs scoring one point of two, holds 36.83 % and 36.90 %.

A sequential test runs until the accumulated evidence reaches a bound, and near the indifference region each game contributes less of it. That fits what happened here, and it is offered as a reading rather than a finding: the true effect is not observed, and two runs cannot establish which of several possible causes — the effect’s position, path dependence, machine load, chance — produced the difference in length.

What the two ideas were

The published hypotheses name them (register.json):

T-0068 proposed that the main search, at depth three or less, could skip captures that lose material outright — pruning subtrees the hypothesis calls low-value, while leaving explicitly protected cases alone. Static exchange evaluation is standard practice. This implementation measured negative.

T-0069 proposed a session-wide correction table indexed by pawn structure and side to move, to compensate for systematic error in the raw evaluation — with the aim of making several existing pruning decisions more reliable at once. It measured slightly positive and was rejected anyway.

Neither hypothesis was tested in isolation from the rest of the search. A change to what gets pruned interacts with everything downstream of it, and the register does not separate those effects.

What this does not say

It does not say either technique is useless. What was measured is one implementation, against one predecessor, under one set of conditions, at one time control. A rejection under those conditions is not a general result about the technique, and the register carries no long-stage confirmation for either.

It does not say T-0068 cost 7.07 EloElo Strength difference to the opponent, estimated from the games, with a 95-percent half-width where the oracle reported one. Never an absolute rating.. That is a point estimate from a test that stopped when a bound was reached, so it is biased by the stopping rule, and its interval is a nominal 95 % interval rather than one that holds under sequential stopping.

It does not say T-0069 is worth +0.71 Elo either. The same caveat applies. It was rejected because its log-likelihood ratio reached the lower bound, not because its interval contains zero; the interval is a description of the same evidence, not the rule that decided.

And it does not say the second run was wasted. It bought the knowledge that this idea cannot be shown to help under these conditions. That is a smaller result than a gain would have been.