♚افتح في التطبيق

Balanced start pairs in Double Fischer Random Chess: a computational search with multi-engine verification

Roman Morozov · CH960 (ch960.com) · 21 October 2026 · 1.2 · CC BY 4.0

Abstract

In Double Fischer Random Chess (DFRC) White and Black start from independently chosen Chess960 back ranks, so a game is defined by an ordered pair of setups, 921,600 in all. Most such pairs give one side a clear starting edge. We asked whether there are pairs in which neither side starts better, and how many of the candidates survive when the search budget keeps growing and when other engines look at the same position. 460,800 pairs, one representative of each pair and its left–right mirror image, were screened with Stockfish 17.1 in a staged search whose windows were calibrated on the ordinary symmetric Chess960 starts as a built-in control; a mirror image is a different position, because castling always ends on fixed files, so this covers half of the space. Candidates were then taken through a budget ladder of 400 M, 800 M, 1.6 G and 4 G nodes, a subset to 8 G and 16 G nodes, and every candidate was re-evaluated by two engines of a different construction, Berserk 14 and Lc0 0.32.1. The admission rule is a corridor, |cp| ≤ 5 at every budget at which a pair has been evaluated, with no consistently one-sided reading, confirmed by both other engines below their own median on symmetric positions. 436 pairs satisfy it. The verified pairs behave differently from the best symmetric starts: when the budget grows, the symmetric starts drift towards White only, while the pairs that leave zero move both ways; swapping the two setups of a balanced pair gives White a median edge of +76 cp; and in 6,450 engine games from 43 of the pairs White scores 49.2 %, a small deficit that is statistically visible game by game and borderline when games are clustered by pair. The mirror twins of the balanced pairs are not balanced: only 22 of 236 stay within the corridor on the whole ladder, and at 4 G nodes they read 21.3 cp from zero on average. A funnel audit of 1000 pairs rejected just outside the windows finds 4.1 % of them even at 400 M nodes. All data, run passports and the exact engine builds are published, so every number here can be recomputed.

1. Introduction

Chess960 removes opening preparation by shuffling the back rank, but it keeps the two armies mirror images of each other. Double Fischer Random Chess goes one step further and draws the two back ranks independently [1]. The space of starts grows from 960 to 960 × 960 = 921,600, and with it grows the chance that a start is unfair: at a quick glance about a quarter of the pairs favour Black, and most of the rest favour White by more than 10 centipawns.

That unfairness is also an opportunity. If White's first move is worth a few tens of centipawns, then somewhere in the space there should be pairs where Black's back rank is worth exactly that much more than White's, so that the two sides start even. Such pairs would be interesting for play, since they remove the first-move advantage without changing the rules of movement, and they are interesting as objects of study, because they let one ask what a setup is "worth" on its own.

This paper reports a staged search for such pairs over half of the DFRC space and the programme that was run to find out whether the balance found by one engine at one budget is a property of the position or an artefact of the measurement. The work was done for CH960, a free Chess960 platform, between 18 September and 9 October 2026; the search continues, and this text describes the state of the data on the snapshot date given in Section 9. A first draft of this text, circulated on 8 October 2026, treated a position and its mirror image as equivalent and compared spreads across sets of different composition; both errors are corrected here (Sections 3.1 and 5.1), and the title no longer calls the search exhaustive.

Three things are, to our knowledge, new. First, the target: previous engine passes over the DFRC space selected positions by their draw rate, which is what engine testing needs, not by closeness to equality. Second, the method: a staged search whose promotion windows are set by a control experiment rather than by guesswork, a budget ladder that goes an order of magnitude beyond the usual evaluation depth, and an admission rule that requires two engines of a different design to agree. Third, the published result: a verified, reproducible set of balanced pairs together with the raw data behind every claim.

2. Related work

The DFRC scheme itself is not ours. The Top Chess Engine Championship has run DFRC events since 2022 and later made DFRC the format of its "FRD" finals [1, 2]; its rules say that start positions are generated at random or picked by a bookmaker, and no selection criterion is published. An early request for the same variant on Lichess, under the name "Chess 960²", notes that in a small sample White was almost always ahead by at most half a pawn [3].

The closest prior work is in the opening-book repository of the Stockfish project. In July 2022 it received a file with all 921,600 DFRC start positions, with the remark that some of them are "really unbalanced by nature" [4]. In May 2025 R. Nürnberg added a book built from a complete pass over all 921,600 positions with Stockfish 17.1 at depth 22: 105,883 positions had a model draw rate between 40 % and 60 % and 20,797 between 48 % and 52 %; the latter set was published as DFRC_4852_v1.epd [5]. That pass filtered on the draw share of Stockfish's win–draw–loss model, the quantity that matters for engine testing, not on the centipawn evaluation, and it covered the whole space rather than one representative per mirror pair.

Balance of the 960 symmetric starts has been studied several times. Gunderson evaluated all 960 with Stockfish 9 at depth 39–40 and found a range from 0.00 to +0.57 pawns with 27 positions at 0.00 and the classical start at +0.22 [6]. Barthelemy evaluated them with Stockfish 17.1 at depth 30, reports a mean White advantage of +0.33 ± 0.12 pawns, and studies their complexity [7]. Molas analysed more than fourteen million Lichess games and found no start position whose practical results differ from the others by more than noise [8]. Our own baseline for the symmetric starts is the Morozov Balance Index, described in Section 3.5.

Castling in Chess960 and the numbering of the 960 setups follow Scharnagl's scheme [9]; the notation for castling rights follows Shredder-FEN, in which the rights are written as the files of the rooks [10].

3. Definitions and conventions

3.1 Pairs, mirror images and coverage

A pair (w, b) means that White starts from Scharnagl setup w and Black from setup b, each numbered 0–959, with the classical start being 518. Pawns stand on their usual ranks, castling is ordinary Chess960 castling for each side on its own rank, and all other rules are unchanged.

Reflecting the board left to right maps a setup w to its mirror image mirror(w) and a pair (w, b) to (mirror(w), mirror(b)). The reflected position is not equivalent to the original, because Chess960 castling always ends on fixed squares — king on c1 and rook on d1, or king on g1 and rook on f1 (c8/d8 and g8/f8 for Black) [12] — and the pairs of files c/d and g/f are not mirror images of each other. A position and its reflection therefore have different castling options: with the back rank BBQNRKRN White can castle on the first move, while with its reflection NRKRNQBB no castling is possible until pieces have moved. The difference is not a curiosity. All 960 symmetric starts were evaluated individually at 400 M nodes for the Morozov Balance Index (Section 3.5), which gives 480 pairs of mirror twins measured by the same engine with the same budget: the twins differ by 9.8 cp on average (median 8, maximum 53); 61 % of the twin pairs differ by more than 5 cp and 37 % by more than 10 cp; the correlation between twins is only r = 0.50; the classical start reads +11 cp and its mirror image, setup 534, reads +49 cp. For comparison, raising the budget from 400 M to 1.2 G nodes moves the evaluation of one and the same position by 5.3 cp on average, so twins differ about twice as much as a tripling of the budget does.

The search nevertheless treated a pair and its reflection as one class and evaluated only the lexicographically smaller of the two. No setup is its own mirror image (0 of 960, because the king and queen change sides), so the 921,600 ordered pairs fall into 460,800 such classes, and the search covers exactly 460,800 representatives: one half of the space. Every published pair is a pair that was actually evaluated; the value of its mirror twin is not inferred from it. The mirror twins of the published pairs were evaluated directly on the same budget ladder, and they are not balanced (Section 5.8); the other half of the space is scheduled for the same staged search and will be reported in a later version. The 480 classes with w = b are the ordinary Chess960 starts; they are kept in the search as a control and are excluded from the results.

3.2 Notation and sign convention

Positions are written in Shredder-FEN; the castling field lists the file letters of each side's rooks on its own rank. All evaluations are in centipawns from White's point of view: a negative value means that Black stands better.

