Runs T-0248 to T-0251: a smaller margin passes, the node count reaches the main branch

On 09.11.2026 and 10.11.2026 the oracle recorded the entries T-0248 to T-0251, all in epoch 16. Two changes to the search were rejected. A change to the reverse futility cut passed both stages, and a correction brought the line with the node count before bestmove into the main branch. The stages and decision rules are described in the post "Epoch 14: node count before bestmove and a rule for tempo bundles"; the post "Epoch 16: the network class is withdrawn and a run is cancelled" records that they still apply unchanged.

The runs at a glance

The post "Run T-0247: a network with threat features passes" explains the ± values. The register predecessor was the engine of T-0247 for T-0248 and T-0249, and that of T-0249 for T-0250 and T-0251. The rejected candidates carried the same version numbers as the passed ones: T-0248, like T-0249, carried 0.2.28, and T-0250, like T-0251, carried 0.2.29.

T-0249: a smaller margin

The reverse futility cut and its margin are described in the post "Runs T-0224 to T-0236: a bundle of 6 changes passes", in the section on run T-0224. The cut applies at remaining depths up to 8. Until now the base margin grew linearly, by 54 per ply of remaining depth. T-0249 replaces it with ⌊54 · t · (t + 8) / (2 · 8)⌋, where t is the remaining depth. At remaining depth 8 the base margin stays at 432; at every smaller remaining depth it is smaller, at remaining depth 1 for example 30 instead of 54. The deductions for a rising evaluation and for an opponent who harmed themselves with their last move are subtracted as before, and the margin never drops below zero. With the same remaining depth, deductions and other cut conditions, the new margin permits every cut the old margin permitted. Additional cuts are possible where the final margin is strictly smaller; at remaining depth 8, and when both margins are clamped to zero, the permitted cuts are unchanged.

The bench figure in the register drops from 713 for the predecessor to 635. Both stages passed. In the long stage the recomputed LLRLLR Log-likelihood ratio, the running evidence of an SPRT. It starts at 0; a stage passes at the upper bound and fails at the lower one. was 2.710, just above the upper bound of 2.708; rounded to two places both are 2.71. Since 10.11.2026, T-0249 has been in the main branch.

Two rejections

T-0248: malus for extended quiet moves

The search keeps history values by which it orders and reduces quiet moves. When a node ends with a best move, that move receives a bonus; if it is quiet, the other quiet moves searched at the node receive a malus that decays for moves searched later. When a node ends without a best move, this learning at the end of the node is skipped. In principal-variation nodes, T-0248 additionally gave every quiet move that the search had extended (singular extension of the table move or extension of a check) the full malus that depends on the remaining depth at the end of the node, provided it was not the best move. This also applied when the node ended without a best move. The short stage rejected.

T-0250: correction learning with only one legal move

At the end of a node, under certain conditions, the search learns a correction of its static evaluation from the search result. T-0250 additionally skipped this learning when the node ended without a beta cutoff and had exactly one legal move. Nodes with a beta cutoff and nodes with at least 2 legal moves retained the existing learning conditions; exclusion searches remained excluded from correction learning. The short stage rejected; the 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. estimate was close to zero.

T-0251: the node count in the main branch

T-0251 adds to the engine of T-0249 the same 2 lines that T-0238 had introduced, and raises the version to 0.2.29. The line, its purpose and the correction route with differential gate and fixing record are described in the post "Epoch 14: node count before bestmove and a rule for tempo bundles". Like T-0238, T-0251 went this route without games and passed. The bench figure stays at 635.

The build of T-0238 was the register predecessor of T-0247. The post "Run T-0247: a network with threat features passes" describes the base of T-0247 and the missing line. All candidates from T-0247 to T-0250 lacked it. The commit of T-0238 never reached the main branch; the line first appears there with T-0251.

State of the register

The export of 10.11.2026 contains 251 entries. Since 10.11.2026 the engine in the main branch has been that of T-0251, version 0.2.29.

What this post does not say

The register does not break down the result of T-0249 by remaining depth.

How this was checked

• Sources: oracle register, export of 10.11.2026, entries T-0238 and T-0247 to T-0251; the code changes of the four candidates against their register predecessor, with excerpts of the affected functions; the commit history of the main branch; the pipeline events.
• Recomputation: a script recomputes Elo, nElonElo Normalised Elo: relates the mean score excess of a colour-swapped pair to the spread of the pair results. and LLR of all stages from the pair counts, as well as the margins per remaining depth and the durations, and checks the numbers in the text in both languages.
• Independent check: text, metadata, figure and both language versions were checked independently against the same sources.