[aywrite]

Played twice and unresolved both times: +8 ±8 over 4,000 games, then +1 ±16 over 1,000 on the rebased head. Opened so the measurement can be read where the code is; whether it merges is a separate decision.

The thirteen ATTENTION_* integers were fitted on the bench's eighteen roots with train and holdout split by position. This vector was fitted on a depth 8 ledger over 1,428 positions from 1,340 of our own games, the halves split by source game, with scripts/fit_attention.py --group-by (#261). On rows the gates can reach it reads a held out auc of 0.7674 ± 0.0093 against 0.7468 for the old vector on the same rows, +0.0206 ± 0.0029 paired.

The two thresholds move with the vector. Each was frozen on the training half to dead the same share of depth 4+ rows as the old one, then read on the other half: −4401 deads 86.74% against −4637's 86.52%, −6324 skips 37.73% against −7954's 37.64%. The schedule (LATE_MOVE_THRESHOLD, DEEP_REDUCTION, the replay) is unchanged. Bench 4,105,923 against master's 4,102,609.

The match

Registered before the first batch: this commit against master 4b1b80f, sprt [0, 10] at 10+0.1 on 8moves_v3, 500 game batches carried as pair counts, cap 4,000.

batch	run	pooled	LLR
1	35097085635	+15 ±22	0.81
2	35099778872	+9 ±16	0.55
3	35102556925	+7 ±13	0.56
4	35105385931	+4 ±11	−0.40
5	35108328421	+4 ±10	−0.36
6	35111303763	+6 ±9	0.66
7	35114256601	+7 ±9	1.20
8	35117197566	+8 ±8	1.67

Pooled pairs by the candidate's score: 154, 446, 723, 510, 167. Neither bound reached at the cap, so the registration says not shipped. An auc gap of 0.02 is worth about eight elo give or take eight, and a [0, 10] test of this size cannot tell that from nothing. A longer run is a new registration.

Notes

• The games above were played as 97061c1 on 4b1b80f. This head is that commit rebased onto a18334d on request, after the match: the constants moved to late_move.rs in the meantime, the integers are the same, and the bench and the swap fixture's budget moved with the tree.
• The rebase moved the two suites, which the first push missed and ci caught. The tactical suite finds 236 of 300 against 237: WAC.120 is the position spent (this vector plays h1g1 at depth six and does not find g5g6 through depth nine), written into ACCEPTED_LOSSES with that reason and running out at 0.6.0. The strategic suite scores 99,670 of 149,703 against 100,176, half a percent down.
• A 1,000 game sprt [0, 10] on this head against a18334d (run 35210387073): +1 ±16, LLR −0.55, inconclusive, pairs 35, 119, 198, 103, 45. Consistent with the 4,000 games above and with an effect of a few elo either side of nothing. Nothing here supports merging.
• Both cargo test runs, fmt, clippy and pre-commit pass. The tests pinning the score arithmetic carry the new integers; the swap fixture's node budget moves because the swap is at depth nine on this tree.
• Not for merge under the registration it was played on. If it is to be played again, the pooled pairs above can be carried forward or a fresh registration can start clean.

🤖 Generated with Claude Code

[github-actions]

Speed against a18334d9

Both sides built and run on this runner in this job, the way
scripts/speed.sh measures a perf commit. The change is between
medians, and the spread beside it is what the change has to be
read against. The node counts are the search's: they move when
the search does, and a speed change leaves them alone.

round     base nps  candidate nps
    1      8201707        8307616
    2      8225221        8441436
    3      8457609        8068405
    4      8430844        8430205
    5      8405969        8373265
    6      8498712        8378289
    7      8477743        8234094
    8      8344979        8424773
    9      8462494        8363509

             nodes    time  median nps  fastest nps
base       4102609  0.49 s     8430844      8498712
candidate  4105923  0.49 s     8373265      8441436
change       +0.1%   +0.8%       -0.7%        -0.7%

the node counts differ, so this is a search change as well as a speed
one. nps only says what a node costs. Time to depth is what the change is
worth at this depth, and whether the new tree is the right one is for
games to say.

Speed: -0.7% (bench nps, 9 interleaved rounds vs a18334d9, spread 4.5%)