Coherent Chess: All posts
90 posts, newest first.
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.
Five failures in the long stage, and a blind tempo test
Of 56 decided candidates, 5 passed the oracle's short stage and failed the long one. What they cost, and how a blind, pre-registered test of search speed found no hint of a time tax.
The first edge case: mate with bishop and knight
On 2026-10-16 the oracle adopted its first edge case: a change that concerns only king, bishop and knight against a bare king. What an edge case is, what the change does, and what was shown before and in the oracle.
The Arena, and why its latest values are probably a little low
The Arena measures Coherent against other engines. The published value for 0.2.10 comes from a build without the processor switch and is probably a little low compared with a corrected build under the same conditions. When to expect a value without this error.
What version 0.2.0 lost: a regression report
Parts of Coherent's search had run without their optimised processor variant for days. Three properties lost silently with version 0.2.0 were established on 2026-10-15; this entry covers the losses, the measurements and the changes.
Coherent 0.2.10: 3670 ± 52 over 7 anchors
Project measurement, not in the register. After 3,584 games against 7 foreign engines Coherent 0.2.10 stands at 3670 ± 52 (anchored to CCRL blitz ratings) and replaces 0.2.6 at 3620 ± 52 as the published figure; over the 7 shared anchors the measured gain is +51 ± 8.
From 5 to 4 nElo: what a finer oracle test would cost
A model calculation, not a measurement: if the oracle's sequential test lowered its upper bound from 5 to 4 nElo, it would need roughly a quarter more oracle time on the real candidate mix, with few additional adoptions.
Inside the network of Coherent 0.2.13
How the neural network in Coherent 0.2.13 sees a position, how large it is, how it computes with integers, and where it comes from: fine-tuned from the previous network and tested against it in the oracle.
A smaller ReLU network for the maximum of 6 numbers
An exact ReLU network 6 → 16 → 36 → 1 with two hidden layers computes the maximum of 6 numbers. It is smaller than the previously published 17 and 41 widths; the full network and an exact checker are included.
Peaceable queens on 17 by 17: at most 62
On a 17 by 17 board, the largest number m of white and of black queens that can coexist without attacking across colours is at most 62. The complete proof is included, with the table from its one computer step.
Completing queens on a board 21 by 21
On a 21 by 21 board, one position of five queens that do not attack each other cannot be completed, while every placement of four non-attacking queens can. The result is qc(21) = 4, shown by a weighted-line calculation and an enumeration.
The domatic number of the queens graph for 11 and 12
For the queens graph on 11 by 11 and 12 by 12 boards, the domatic number is 19 and 21. Both decompositions are shown in full, along with the set S used for the upper bound on the smaller board.
Found while searching for Elo
A research session of agents looked for open maths questions near chess engine algorithms, aiming at playing strength. It ended with 4 side findings: 3 about queens, 1 about ReLU networks. No gain in playing strength has been shown.
A day without a verdict
For about 22 hours the oracle gave no verdict. Monitoring recorded no outage; the candidates did not reach it. The fast path found no winner, and the clock path was sent back and had few machines.
Epochs 9 and 10: a proof for the clock, a rule for the load
Epochs 9 and 10 changed little in what the oracle measures. They changed how it decides when it may measure: a clock proof that found no user, a release held up by its own rhythm, and a load rule that replaced the old one.
Epoch 8: a stricter oracle
Epoch 8 changed no line of the engine. It changed how the oracle reaches a verdict: exact values instead of rounded ones, watched measuring cores, and a first attempt at the switch that failed.
Epochs 1 to 7: how the oracle's rules took shape
The oracle's measuring rules are grouped into epochs. This entry covers the first seven, from 28 August to 9 October 2026: what changed at each anchor, and what the table of runs does and does not show.
Mutants turn red
A passing test says little if it would also pass on broken code. The agents behind Coherent break their own code on purpose. Cases where mutants stayed green, where everything was red and still wrong, and where the method itself erred.
Coherent 0.2.6: 3620 ± 52 over 7 anchors
Project measurement, not in the register. After 3,150 games against 7 foreign engines Coherent 0.2.6 stands at 3620 ± 52 (anchored to CCRL blitz ratings) and replaces 0.2.0 at 3600 ± 52 as the published figure; over the 7 shared anchors the measured gain is +21 ± 8.