3.3 Engine settings and the score rule

The main engine is Stockfish 17.1 (SHA-256 5f95eaea0d4eb697381989187ce6eb4d6ad59283c34421765ecc73cdb09ba766) with Threads = 1, UCI_Chess960 = true, a hash size fixed per stage, and a search limited by nodes rather than time. Each position is searched after ucinewgame. The score taken is the one reported at the last completed iteration of iterative deepening; an iteration cut short by the node limit reports a bound rather than a settled value and is not used. With these settings a search is a deterministic function of the position and the budget: the same build on any machine returns the same number, so every value in this paper can be reproduced bit for bit.

3.4 Budgets

Budgets are written in nodes. The screening stages run from 100 k to 128 M nodes (Section 4); the verification ladder runs at 400 M, 800 M, 1.6 G and 4 G nodes, with a subset at 8 G and 16 G. For orientation, 400 M nodes of single-thread Stockfish 17.1 reach a nominal depth in the mid-forties on these positions, and 16 G nodes reach depths of 60–74.

3.5 The symmetric baseline: the Morozov Balance Index

The 960 symmetric starts were evaluated with the same engine, settings and score rule at 400 M nodes; the result is published as the Morozov Balance Index (MBI) [11]. Over the 960 positions the evaluation runs from −7 to +83 cp with median +26 and mean +27.3; 947 positions favour White, 9 read exactly 0 and 4 read slightly negative. 64 positions lie within ±10 cp and 147 within ±15 cp; the classical start reads +11. The mean over the 480 lexicographically smaller twins — the representatives the search uses — is +25.5 cp, over their mirror images +29.1 cp: the two halves of the space are not interchangeable even on average. The index itself is the rank of a position by |cp|, rescaled so that 100 is the most balanced. It was validated against 3,000 engine games from 60 positions (Spearman 0.76, bootstrap 95 % interval 0.63–0.85). In this paper the symmetric starts serve two purposes: their 400 M evaluations calibrate the search windows, and the most balanced of them are the control group against which the pairs are compared at every deeper budget and with every other engine.

4. The search

4.1 Stages and windows

The search is staged. Stage 0 evaluates every one of the 460,800 representatives with a small budget; a pair advances to the next stage only if its |cp| at the current stage is within that stage's window. Table 1 lists the stages.

Stage Nodes Hash, MB Window |cp|
0 100 k 16 50
1 500 k 16 40
2 2 M 32 41
3 8 M 64 32
4 32 M 128 30
5 128 M 256 35

Table 1. Screening stages: node budget, hash size and the promotion window.

Beyond the last regular stage, verified pairs are deepened from 400 M nodes, the MBI budget, by doubling: 800 M, 1.6 G, 3.2 G. Stage 0 traverses the representatives in a fixed seeded shuffle (seed 20,260,918), so that the set of pairs evaluated at any moment is reproducible. Waves are served depth-first: the deepest stage with pending candidates goes first, and within a stage the most even candidates go first, so verified pairs appear while screening is still running. Since 26 September 2026 most of the compute has gone into deepening rather than into screening new pairs; the actual share is written into the run passport (0.97 at the snapshot).

4.2 Why the windows are wide: the control experiment

A shallow search does not simply add noise to the deep value; it is biased. To measure both the noise and the bias, the 93 symmetric representatives with |MBI evaluation| ≤ 15 cp were run through every stage outside the windows and compared, each with its own 400 M value. Table 2 gives the result.

Stage Nodes Shift vs 400 M, cp Noise σ, cp r within ±15 Median |cp| of even 95th pct of even
0 100 k +17.9 ±12 −0.08 27.5 45
1 500 k +17.4 ±10.6 0.04 26 39
2 2 M +15.4 ±10.7 −0.00 23.5 36
3 8 M +13.7 ±8.8 0.12 21 32
4 32 M +9.8 ±7.4 0.20 18 27
5 128 M +7.1 ±6 0.44 16 23

Table 2. The control experiment: 93 symmetric starts within ±15 cp at 400 M nodes, run through every stage outside the windows. The shift is the mean difference from the 400 M value; r is the correlation with it; the last two columns are over the 48 starts within ±10 cp at 400 M.

2026-10-21T20:51:45.977780 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 100 k 500 k 2 M 8 M 32 M 128 M Stage budget, nodes 0 5 10 15 20 25 30 Shallow − 400 M, cp +17.9 +17.4 +15.4 +13.7 +9.8 +7.1
Figure 2. The control experiment: mean shift of the shallow evaluation from the 400 M value (line) and its spread (band) across the stages, for the symmetric starts within ±15 cp.

At 100 k nodes the engine reads an even position at about +18 cp for White with a spread of ±12 cp; at 128 M nodes the shift is still +7 cp. Within the band of near-even positions the shallow value hardly correlates with the deep one (r from −0.08 to 0.20 up to 32 M nodes), so early stages can only reject a clear advantage; which of two near-even positions is the more even is decided by depth alone. The practical rule follows: a stage's window is the 95th percentile of |cp| among the control positions that are even at depth (|cp| ≤ 10 at 400 M, 48 representatives), so that at least 95 % of the even control positions survive each stage. The windows in Table 1 were set this way on 19 September 2026; the original, narrower windows of 30/25/20/15/12/10 cp would have discarded most of the even control positions at the middle stages, and all pairs were re-screened after the change.

Two caveats belong here. The 95 % figure holds per stage; over five promotions, if misses were independent, about 77 % of even positions would reach the end of screening. And the calibration set consists of symmetric positions; that the same windows retain even asymmetric pairs at the same rate was tested directly by a funnel audit: 1000 pairs rejected at stages 0–3 with |cp| at most 30 cp beyond the window were re-evaluated at 400 M nodes regardless of the filter. 41 of them (4.1 %) read within ±5 cp and 72 within ±10 cp; the mean |cp| is 40.7. Even right at the window edge the windows lose few pairs that are even at 400 M, and an even reading at 400 M is only the first rung of the admission ladder.

The bias is visible already at stage 0 across the 480 symmetric representatives: the mean |cp| is 36.6 at 100 k nodes against 25.5 at 400 M, and the correlation between the two is only r = 0.41.

4.3 The shape of the space at stage 0

Stage 0 was completed on the first evening of the search. Over all 460,800 representatives the median |cp| is 63, the 90th percentile 137 and the maximum 499; 330,760 pairs favour White by more than 10 cp, 100,028 favour Black by more than 10 cp, and 30,012 lie within ±10 cp. 114,115 pairs (24.8 %) read negative, which never happens at depth for a symmetric start: Black's back rank can indeed outweigh White's move.

2026-10-21T20:51:46.130112 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 100 k 500 k 2 M 8 M 32 M 128 M 400 M 800 M 1.6 G 3.2 G Deepest stage (nodes) 1 0 0 1 0 1 1 0 2 1 0 3 1 0 4 1 0 5 1 0 6 Pairs 277,515 170,410 684 1,504 901 5,883 30 1,825 1,953 95
Figure 1. Pairs by the deepest stage reached at the snapshot (log scale). Screening stops most pairs at 100 k or 500 k nodes; the right-hand bars are the verification ladder.

4.4 Compute

At the snapshot the search journal holds 699,622 evaluations of all 460,800 representatives; the sum of their nominal budgets is 1.0 × 1013 nodes. The verification runs of Section 5 add about half as much again. All runs were made on one workstation with 24 physical cores and 128 GB of memory, with 10 to 20 single-thread engine processes in parallel; parallelism changes only the wall-clock time, never a number.

5. Verification

Verification started from the first published set of 52 pairs, frozen on 21 September 2026 after reading exactly 0.00 cp at 400 M, 800 M and 1.6 G nodes. The question was whether that zero is a property of the pair. Unless stated otherwise, the experiments of this section concern this first set and its subsets, not the full published set of Section 6.

5.1 The budget ladder: 4 G, 8 G, 16 G

All 52 pairs were re-evaluated at 4 G nodes, the survivors at 8 G, and their survivors at 16 G (Table 3).

