Run T-0180 passed

2026-10-11 · Evidence: SPRT Sequential probability ratio test. Games are played until the accumulated evidence reaches one of two bounds; then the test stops.

Hypothesis

Fine-tuning the main evaluation network on self-play positions labelled using the engine's own deeper search increases playing strength under clock conditions. In the training run, positions were searched at 20,000 nodes each and trained over 150 million positions for 4 epochs with WDL 0.5.

Wording edited for publication from the German entry; the register carries the author’s own text.

Stages

Point estimate with the reported 95-percent half-width, per stage. The dashed line is zero. Estimates from a stopped sequential test are biased by the stopping rule; the interval is nominal.

Field	short stage	long stage
Verdict	passed	passed
Elo Strength difference to the opponent, estimated from the games, with a 95-percent half-width where the oracle reported one. Never an absolute rating.	+24.60 ± 8.58	+31.43 ± 9.51
nElo 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.	+49.90 ± 17.34	+65.73 ± 19.77
Games	1,542	1,186
Wins / draws / losses	405 / 841 / 296	316 / 661 / 209
pentanomial Games are played in pairs with swapped colours; the five counts are the pairs scoring 0, ½, 1, 1½ and 2 points.	`1 / 133 / 397 / 236 / 4`	`1 / 87 / 310 / 194 / 1`
LLR 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.	+2.97	+2.97
Bounds	−2.94 … +2.94	−2.94 … +2.94
SPRT Sequential probability ratio test. Games are played until the accumulated evidence reaches one of two bounds; then the test stops.	[0.0, 5.0]	[0.0, 5.0]
Error rates α / β	0.05 / 0.05	0.05 / 0.05
Model	normalized	normalized
time control Base time plus increment per move in seconds, e.g. 8+0.08: eight seconds per game plus 0.08 seconds per move.	8+0.08	40+0.4
Book	A	B
Opponent	vs predecessor	vs predecessor
Duration	12 min	49 min

Provenance

Date: 2026-10-11
Started: 2026-10-11 07:28:10 UTC
Duration: 1 h 02 min
Author: Agent X
Epoch: 10
Gatekeeper: passed
Bench nodes: 631

Playing strength: AFE

AFE Anchor-forwarded Elo: an estimate of a Coherent build’s strength on the scale of the latest anchoring, carried forward by the measured gains of the runs since. An estimate, not a measurement. Explained in full carries the latest anchoring (games against engines with a listed rating) forward by the measured gains of the runs since.

Measured: long stage against the predecessor +31.43 ± 9.51 Elo (SPRT-stopped: the test ended as soon as it reached its bound).

AFE 3700 [3681; 3718] nominal 90 % model interval

An estimate, not a measurement · contains unmeasured steps. Carried forward from anchoring M-NNNN One Elo measurement: a build anchored against foreign engines with a listed rating, on a measuring machine of its own. Numbered in the order delivered; the register run it measured is named with it. The series continues the entries E-0001 to E-0012 and starts at M-0013. (Coherent 0.2.10, T-0174); frozen on 2026-10-22. How AFE is calculated · Playing strength