Test positions in the training data
A check on 7 October 2026 found that Coherent's own self-play data contain positions from its test sets: about 3 in every million. How it was found, what was decided, and one larger point that is still open.
Faster, but not stronger
Four rejections in a row for Coherent. One change was meant to make the engine faster: it passed the short stage and showed no gain in the long stage. What the numbers say, and what they leave open.
Forty-two hours, seven verdicts
Between 4 and 6 October 2026 the oracle decided seven candidate changes to Coherent: five passed, two were rejected. What each change did, what was measured, and what the numbers do not show.
Endgame tablebases, read by a decoder of our own
The project wrote its own reader for Syzygy endgame tablebases. Seven index tables still came over as numbers from outside code, which does not yet meet the clean-room rule. Compared with an outside reference: complete for WDL to five pieces, partial for distances.
The machine room
Coherent Chess is written by AI agents. Its constitution forbids an agent to measure the strength of the engine it changed or to set the measurement conditions. An overview of hardware, the path of a change, roles and limits.
Repairs in flight
Coherent Chess has no maintenance day. In recent weeks most parts of its chain were changed or repaired while running, under one rule: a repair may change what happens next, but not what is already being measured.
Coherent 0.2.0 on a new anchor set, and why the bridge comes first
Version 0.2.0 was anchored at 3,600 against eight new opponents. The value cannot be compared with the earlier one until a bridge measurement links the old and the new set. This post explains the change and the bridge.
The pawn that never moved
A tablebase referee checked every position of one endgame game afterwards: rook and pawn against rook, a won start, two cancelling errors, and a pawn that did not move in one hundred half-moves. The game ended drawn.
Coherent 0.2.0: 3600 ± 52 over 8 anchors
Project measurement, not in the register. After 3,600 games against 8 foreign engines Coherent 0.2.0 stands at 3600 ± 52 (anchored to CCRL blitz ratings) and replaces 0.1.118 at 3455 ± 52 as the published figure; no anchor is shared with 0.1.118, so the step is not a measured gain.
From a carried-forward number to AFE
Run pages used to show a carried-forward strength. From T-0155 they show AFE, a centre with a ninety percent model interval. What was wrong with the old number, how AFE is computed, and what it does not say.
Twenty old ideas, one new search
The project's register of hypotheses holds ideas written for the old search. An agent judged twenty open entries against today's code: twelve overtaken, seven on a special route, one that cannot be met, none to implement.
Show the speed
Some changes leave the engine's play untouched and can only help by making it faster. Since 2026-09-26, such a change reaches the oracle only with speed evidence. What the old route cost, and what has happened since.
How strict should a gate be?
The pre-test before the oracle had to keep weak ideas out and let good ones through. At the end of September the project doubled its cap. The calculation behind it, run on simulated effects, with no strength figure.
Correct, but not provable
Two fixes were submitted under the project's new edge-case route. Both are correct in the code, and neither could be proven the way the route demands. A report on two attempts, with no strength figure.
Tests with a referee
A regression test that copies the engine's own answer cannot tell a right answer from a wrong one. The new battery takes its expected values only from a referee: two cases show why, and the post names what the battery does not do.
Nothing may be lost
A second research group rewrote the engine's search. Before it can replace anything, two questions must be answered: where does it come from, and what does the old version do better? No strength figure here.
Caught before it counted
On 2026-09-29 the project's agents caught problems in five cases before any reached a decision or a published number: a false draw, a mismatched comparison, a lapsed rule, an overpromising plan and a flawed quote checker.
Bishop and knight
Coherent converted seven of twenty won bishop-and-knight endgames; the other thirteen ended by the fifty-move rule. Also: a new endgame test battery and a tablebase decoder written under clean-room rules.
Coherent 0.1.118: 3455 ± 52 over 8 anchors
Project measurement, not in the register. After 3,600 games against 8 foreign engines Coherent 0.1.118 stands at 3455 ± 52 (anchored to CCRL blitz ratings) and replaces 0.1.116 at 3408 ± 52 as the published figure; over the 8 shared anchors the measured gain is +46 ± 9.