Budget Pairs Exactly 0 Drifted to Black Drifted to White Mean, cp SD, cp Depth
4 G 52 43 7 2 −0.83 3.6 54–67
8 G 43 39 1 3 0.47 2.1 59–70
16 G 39 35 2 2 −0.05 1.7 60–74

Table 3. The budget ladder for the first set of 52 pairs.

2026-10-21T20:51:46.238436 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ 4 G 8 G 16 G 0 10 20 30 40 50 60 Pairs at exactly 0 43 / 52 39 / 43 35 / 39 0 1 2 3 4 SD of the set, cp 3.6 2.1 1.7
Figure 3. The budget ladder: pairs still reading exactly 0 (bars) and the standard deviation of the whole set (line) at 4 G, 8 G and 16 G nodes.

At each doubling roughly one pair in ten leaves exact zero: 52 → 43 → 39 → 35. The standard deviation in Table 3 falls from 3.6 to 2.1 to 1.7 cp, but the three rows describe sets of different composition, since only survivors went deeper; the fall therefore says nothing about depth by itself. On the fixed set of 52 pairs evaluated at all three budgets the standard deviation is 3.56 / 4.43 / 5.40 cp, with 43 / 40 / 37 pairs at exactly zero. The honest reading is that "exactly zero" was never the right description of these pairs: a deeper search redistributes a residual of a few centipawns, and whether that residual grows or shrinks with depth has to be measured on one and the same set — the fixed set of all 52 pairs above. On it the residual does not shrink: the spread grows with the budget (3.6 → 4.4 → 5.4 cp) while the mean stays within a centipawn of zero (−0.8 / −0.4 / −0.8 cp). Deeper search finds small errors on both sides of a balance that stays centred, which is why the admission rule asks for a band at every budget rather than for exact zero.

5.2 Direction of drift: pairs versus symmetric starts

What distinguishes a balanced pair from a merely lucky symmetric start is the direction in which the value moves with depth. The 13 most balanced symmetric starts, re-evaluated at the same 4 G nodes, moved to a median of +9 cp (range 0 to +20); the 9 of them that read exactly 0 at 400 M now read between +5 and +20, and not one went negative. A further 134 symmetric starts re-evaluated at 1.2 G nodes moved by a median of +3 cp, again with none going negative. The pairs that left zero moved both ways: at 4 G, 7 towards Black and 2 towards White; at 8 G, 1 and 3; at 16 G, 2 and 2. A symmetric start has an unpaid tempo that depth gradually prices in; a balanced pair has a structural compensation, and depth finds small errors on either side of it.

2026-10-21T20:51:46.283671 image/svg+xml Matplotlib v3.10.8, https://matplotlib.org/ −20 −15 −10 −5 0 5 10 15 20 25 Evaluation, cp (White's point of view) pairs, 16 G pairs, 8 G pairs, 4 G symmetric, 4 G symmetric, 400 M
Figure 4. Where the value goes with depth. Top: the most balanced symmetric starts at 400 M (grey) and the same positions at 4 G nodes (orange). Below: the first set of pairs at 4 G, 8 G and 16 G nodes (blue). Each dot is one position.

5.3 Depth predicts survival; self-play does not

At 4 G nodes the non-survivors had reached nominal depths of 54–60, while all 29 pairs that reached depth 61 or more held zero. Already at 1.6 G nodes a depth threshold of 56 would have kept 37 pairs of which only 2 later left zero, at the cost of discarding 8 good ones. Engine self-play at a modest budget, by contrast, does not separate the two groups at all: every one of the 9 pairs that left zero at 4 G had scored 10 draws out of 10 games in the earlier self-play.

5.4 Reversed pairs

If the balance of a pair is a real compensation, swapping the two setups should destroy it. For the 43 pairs that held zero at 4 G, the reversed pair (b, w) was evaluated at 400 M nodes: every one of them favours White, with a median of +76 cp, mean +79.1, range +22 to +175. The same set of two back ranks, handed out the other way round, turns an even start into one that is close to won. At stage 0 the reversed pairs had read a median of +93 cp, so this property is visible even to a shallow search. The result is established for these 43 pairs; it has not been measured for the pairs admitted later.

5.5 Independent engines

Two engines of a different construction were run on the same positions: Berserk 14, an alpha-beta engine with its own, independently trained network, at 400 M nodes; and Lc0 0.32.1 with network 791556, a Monte-Carlo tree search engine with a large neural network, at 100 k tree nodes. Each engine was first run on the 13 most balanced symmetric starts to learn how it reads an ordinary even Chess960 position: Berserk gives them a median of +10 cp (range 0 to +22), Lc0 a median of +18 cp (range +14 to +20). Neither engine puts any symmetric start below zero.

On the 43 pairs that held zero at 4 G, Berserk reads a mean of 0.1 ± 0.8 cp (median 0) and Lc0 a mean of +0.5 ± 2.1 cp (median +1); Berserk puts 9 of them below zero and Lc0 19. The two engines agree with each other on the pairs with r = 0.45. The claim this supports is the one used in the admission rule: by the reading of an independent engine, a balanced pair is closer to equality than a typical balanced symmetric start.

5.6 Setup weights

Is the balance of a pair the sum of two independent setup values plus a tempo? To test this, every setup X was given a weight: the evaluation of "X as White against the classical setup 518 as Black" at 400 M nodes. Weights run from −61 to +183 cp (mean +54.3, median +49, SD 44.8); 91 setups have a negative weight, that is, as White against the classical array they stand worse; the classical array itself weighs +11 cp, which is simply the evaluation of the classical start at this budget. A linear model cp ≈ k·(W_w − W_b) + T fitted over all 460,800 stage-0 evaluations gives k = 0.80, a tempo T = +47.2 cp and r = 0.73: the additive model explains 54 % of the variance of the shallow evaluation, and the residual standard deviation is 47 cp against 70 cp for the raw values. The balanced pairs have, as the model requires, a systematically negative weight difference: mean −48.4 cp over the 43 pairs held at 4 G and median −33 cp over the published set, where 349 of 436 pairs give the heavier setup to Black. The model is useful as a prior, not as a judge: the window of predictions within ±10 cp holds 10.1 % of the space but 20.6 % of the published set, an enrichment of ×2.0; the rest of the balance is interaction between the two ranks that no sum of weights captures.

5.7 Self-play

The 43 pairs held at 4 G were played out by the engine against itself, 150 games per pair at 500,000 nodes per move with four random opening plies kept within 15 cp of the best line: 6,450 games, 116 White wins, 6,114 draws, 220 Black wins, a White score of 49.19 %. The uncertainty of that score must be computed from the actual outcomes 0, ½ and 1, not from a two-outcome binomial model: with 95 % draws the per-game standard error gives a 95 % interval of ±0.28 percentage points, and the deficit of White is 5.7 standard errors from equality. Games from the same pair are not independent samples of the population of positions, however; clustering by pair (43 clusters) widens the interval to ±0.94 points and the deficit to 1.7 standard errors. The direction is consistent: 21 pairs score below 50 % for White and 8 above, and Black wins the decisive games about two to one (116 against 220). Self-play at this budget measures practical play, not the evaluation; what it shows is a small but real practical edge for Black in these pairs at 500,000 nodes per move, of the order of one percentage point, not equality within noise.

5.8 Mirror twins of the balanced pairs

Section 3.1 showed on the symmetric starts that a position and its mirror image are different positions. The direct test is to take the balanced pairs themselves and evaluate their mirror twins on the same ladder. This was done for the 236 pairs of the edition of 10 October 2026 at 400 M, 800 M, 1.6 G and 4 G nodes (Table 4).

Budget Twins within ±5 cp within ±10 cp mean |cp| median above +5 below −5
400 M 237 34 57 24.7 0 95 108
800 M 237 44 62 23.3 0 94 99
1.6 G 236 48 71 22.5 0 87 101
4 G 236 52 81 21.3 0 86 98

Table 4. Mirror twins of the balanced pairs on the budget ladder (evaluation in cp from White's point of view).

