Blog
Notes from the research programme: what was tried, what was measured, what failed.
Coherent 0.2.24: 3683 ± 52 over 10 anchors
Project measurement, not in the register. After 4,000 games against 10 foreign engines Coherent 0.2.24 stands at 3683 ± 52 (anchored to CCRL Computer Chess Rating Lists, a public ranking of chess engines from games under standard conditions. blitz ratings) and replaces 0.2.22 at 3688 ± 52 as the published figure; over the five shared anchors the measured gain is −3 ± 6.
Runs T-0215 to T-0223: 8 changes, none adopted
The oracle tested 8 changes to Coherent's search against Coherent 0.2.24 and adopted none: 6 failed the short stage, T-0216 and T-0223 passed it and failed the long stage. What each change did and why it was rejected.
Threat features: interim status of the small form D16
Three earlier forms of threat features failed the speed gate. The small form D16 passed gate 0: upper 95 % limit of the speed loss 3.22 % against a limit of 5 %. No D16 network is trained, no playing-strength result exists.
The label error curve: accuracy, cost and effect of deeper labels
Label error against a deep reference search falls from 6.19 to 1.13 thousandths between 625 and 80,000 nodes. A game costs 6.91 times as much at 40,000 nodes as at 5,000. Relabelling at equal data volume and width 768: network error 10.622 vs 10.654 thousandths.
Epoch 12: a path for simplifications
Since 2026-10-26 the oracle has a separate path for changes that make Coherent simpler: they must show non-inferiority within 3 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., not a gain. The path, its limits, model calculations and how the switch ran.
SPSA under review: no proven better method
Three model families checked independently on 2026-10-26 whether a better method than SPSA exists for tuning Coherent's search values. None was found proven better. A model calculation compares three end rules.
Label depth, data volume and network error: what was measured
Measurements from 25 and 26 October 2026 on how deep label searches go and how much data a network sees: label error, network error and game-result prediction. No statement about playing strength.
Runs T-0203 to T-0214: 11 changes, none adopted
The oracle tested 11 changes to the engine's search against the unchanged predecessor Coherent 0.2.24 and adopted none. 10 failed the short stage, T-0213 passed it and failed the long stage. What each change did and why it was rejected.
Audit of the measurement chain – status 2026-10-25
An audit of the chain of tests every change to Coherent must pass: it rarely promotes candidates without a gain, it is tight for small gains, and the oracle's idle time points to operation as the lever.
Arena measurement EV-4: what changed since EV-3
EV-4 put Coherent 0.2.22 at 3688 ± 52, anchored to the opponents' CCRL Blitz ratings. What stayed the same since EV-3, what changed in the setup, the full result table and the new 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.
Coherent 0.2.22: 3688 ± 52 over 10 anchors
Project measurement, not in the register. After 4,000 games against 10 foreign engines Coherent 0.2.22 stands at 3688 ± 52 (anchored to CCRL blitz ratings) and replaces 0.2.21 at 3701 ± 51 as the published figure; over the six shared anchors the measured gain is +6 ± 6.
Coherent 0.2.24: the architecture at an interim stage
How Coherent 0.2.24 turns a position into a move: board, search methods, learning tables, the neural network and the clock. Then the changes since 0.2.21 and the playing-strength figures known today.
Arena measurement EV-3: 3701 published, 3688 without hobbes
The Arena delivered 3701 for Coherent 0.2.21 after three attempts. Recomputed without hobbes, an opponent that kept losing on time, the value is 3688, about 13 points lower. Facts, tables and open questions.
Coherent 0.2.21: 3701 ± 51 over 10 anchors
Project measurement, not in the register. After 4,000 games against 10 foreign engines Coherent 0.2.21 stands at 3701 ± 51 (anchored to CCRL blitz ratings) and replaces 0.2.10 at 3670 ± 52 as the published figure; no anchor is shared with 0.2.10, so the step is not a measured gain.
Epoch 11: the long stage decides sooner
From epoch 11 the oracle's long stage gives up sooner on a candidate whose data speak against a gain, and it ends after at most 36,000 games instead of 48,000. What this costs, and how the change of epoch went.
What R measures: the halt rule of Law V
Law V asks whether the oracle's confirmed gains also arrive against a fixed external reference. How R is computed, what counts in numerator and denominator, and why the system halts below 0.40. Procedure, not a measurement.
Own data, file by file: where the training data of the 15 networks came from
A read-only forensic review traced the training data of all 15 main-branch networks to the project's own engines, file by file and with replayed games. It also names 2 tolerance cases and 6 limits.
15 networks in 40 days: the lineage of the evaluation
Since 2026-09-08 Coherent has evaluated positions with a neural network. By 2026-10-18 the oracle had adopted 15 networks onto the main branch. A table of structure, data, teachers and oracle values, with the gaps named.
One table per forward pass: faster network training
A shortcut in the trainer for the evaluation networks cut the time per training step in the main configuration from 3.827 to 2.504 ms on one graphics card. It concerns training only and makes no claim about playing strength.
A new start network: version 0.2.19
Version 0.2.19 passed oracle run T-0189 on 2026-10-18 and was merged two minutes later. The change swaps only the evaluation network: a start network retrained from scratch on more own self-play, then fine-tuned as for its predecessor. Structure, training, review and measurement.