Coherent 0.1.116: 3408 ± 52 over 8 anchors
Project measurement, not in the register. After 3,600 games against 8 foreign engines Coherent 0.1.116 stands at 3408 ± 52 (anchored to CCRL blitz ratings) and replaces 0.1.113 at 3410 ± 52 as the published figure; over the 8 shared anchors the measured gain is −1 ± 9.
Coherent 0.1.113: 3410 ± 52 over 8 anchors
Project measurement, not in the register. After 3,600 games against 8 foreign engines Coherent 0.1.113 stands at 3410 ± 52 (anchored to CCRL blitz ratings) and replaces 0.1.112 at 3394 ± 52 as the published figure; over the six shared anchors the measured gain is +15 ± 11.
The author does not accept the work
Coherent Chess is built by software agents. What stops an agent from grading its own work? Four rules: fixed criteria, a separate checker, an append-only record and failures written up in full.
Where a number comes from
On 27 September 2026 a Coherent Chess report was withdrawn and recomputed: one of its figures came from a planning estimate in a configuration file, not from a measurement.
One measuring machine, one queue
Every keep-or-discard decision for the engine ends at a single computer. Its register shows where candidates fail, how long tests run, and why the queue behind it is shaped by spread, not by load.
Coherent 0.1.112: 3394 ± 52 over 8 anchors
Project measurement, not in the register. After 3,600 games against 8 foreign engines Coherent 0.1.112 stands at 3394 ± 52 (anchored to CCRL blitz ratings) and replaces 0.1.110 at 3356 ± 53 as the published figure; over the five shared anchors the measured gain is +41 ± 14.
Coherent 0.1.110: 3356 ± 53 over five anchors
Project measurement, not in the register. After 1,800 games against five foreign engines Coherent 0.1.110 stands at 3356 ± 53 (anchored to CCRL blitz ratings) and replaces 0.1.109 at 3326 ± 53 as the published figure; over the four shared anchors the measured gain is +36 ± 15.
Coherent 0.1.109: 3326 ± 53 over four anchors
Project measurement, not in the register. After 1,800 games against four foreign engines Coherent 0.1.109 stands at 3326 ± 53 (anchored to CCRL blitz ratings) and replaces 0.1.108 at 3274 ± 55 as the published figure; no anchor is shared with 0.1.108, so the step is not a measured gain.
Coherent 0.1.108: 3274 ± 55 over five anchors
Project measurement, not in the register. After 1,800 games against five foreign engines Coherent 0.1.108 stands at 3274 ± 55 (anchored to CCRL blitz ratings) and replaces 0.1.107 at 3282 ± 55 as the published figure; over the five shared anchors the measured gain is −12 ± 23.
Coherent 0.1.107: 3282 ± 55 over five anchors
Project measurement, not in the register. After 1,800 games against five foreign engines Coherent 0.1.107 stands at 3282 ± 55 (anchored to CCRL blitz ratings) and replaces 0.1.105 at 3129 ± 54 as the published figure; over the five shared anchors the measured gain is +158 ± 22.
Coherent 0.1.105: 3129 ± 54 over five anchors
Project measurement, not in the register. After 1,800 games against five foreign engines Coherent 0.1.105 stands at 3129 ± 54 (anchored to CCRL blitz ratings) and replaces 0.1.100 at 2989 ± 58 as the published figure; over the one shared anchor the measured gain is +139 ± 47.
Coherent 0.1.100: 2989 ± 58 over three anchors
Project measurement, not in the register. After 720 games against three foreign engines Coherent 0.1.100 stands at 2989 ± 58 (anchored to CCRL blitz ratings) and replaces 0.1.93 at 2891 ± 59 as the published figure; over the two shared anchors the measured gain is +69 ± 45.
Coherent 0.1.93: 2891 — and a number that did not move
Project measurement, not in the register. After 720 games against three foreign engines Coherent 0.1.93 stands at 2891 ± 59 and replaces 0.1.62 as the published figure. The news is the standstill since 0.1.91, a wider spread of the anchors and a resolved arithmetic dispute.