Only 22 of the 236 twins (9 %) stay within ±5 cp at every budget, as every pair of the set does by construction; at 4 G nodes the twins read 21.3 cp from zero on average, 86 of them favour White by more than 5 cp and 98 favour Black. The median stays near zero, so the twins are not systematically better for one side; they are simply not balanced. The mirror twin of a balanced pair is therefore no more likely to be balanced than a pair near the search windows, and the balance of a pair cannot be transferred to its reflection: each half of the space has to be searched on its own.

6. The admission rule and the published set

6.1 Criterion v2

The first set had been admitted by "exactly 0.00 at 400 M, 800 M and 1.6 G nodes, one engine". Section 5 shows why that rule is wrong: it selects the luckiest pairs rather than the most balanced ones, since about one in ten leaves exact zero at every doubling. The rule that replaced it on 26 September 2026, extended on 9 October 2026 to every evaluated budget, requires all at once:

Conditions C and D are not arbitrary thresholds: each engine is compared with its own reading of ordinary even Chess960 positions, so the statement is the same for all three engines and can be checked by anyone with the published builds.

6.2 Calibration on the first set and the published composition

Applied to the 52 pairs of the first set, the core rule (the four rungs and the two engines) kept 33 and rejected 19; the extension to deeper budgets removed a further 5 pairs that read +6 to +10 cp at 8 G nodes. The rejections are informative: of the 24 pairs removed in total, 11 fail the Lc0 condition, 7 the Berserk condition, 14 leave the corridor at 4 G or deeper and 3 fail condition B (a pair can fail several). Pairs admitted by the same rule as their verification completed bring the published set to 436 pairs (Table 5); there is one set and one rule, the search's candidates join it as their verification completes, and the 24 former pairs remain published with the reason for their removal.

Date Admitted Set size
21 September 2026 28 28
3 October 2026 122 150
9 October 2026 73 223
10 October 2026 13 236
15 October 2026 87 323
21 October 2026 113 436

Table 5. Growth of the published set: pairs admitted by criterion v2 on each date as their verification completed, and the size of the set after that date.

6.3 Properties of the set

Over the 436 pairs, the 4 G evaluation reads exactly 0 for 385, negative for 33 and positive for 18 (range −5 to +5, mean −0.04, SD 1.08); 254 pairs read 0 at all four rungs. Berserk reads the set with median 0, SD 3.5 and range −9 to +9 (104 below zero, 105 above); Lc0 with median +1, SD 8.8 and range −17 to +17 (179 below, 224 above). 10 engine games per pair at 500,000 nodes per move, available for 436 of the 436 pairs at the snapshot, give 4,360 games with a White score of 49.30 % (99 wins, 4,101 draws, 160 losses), 2.8 standard errors below equality when clustered by pair — the same small practical edge for Black as in Section 5.7. The full list, with the deeper budgets where they exist, is given in Appendix A.

7. Discussion and limitations

What the evaluation is. An engine evaluation is a statement about best play as the engine sees it at a given budget, not the game-theoretic value of the position and not a prediction of human results. The verification shows that the statement is stable in the ways that matter: the balance survives a forty-fold increase in budget to within a few centipawns, it is confirmed by engines of a different design, and it disappears when the setups are swapped.

What "balanced" means here. We do not claim that any pair is exactly zero in a theoretical sense, and we do not claim that the residual shrinks with depth. The defensible claim, and the one encoded in the admission rule, is that each pair is balanced to within 5 cp at every budget at which it has been evaluated, with the residual scattered on both sides of zero, in contrast to the symmetric starts, whose residual lies on one side only.

Coverage is half of the space. The search evaluated one representative of each mirror pair, and mirror twins are different positions (Section 3.1). Nothing is claimed about the pairs that were not evaluated; in particular, the mirror twin of a published pair is not a balanced pair until it has been verified itself.

One engine family dominates. Stockfish does the screening and the ladder; the other two engines act as referees, not as independent searchers. A pair that Stockfish misreads at every budget would never reach the referees. The control experiment of Section 4.2 bounds this risk at the screening stage for symmetric positions, and the wide windows were chosen for that reason, but it cannot remove it; the funnel audit (Section 4.2) puts the loss at the window edge at 4.1 % of rejected pairs.

Budgets are not comparable across engines. Lc0's 100 k tree nodes and Berserk's 400 M alpha-beta nodes are different quantities; each engine is therefore compared only with its own reading of the symmetric control, never with another engine's number.

Self-play is a weak but not an empty instrument. At the budgets affordable for thousands of games the draw rate is about 94 %, and the few decisive games are decided by play. Still, both self-play experiments show a practical edge for Black of about one percentage point with Black winning the decisive games about two to one. That is compatible with an evaluation of zero at the root — the evaluation measures the position, the games measure how two copies of the same engine handle it at 500 k nodes per move — but it is a finding, not noise, and it is published as such.

The search continues. New pairs reach the ladder and some will be admitted; counts, ranks and the deepest stage of a given pair change over time. Every number in this paper is computed from the data snapshot of Section 9, and the site shows the live state. The mirror twins were evaluated for the edition of 10 October 2026 (236 pairs); the twins of pairs admitted later, and the other half of the space, will be reported in a later version.

8. Conclusions

A staged search over 460,800 representatives of the 921,600 DFRC start positions, one from each pair of mirror images, with windows calibrated on the symmetric starts as a control, finds a set of 436 pairs in which neither side starts better by the reading of three engines, to within 5 cp at every budget examined. The set has properties that no selection by a single evaluation could guarantee: the pairs that leave zero at deeper budgets move both ways while symmetric starts move only towards White, swapping the setups of a verified pair hands White a large advantage, and a simple additive model of setup weights explains about half of the balance and no more. Mirror images are different positions, so the other half of the space remains to be searched. The data and the engine builds are published so that each of these statements can be recomputed.

9. Data and code availability

All primary data are published under CC BY 4.0 together with this paper: the search journal with every evaluation, the run passports with engine builds and settings, the ladder, reverse-pair, weight, Berserk, Lc0 and self-play runs, the symmetric control, and the composition with the reasons for every removal. The snapshot used for this text was taken on 2026-10-21 (699,622 journal records). The live rating, the encyclopedia of pairs with a page per pair, and the methodology are at ch960.com/asym; the symmetric baseline is at ch960.com/balance. The tools that run the search, the ladder and the dataset build, and the script that computes every number in this text from the data, are part of the CH960 repository.

Acknowledgements

The computational pipeline (search scheduler, verification runs, dataset builds) and the draft of this text were prepared with the assistance of Claude Code, an AI assistant by Anthropic, working under the author's direction; the research questions, the design decisions and the responsibility for the content are the author's. The author thanks the reviewer of the first draft for identifying the mirror-image error, the selection effect in the spread comparison and the mistaken confidence interval of Section 5.7. The engines are the work of their respective teams: Stockfish, Berserk (Jay Honnold) and Leela Chess Zero.

References

  1. Wikipedia, "Chess960", section Tournaments: DFRC/FRD at TCEC since 2022. https://en.wikipedia.org/wiki/Chess960
  2. TCEC wiki, "TCEC DFRC rules" and "TCEC FRD rules". https://wiki.chessdom.org/TCEC_DFRC_rules, https://wiki.chessdom.org/TCEC_FRD_rules
  3. Lichess forum, "Chess 960² (feature request)". https://lichess.org/forum/lichess-feedback/chess-960-feature-request
  4. J. VandeVondele, "All DFRC opening positions", official-stockfish/books, pull request #31, 3 July 2022. https://github.com/official-stockfish/books
  5. R. Nürnberg, "add two DFRC books", official-stockfish/books, pull request #67, 24 May 2025 (file DFRC_4852_v1.epd). https://github.com/official-stockfish/books/pull/67
  6. S. H. Gunderson's evaluation of all 960 positions, reported in M. Klein, "What's The Most Unbalanced Chess960 Position?", Chess.com, 14 March 2018. https://www.chess.com/article/view/whats-the-most-unbalanced-chess960-position
  7. M. Barthelemy, "Not all Chess960 positions are equally complex", arXiv:2512.14319, December 2025 (v3 March 2026). https://arxiv.org/abs/2512.14319
  8. A. Molas, "Analyzing Chess960 Data", January 2023. https://alexmolas.com/2023/01/11/chess-960-initial-position.html
  9. R. Scharnagl, Fischer-Random-Schach (FRC/Chess960) – Die revolutionäre Zukunft des Schachspiels, 2004, ISBN 3-8334-1322-0; numbering scheme: https://en.wikipedia.org/wiki/Chess960_numbering_scheme
  10. Shredder-FEN and X-FEN notation for Chess960 castling rights: https://en.wikipedia.org/wiki/X-FEN
  11. R. Morozov, "Chess960 Balance Index (MBI)", ch960.com/balance, 2026, CC BY 4.0. https://ch960.com/methodology/balance
  12. FIDE Laws of Chess, Guidelines II: Chess960 rules (castling ends with the king on c1/g1 and the rook on d1/f1). https://handbook.fide.com/chapter/E012023
  13. Stockfish 17.1, released 30 March 2025. https://github.com/official-stockfish/Stockfish/releases/tag/sf_17.1
  14. Berserk 14, released 24 May 2026. https://github.com/jhonnold/berserk/releases/tag/14
  15. Leela Chess Zero v0.32.1, released 23 November 2025. https://github.com/LeelaChessZero/lc0/releases/tag/v0.32.1