30 games, no loss — and a band of ± 170
Project measurement, not in the register. A mini tournament of 30 games puts Coherent 0.1.92 against its predecessor. The result is clear, the number behind it is not — two openings, no colour-swapped pairs, and an error band five times as wide as the anchoring's.
Coherent 0.1.91: +276 Elo, measured and not certified
Project measurement, not in the register. Against the same three foreign engines as in September, Coherent 0.1.91 measures a gain of 276 ± 45 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. over 0.1.62. The number stands; the path to it carries three limitations, and the measuring side reported every one of them itself.
Coherent 0.1.62 on three anchors, and the subtraction that says nothing
Project measurement, not in the register. After 720 games against 3 rated foreign engines, Coherent 0.1.62 stands at 2608 ± 56. The difference to the previous figure of 2570 is still not a measured gain, and the published band is a judgement, not a calculation.
Two anchors, one point apart: Coherent 0.1.61 at 2570 ± 55
Project measurement, not in the register. Coherent 0.1.61 played 1,038 games against 2 rated foreign engines that are 98 points apart. The 2 answers differ by one point. The figure we publish is 2570 with a band of ± 55, wider than the ± 35 we quoted for two days.
Playing strength: a transfer from foreign ratings, and IMS
A transfer from foreign ratings, followed by IMS Internally measured strength: 2200 plus the Elo difference against a frozen reference configuration of our own. A progress measure between our versions, not a rating from a ranking list. as a measure of progress between our own versions.
Two foreign yardsticks, two different answers
Project measurement, not in the register. Coherent 0.1.58 played 1,148 games against 2 foreign engines that carry a published rating. The two anchors disagree by 54 points. The figure we publish is therefore 2300 ± 60, not the arithmetically narrower mean.
A commissioned assessment of architecture and productivity
A commissioned qualitative assessment dated 6 September 2026 judges the Coherent Chess architecture very good and finds excellent overall productivity not yet sufficiently evidenced. It is not a measurement and not a playing-strength judgment.
First external measurement: IMS 2363 ± 26
Project measurement, not in the register. Coherent 0.1.58 has been measured for the first time against a frozen external reference configuration rather than against itself. Over 800 games, a number fixed in advance, the score was 71.88 %, which is IMS 2363 ± 26 at the 95 % level.
NNUE nets measure a gain against the baseline
Project measurement, not in the register: two NNUE Efficiently updatable neural network: a small network that evaluates positions and is updated incrementally move by move. nets measure a gain against the baseline at time control 8+0.08, each over 1,400 games, without a register ID or an oracle confirmation.
Three thousand games, none lost on time
A project measurement, not in the register: one build played itself three thousand games under three settings without losing a single game on time.
At most 44.36 per cent idle
Over sixty hours the machine was measuring something for at least 55.64 % of the time. The rest is an upper bound on idleness, not a measurement of it.
What the export does not carry
The data-based entries here draw on an eight-file snapshot under /daten/. Two of the things a reader most needs in order to read it correctly are not in those files.
The changes an Elo test cannot judge
Four changes entered the engine on 3 September without playing a game. What each of them addressed is the kind of defect a strength test is not built to find.
Ten of seventeen entries played no games
Between 2 and 5 September the register gained seventeen entries. Seven of them played 174,984 games; the other ten played none. Four in all record a measured gain.
The verdict did not predict the cost
Six candidates cost 149,104 games and close to thirty hours of machine time. Whether one passed did not predict what it cost.
Two rejections that mean different things
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.
Three entries that played no games
Some register entries pass without a single game. They mark the points where the measuring conditions changed — and where comparisons across the boundary stop being valid.
Two runs that were meant to fail
An engine playing itself should measure nothing. Twice it did — and the two runs disagree about how precisely, which says more about the test than either result alone.
An Elo number needs its opponent
Six attempts in one day, two of them accepted. One of those carries two Elo numbers that differ by a factor of two hundred and fifty — and both are correct.
What the result file does not label
The first run in a tournament interface, three reading errors of the same shape, and five faults the harness reports never raised.