Appendix A. The published set

# White Black Since 400 M / 800 M / 1.6 G / 4 G 8 G / 16 G Berserk Lc0
1 4 837 2026-10-09 0 / −1 / 0 / −1 — / — +1 −6
2 5 16 2026-10-03 0 / 0 / −4 / −4 — / — 0 −7
3 6 99 2026-10-21 0 / +4 / 0 / 0 — / — 0 +4
4 7 512 2026-10-21 0 / +3 / 0 / 0 — / — +6 +9
5 7 592 2026-10-21 0 / −2 / −3 / −1 — / — +2 −2
6 7 595 2026-10-10 0 / 0 / 0 / 0 — / — 0 +6
7 7 850 2026-10-15 +3 / −1 / 0 / 0 — / — 0 +10
8 8 513 2026-10-10 0 / 0 / 0 / 0 — / — 0 +13
9 9 60 2026-10-21 −3 / 0 / 0 / +1 — / — −9 −2
10 9 772 2026-10-10 0 / 0 / −3 / 0 — / — 0 +8
11 9 850 2026-10-10 0 / 0 / 0 / 0 — / — −6 −1
12 9 949 2026-10-10 0 / 0 / +3 / 0 — / — −5 −14
13 12 7 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +9
14 13 297 2026-10-21 0 / 0 / 0 / 0 — / — +8 +9
15 14 641 2026-10-10 +3 / 0 / 0 / 0 — / — 0 −1
16 15 203 2026-10-21 +1 / 0 / 0 / 0 — / — +1 +6
17 15 658 2026-10-03 0 / 0 / +1 / 0 — / — 0 −2
18 17 823 2026-10-21 0 / 0 / 0 / 0 — / — +8 +1
19 19 400 2026-10-09 +1 / 0 / 0 / −3 — / — 0 0
20 20 846 2026-10-10 0 / 0 / 0 / 0 — / — +1 0
21 21 392 2026-10-09 0 / 0 / 0 / −2 — / — +1 −9
22 23 115 2026-10-21 0 / 0 / 0 / 0 — / — +3 −8
23 24 98 2026-10-21 0 / 0 / 0 / 0 — / — +3 +8
24 26 743 2026-10-21 0 / 0 / 0 / 0 — / — +1 +6
25 27 43 2026-10-03 −2 / 0 / 0 / 0 — / — −5 +4
26 27 488 2026-10-03 0 / 0 / 0 / 0 — / — −8 +5
27 28 132 2026-10-21 +4 / 0 / 0 / 0 — / — 0 −9
28 29 723 2026-10-09 0 / 0 / 0 / 0 — / — +8 +13
29 30 357 2026-10-03 +3 / 0 / −2 / 0 — / — +4 −3
30 30 848 2026-10-09 0 / 0 / 0 / 0 — / — −6 0
31 31 573 2026-10-09 0 / 0 / 0 / 0 — / — 0 −7
32 32 211 2026-10-09 0 / 0 / 0 / 0 — / — 0 −7
33 33 35 2026-09-21 0 / 0 / 0 / 0 0 / 0 −5 −12
34 33 38 2026-10-10 0 / 0 / 0 / 0 — / — 0 +7
35 33 483 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −9
36 35 401 2026-10-21 0 / 0 / 0 / 0 — / — −8 −10
37 36 753 2026-10-03 0 / 0 / 0 / 0 — / — +6 −3
38 36 817 2026-10-03 0 / 0 / 0 / 0 — / — 0 −6
39 37 822 2026-10-03 0 / 0 / −1 / 0 — / — 0 +4
40 39 164 2026-10-21 0 / 0 / 0 / 0 — / — +7 +2
41 40 289 2026-10-10 −5 / 0 / 0 / 0 — / — 0 +16
42 40 838 2026-10-21 0 / −2 / 0 / 0 — / — 0 +1
43 42 528 2026-10-21 0 / 0 / 0 / 0 — / — 0 +11
44 44 865 2026-10-21 0 / 0 / 0 / 0 — / — +6 +10
45 45 146 2026-10-03 0 / 0 / 0 / −3 — / — 0 −7
46 45 180 2026-10-21 0 / 0 / 0 / 0 — / — 0 +3
47 45 459 2026-10-09 0 / 0 / 0 / 0 — / — 0 +11
48 46 147 2026-10-10 0 / 0 / 0 / 0 — / — 0 +5
49 46 583 2026-10-10 0 / 0 / +1 / −2 — / — −7 +1
50 46 643 2026-10-10 0 / 0 / 0 / 0 — / — +6 −1
51 48 902 2026-10-10 0 / 0 / 0 / 0 — / — 0 +9
52 49 103 2026-10-03 0 / 0 / 0 / 0 — / — +2 +4
53 49 563 2026-10-21 0 / +4 / 0 / 0 — / — 0 −7
54 51 499 2026-10-03 0 / 0 / 0 / 0 — / — −2 −5
55 51 516 2026-10-15 0 / 0 / 0 / 0 — / — 0 +17
56 54 116 2026-10-09 0 / 0 / 0 / 0 — / — −8 +6
57 54 117 2026-10-21 0 / 0 / 0 / 0 — / — −5 +13
58 56 37 2026-10-09 +2 / 0 / 0 / 0 — / — +8 +3
59 56 854 2026-10-09 0 / 0 / 0 / 0 — / — 0 +13
60 58 547 2026-09-21 0 / 0 / 0 / 0 0 / 0 +1 +8
61 58 593 2026-10-15 0 / 0 / 0 / 0 — / — 0 +12
62 59 42 2026-10-03 0 / 0 / 0 / 0 — / — −7 +1
63 59 290 2026-10-21 0 / 0 / 0 / 0 — / — +5 +17
64 60 216 2026-10-03 0 / 0 / 0 / 0 — / — −2 0
65 60 340 2026-10-09 0 / 0 / 0 / 0 — / — 0 +15
66 61 284 2026-10-03 −2 / 0 / 0 / 0 — / — −4 +12
67 61 423 2026-10-09 0 / 0 / 0 / −2 — / — −6 −1
68 61 429 2026-10-21 0 / −2 / +2 / 0 — / — 0 +12
69 61 903 2026-10-15 +4 / 0 / +1 / 0 — / — 0 −10
70 62 109 2026-10-21 0 / 0 / 0 / 0 — / — −3 −4
71 62 543 2026-10-03 0 / −3 / −3 / −1 — / — 0 −10
72 63 31 2026-09-21 0 / 0 / 0 / 0 0 / 0 −1 0
73 63 609 2026-10-03 −2 / +1 / 0 / 0 — / — 0 +4
74 64 35 2026-10-21 0 / 0 / 0 / 0 — / — 0 −1
75 65 39 2026-10-09 −2 / 0 / 0 / −1 — / — −8 +13
76 65 103 2026-10-21 −5 / +1 / 0 / 0 — / — 0 −3
77 65 213 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +6
78 65 955 2026-10-21 +3 / +3 / 0 / 0 — / — 0 −8
79 66 242 2026-10-09 0 / 0 / 0 / 0 — / — 0 −5
80 66 880 2026-10-15 0 / 0 / 0 / 0 — / — −3 −1
81 69 740 2026-10-09 0 / 0 / 0 / 0 — / — +4 +2
82 70 639 2026-09-21 0 / 0 / 0 / 0 0 / 0 −5 −11
83 70 658 2026-10-03 0 / −1 / +1 / 0 — / — 0 +3
84 70 850 2026-10-03 0 / 0 / 0 / 0 — / — +1 +10
85 71 98 2026-10-15 0 / 0 / 0 / 0 — / — 0 −14
86 71 851 2026-10-21 0 / 0 / 0 / 0 — / — 0 +17
87 72 12 2026-10-21 0 / 0 / 0 / −2 — / — 0 +10
88 73 782 2026-10-21 +5 / 0 / 0 / 0 — / — 0 −13
89 73 926 2026-10-15 −5 / 0 / 0 / 0 — / — 0 −12
90 74 319 2026-10-03 0 / 0 / 0 / 0 — / — 0 +5
91 74 365 2026-10-03 0 / +2 / +3 / 0 — / — −6 +2
92 74 692 2026-10-03 0 / 0 / 0 / 0 — / — 0 −11
93 74 835 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −14
94 75 1 2026-10-15 0 / 0 / 0 / 0 — / — −8 −7
95 76 140 2026-10-21 +2 / 0 / 0 / 0 — / — +4 +16
96 76 692 2026-10-15 0 / 0 / 0 / 0 — / — 0 0
97 78 126 2026-10-21 0 / 0 / 0 / 0 — / — 0 +4
98 78 527 2026-10-03 0 / 0 / 0 / 0 — / — 0 +5
99 79 927 2026-10-09 0 / 0 / 0 / 0 — / — 0 +13
100 81 651 2026-10-09 −5 / 0 / 0 / 0 — / — −9 0
101 82 609 2026-09-21 0 / 0 / 0 / 0 0 / 0 +3 +12
102 82 815 2026-10-21 0 / 0 / 0 / −5 — / — 0 +9
103 83 919 2026-10-15 0 / 0 / 0 / 0 — / — 0 +1
104 84 120 2026-10-03 0 / 0 / 0 / 0 — / — 0 −11
105 84 653 2026-10-09 −3 / 0 / 0 / 0 — / — +4 −15
106 85 68 2026-10-15 0 / 0 / 0 / 0 — / — −3 −8
107 85 541 2026-10-03 0 / 0 / −1 / 0 — / — +1 −7
108 86 55 2026-10-09 0 / 0 / 0 / 0 — / — −2 −11
109 86 150 2026-10-15 0 / 0 / −1 / 0 — / — −4 −8
110 86 735 2026-10-15 0 / 0 / 0 / 0 — / — +1 −10
111 87 692 2026-10-03 0 / 0 / 0 / 0 — / — −4 −6
112 87 757 2026-10-09 0 / 0 / 0 / 0 — / — +3 +2
113 88 266 2026-10-21 0 / +1 / 0 / 0 — / — 0 −2
114 89 157 2026-10-15 0 / 0 / 0 / 0 — / — −6 +2
115 90 573 2026-10-09 0 / 0 / 0 / 0 — / — 0 +1
116 90 958 2026-10-21 0 / +4 / 0 / 0 — / — 0 −5
117 92 127 2026-10-03 0 / 0 / 0 / 0 — / — 0 0
118 93 81 2026-10-15 0 / 0 / 0 / 0 — / — +1 −17
119 99 28 2026-10-21 0 / −3 / 0 / 0 — / — 0 +5
120 99 164 2026-10-03 0 / 0 / 0 / 0 — / — 0 −12
121 102 577 2026-10-15 −5 / 0 / 0 / 0 — / — +1 +6
122 102 855 2026-10-21 0 / 0 / 0 / 0 — / — +6 0
123 103 115 2026-10-09 0 / 0 / 0 / −2 — / — +5 −1
124 103 596 2026-10-15 +1 / 0 / 0 / 0 — / — 0 +2
125 103 673 2026-10-21 0 / +2 / +2 / 0 — / — −7 +7
126 104 859 2026-10-15 −3 / 0 / +5 / 0 — / — +8 +14
127 106 40 2026-10-03 0 / 0 / 0 / 0 — / — +9 +15
128 106 499 2026-10-09 0 / 0 / 0 / −4 — / — +6 +1
129 107 59 2026-10-21 0 / 0 / 0 / 0 — / — +5 +15
130 107 291 2026-10-21 0 / −5 / 0 / 0 — / — −2 +15
131 107 811 2026-10-03 0 / 0 / 0 / 0 — / — 0 +4
132 108 144 2026-10-15 0 / 0 / −2 / 0 — / — −6 +2
133 108 180 2026-10-15 0 / 0 / 0 / 0 — / — 0 −1
134 109 28 2026-10-15 0 / +1 / 0 / +1 — / — +3 +11
135 109 53 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −1
136 109 283 2026-10-15 −1 / 0 / 0 / +4 — / — −1 0
137 109 667 2026-10-03 0 / 0 / 0 / +4 — / — −1 +3
138 111 107 2026-10-03 −1 / 0 / 0 / 0 — / — +6 −16
139 111 562 2026-10-15 0 / 0 / 0 / 0 — / — −3 −5
140 112 385 2026-10-03 0 / 0 / 0 / 0 — / — 0 −15
141 115 124 2026-10-15 0 / 0 / 0 / 0 — / — +6 +10
142 121 209 2026-10-03 0 / 0 / 0 / 0 — / — 0 +12
143 123 131 2026-10-15 +2 / −1 / 0 / 0 — / — +8 +13
144 123 596 2026-10-21 −3 / −5 / 0 / 0 — / — −1 +8
145 126 150 2026-10-15 0 / 0 / 0 / 0 — / — −2 −3
146 126 501 2026-10-15 0 / 0 / 0 / 0 — / — −5 −8
147 127 224 2026-10-03 −4 / 0 / 0 / 0 — / — 0 +11
148 127 755 2026-10-21 0 / 0 / 0 / 0 — / — 0 +3
149 129 918 2026-10-03 +4 / 0 / 0 / 0 — / — +1 0
150 134 440 2026-10-15 0 / 0 / +4 / 0 — / — 0 +10
151 135 423 2026-10-21 −2 / −2 / 0 / 0 — / — 0 +11
152 136 839 2026-10-15 0 / 0 / 0 / 0 — / — 0 −11
153 138 40 2026-10-03 0 / 0 / 0 / 0 — / — +2 +9
154 138 58 2026-10-09 0 / 0 / 0 / 0 — / — 0 +5
155 139 304 2026-09-21 0 / 0 / 0 / 0 0 / 0 −1 +9
156 140 488 2026-09-21 0 / 0 / 0 / −1 0 / 0 0 +10
157 140 865 2026-10-03 0 / 0 / 0 / 0 — / — −1 +15
158 141 567 2026-10-21 −2 / −5 / 0 / 0 — / — +3 +11
159 144 355 2026-10-21 0 / 0 / 0 / 0 — / — 0 +17
160 144 821 2026-10-15 0 / 0 / 0 / 0 — / — +8 −13
161 146 608 2026-10-21 0 / 0 / 0 / 0 — / — 0 +16
162 147 115 2026-10-21 0 / −4 / −2 / 0 — / — 0 +5
163 148 949 2026-10-21 0 / 0 / 0 / 0 — / — +7 +15
164 149 349 2026-10-21 0 / 0 / 0 / +4 — / — 0 +13
165 149 939 2026-10-15 0 / 0 / 0 / 0 — / — 0 +1
166 149 941 2026-10-09 0 / 0 / 0 / 0 — / — 0 +2
167 150 104 2026-10-03 +2 / 0 / 0 / 0 — / — +2 +13
168 150 421 2026-10-03 −1 / 0 / 0 / 0 — / — −7 0
169 151 613 2026-10-15 0 / 0 / 0 / 0 — / — +1 +2
170 151 896 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −1
171 155 183 2026-10-09 0 / 0 / 0 / 0 — / — 0 −4
172 155 226 2026-10-21 0 / −3 / 0 / 0 — / — +6 +12
173 155 722 2026-10-03 −2 / 0 / 0 / 0 — / — 0 −1
174 158 525 2026-10-15 0 / 0 / 0 / 0 — / — −1 +2
175 158 549 2026-10-15 −4 / 0 / 0 / 0 — / — 0 −1
176 159 244 2026-10-03 0 / 0 / 0 / 0 — / — 0 0
177 159 546 2026-10-15 0 / 0 / 0 / 0 — / — +1 +4
178 161 163 2026-10-03 −1 / 0 / −3 / 0 — / — −9 −5
179 161 283 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +5
180 162 241 2026-10-03 +1 / 0 / 0 / 0 — / — +6 +3
181 164 216 2026-10-15 0 / 0 / 0 / 0 — / — −7 +5
182 164 482 2026-10-21 +3 / 0 / 0 / 0 — / — 0 +15
183 166 533 2026-10-03 0 / 0 / 0 / 0 — / — 0 −1
184 167 236 2026-10-21 0 / +4 / +2 / 0 — / — 0 +8
185 167 543 2026-10-15 0 / 0 / 0 / 0 — / — 0 −6
186 167 578 2026-10-09 0 / 0 / 0 / 0 — / — +6 +1
187 167 691 2026-10-21 0 / 0 / −2 / −4 — / — 0 0
188 168 226 2026-10-03 0 / 0 / 0 / 0 — / — 0 +4
189 169 629 2026-10-21 +1 / 0 / 0 / +4 — / — −1 +3
190 169 771 2026-10-15 0 / 0 / 0 / 0 — / — 0 −4
191 170 370 2026-10-21 +4 / 0 / −1 / 0 — / — +3 0
192 171 745 2026-10-09 0 / 0 / 0 / 0 — / — −1 −5
193 172 244 2026-10-15 −1 / 0 / +1 / 0 — / — 0 +11
194 173 319 2026-10-03 0 / 0 / 0 / 0 — / — 0 −6
195 173 621 2026-10-21 0 / 0 / 0 / 0 — / — 0 −3
196 174 783 2026-10-09 0 / 0 / 0 / 0 — / — 0 −4
197 176 583 2026-10-03 0 / 0 / 0 / 0 — / — 0 −16
198 177 145 2026-09-21 0 / 0 / 0 / 0 0 / 0 −1 +16
199 177 213 2026-10-03 0 / 0 / −1 / 0 — / — −6 +2
200 177 787 2026-10-15 0 / −1 / 0 / 0 — / — −1 −8
201 178 75 2026-10-21 0 / 0 / 0 / 0 — / — 0 +1
202 178 729 2026-10-21 0 / 0 / 0 / 0 — / — +8 0
203 179 496 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +1
204 179 951 2026-10-03 −1 / 0 / 0 / 0 — / — 0 −1
205 180 717 2026-10-15 0 / 0 / 0 / 0 — / — −3 +13
206 181 589 2026-10-03 0 / 0 / 0 / 0 — / — −6 −6
207 182 519 2026-10-21 −5 / +2 / +2 / 0 — / — 0 0
208 182 935 2026-10-15 0 / −1 / 0 / −2 — / — +7 +3
209 183 597 2026-10-21 −2 / −4 / 0 / 0 — / — 0 −7
210 183 899 2026-10-21 −2 / +3 / −3 / 0 — / — −8 +8
211 185 315 2026-10-03 0 / 0 / 0 / 0 — / — 0 +8
212 186 247 2026-09-21 0 / 0 / 0 / 0 0 / 0 +3 0
213 186 545 2026-10-03 0 / 0 / 0 / 0 — / — +3 0
214 186 562 2026-10-09 −4 / 0 / 0 / 0 — / — 0 −5
215 187 558 2026-10-09 0 / 0 / 0 / +4 — / — 0 +1
216 188 216 2026-10-03 0 / 0 / 0 / 0 — / — −9 +6
217 188 753 2026-09-21 0 / 0 / 0 / 0 0 / 0 +6 −11
218 190 67 2026-10-09 +1 / 0 / 0 / 0 — / — −8 −7
219 190 943 2026-10-15 0 / 0 / 0 / 0 — / — 0 +1
220 191 50 2026-10-21 +4 / −4 / 0 / 0 — / — −6 −17
221 191 496 2026-10-03 0 / 0 / 0 / 0 — / — +6 −3
222 193 113 2026-10-03 0 / 0 / 0 / 0 — / — 0 −7
223 194 99 2026-10-21 +5 / +2 / 0 / 0 — / — +7 +4
224 194 148 2026-10-03 0 / 0 / 0 / 0 — / — −1 +4
225 194 899 2026-10-03 0 / 0 / 0 / −5 — / — 0 +15
226 196 195 2026-10-09 0 / 0 / 0 / 0 — / — 0 −12
227 197 784 2026-10-03 0 / 0 / 0 / 0 — / — 0 −9
228 198 325 2026-10-09 −5 / 0 / 0 / 0 — / — −1 +9
229 198 722 2026-10-21 +5 / −4 / 0 / 0 — / — 0 +4
230 199 291 2026-10-15 −2 / −1 / 0 / 0 — / — 0 +4
231 199 691 2026-10-15 0 / 0 / +2 / 0 — / — −3 −10
232 201 768 2026-10-15 0 / 0 / 0 / 0 — / — −7 −6
233 202 270 2026-10-03 −1 / 0 / 0 / 0 — / — 0 −9
234 202 657 2026-10-03 0 / 0 / −4 / 0 — / — +4 −10
235 202 931 2026-10-15 0 / −1 / +2 / 0 — / — −5 −8
236 204 935 2026-10-15 0 / 0 / 0 / 0 — / — −3 −9
237 205 870 2026-10-03 +1 / 0 / 0 / 0 — / — 0 +1
238 206 646 2026-10-15 −3 / +1 / 0 / 0 — / — 0 +1
239 207 96 2026-10-15 0 / 0 / 0 / 0 — / — 0 +13
240 207 136 2026-10-09 0 / 0 / 0 / 0 — / — 0 +13
241 207 514 2026-10-03 0 / 0 / 0 / 0 — / — +3 +7
242 207 839 2026-10-09 0 / 0 / 0 / 0 — / — +1 −9
243 210 435 2026-10-03 0 / 0 / 0 / 0 — / — +1 +8
244 213 212 2026-10-15 0 / 0 / 0 / 0 — / — 0 0
245 213 403 2026-10-21 0 / +4 / +2 / 0 — / — 0 +11
246 213 731 2026-10-03 0 / 0 / 0 / 0 — / — 0 −13
247 214 356 2026-10-03 −3 / 0 / 0 / 0 — / — 0 −1
248 214 412 2026-10-21 0 / −2 / −1 / 0 — / — 0 −11
249 216 192 2026-10-03 0 / 0 / 0 / 0 — / — 0 +6
250 221 518 2026-10-15 +5 / +1 / 0 / 0 — / — −3 +13
251 221 572 2026-10-15 0 / 0 / 0 / 0 — / — 0 +14
252 222 279 2026-10-15 0 / 0 / 0 / 0 — / — 0 −1
253 223 208 2026-10-09 0 / 0 / 0 / 0 — / — 0 +10
254 226 179 2026-10-15 −1 / 0 / 0 / −3 — / — −9 0
255 226 243 2026-10-03 0 / 0 / 0 / 0 — / — −5 −4
256 227 483 2026-10-15 +2 / 0 / 0 / 0 — / — 0 −9
257 229 162 2026-10-03 0 / 0 / 0 / 0 — / — 0 +10
258 229 687 2026-10-09 0 / 0 / 0 / −2 — / — 0 −16
259 229 812 2026-10-15 0 / 0 / 0 / −1 — / — −3 −16
260 229 849 2026-10-09 0 / 0 / 0 / 0 — / — 0 −3
261 230 43 2026-10-03 +2 / 0 / 0 / 0 — / — 0 −3
262 230 214 2026-10-03 0 / 0 / 0 / 0 — / — 0 +10
263 230 306 2026-10-03 0 / 0 / 0 / 0 — / — −5 +3
264 230 447 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −6
265 231 187 2026-10-03 0 / 0 / 0 / 0 — / — +2 +8
266 232 3 2026-10-09 0 / 0 / 0 / 0 — / — 0 −4
267 232 839 2026-10-15 +4 / 0 / 0 / 0 — / — 0 −7
268 233 866 2026-10-15 0 / 0 / 0 / +1 — / — 0 +8
269 237 87 2026-10-21 0 / +4 / 0 / 0 — / — +6 +15
270 237 446 2026-10-15 0 / 0 / 0 / 0 — / — −6 −3
271 237 569 2026-10-09 0 / 0 / 0 / 0 — / — 0 −16
272 239 68 2026-10-03 0 / 0 / 0 / 0 — / — 0 +2
273 240 211 2026-10-15 0 / 0 / 0 / 0 — / — −8 +2
274 241 535 2026-10-21 0 / 0 / 0 / 0 — / — +6 +17
275 241 543 2026-10-21 0 / 0 / 0 / 0 — / — 0 −14
276 241 596 2026-10-21 0 / 0 / 0 / 0 — / — 0 −3
277 241 613 2026-10-15 0 / −1 / 0 / 0 — / — 0 +3
278 241 916 2026-10-21 0 / 0 / 0 / 0 — / — 0 0
279 242 179 2026-10-15 0 / 0 / 0 / 0 — / — +1 −5
280 242 267 2026-10-21 0 / 0 / −5 / −2 — / — +8 +12
281 242 354 2026-10-21 +5 / 0 / 0 / −3 — / — +2 +7
282 242 516 2026-10-03 0 / 0 / −1 / −1 — / — −2 +4
283 243 103 2026-10-03 0 / 0 / 0 / 0 — / — −6 +3
284 246 705 2026-10-09 +3 / 0 / 0 / −4 — / — 0 −2
285 247 255 2026-10-03 0 / 0 / 0 / 0 — / — +1 +4
286 247 737 2026-10-21 0 / 0 / 0 / 0 — / — 0 +16
287 247 784 2026-10-15 0 / 0 / 0 / 0 — / — 0 −5
288 248 647 2026-10-15 0 / 0 / 0 / 0 — / — 0 −2
289 249 647 2026-10-15 0 / 0 / 0 / 0 — / — 0 +8
290 250 533 2026-10-09 0 / 0 / 0 / 0 — / — +4 +12
291 252 405 2026-10-15 −1 / +1 / +4 / 0 — / — +1 −16
292 253 61 2026-10-21 0 / −5 / 0 / 0 — / — 0 +10
293 254 897 2026-10-09 0 / 0 / 0 / 0 — / — 0 −7
294 255 53 2026-10-09 0 / 0 / 0 / −3 — / — 0 +3
295 255 134 2026-10-03 0 / 0 / 0 / 0 — / — −6 −1
296 255 396 2026-10-21 0 / 0 / 0 / 0 — / — +8 +12
297 255 947 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +14
298 257 43 2026-10-21 0 / −3 / −1 / 0 — / — +1 −13
299 257 353 2026-10-15 0 / 0 / 0 / +3 — / — −4 +4
300 258 119 2026-10-09 0 / 0 / 0 / 0 — / — −5 −11
301 258 834 2026-10-09 −4 / 0 / 0 / 0 — / — 0 −3
302 261 630 2026-10-09 0 / 0 / 0 / 0 — / — −1 −11
303 263 737 2026-10-21 0 / −5 / 0 / +4 — / — 0 +17
304 264 817 2026-10-21 0 / −1 / −2 / 0 — / — 0 −6
305 265 21 2026-10-21 0 / 0 / −2 / 0 — / — 0 −10
306 266 286 2026-10-03 0 / 0 / +2 / +5 — / — 0 0
307 267 672 2026-10-09 0 / 0 / 0 / 0 — / — 0 −4
308 269 914 2026-10-03 0 / 0 / 0 / 0 — / — +1 −8
309 270 45 2026-10-15 0 / 0 / 0 / 0 — / — −1 +16
310 271 831 2026-10-15 0 / 0 / 0 / 0 — / — 0 +5
311 272 705 2026-10-09 0 / 0 / 0 / 0 — / — 0 +14
312 272 861 2026-10-15 0 / 0 / 0 / 0 — / — +6 +15
313 273 42 2026-10-15 0 / 0 / 0 / +5 — / — 0 −13
314 273 541 2026-10-21 0 / 0 / 0 / 0 — / — 0 +15
315 273 611 2026-10-03 0 / 0 / 0 / −5 — / — −7 +9
316 275 403 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −2
317 277 904 2026-10-21 0 / 0 / 0 / 0 — / — +8 −2
318 279 596 2026-10-15 0 / 0 / 0 / 0 — / — 0 −7
319 279 845 2026-10-15 0 / 0 / 0 / 0 — / — 0 +4
320 280 21 2026-10-15 0 / 0 / 0 / 0 — / — 0 −15
321 281 306 2026-10-03 0 / 0 / 0 / 0 — / — +6 +7
322 281 502 2026-10-03 0 / 0 / 0 / 0 — / — 0 −6
323 281 683 2026-10-21 0 / −4 / 0 / 0 — / — −4 −16
324 282 692 2026-10-15 0 / 0 / 0 / 0 — / — +6 −13
325 284 406 2026-10-03 0 / 0 / 0 / 0 — / — 0 −7
326 284 764 2026-10-09 0 / 0 / 0 / 0 — / — 0 −17
327 287 401 2026-10-21 −2 / 0 / 0 / 0 — / — 0 +12
328 287 847 2026-10-15 0 / 0 / 0 / 0 — / — −9 +4
329 289 97 2026-10-03 0 / 0 / 0 / 0 — / — 0 +14
330 291 405 2026-10-15 0 / 0 / 0 / 0 — / — −1 −6
331 293 448 2026-10-15 0 / 0 / 0 / 0 — / — −2 −14
332 295 123 2026-10-15 0 / 0 / −2 / 0 — / — 0 0
333 295 139 2026-10-09 0 / 0 / 0 / −1 — / — +1 0
334 295 751 2026-10-09 0 / 0 / 0 / 0 — / — 0 −14
335 299 567 2026-10-03 −3 / 0 / 0 / 0 — / — 0 −5
336 304 386 2026-10-15 0 / 0 / 0 / 0 — / — +1 −7
337 306 733 2026-10-15 0 / 0 / 0 / 0 — / — +7 −6
338 307 759 2026-10-03 0 / 0 / 0 / 0 — / — 0 +4
339 310 868 2026-10-03 −1 / 0 / 0 / 0 — / — −6 −1
340 313 625 2026-10-09 0 / 0 / 0 / 0 — / — +8 0
341 314 740 2026-10-03 0 / 0 / 0 / 0 — / — 0 −2
342 317 79 2026-10-03 +2 / 0 / 0 / 0 — / — 0 +2
343 317 198 2026-10-03 0 / 0 / 0 / 0 — / — 0 +17
344 318 524 2026-10-15 +2 / +1 / 0 / 0 — / — +8 −10
345 318 624 2026-10-15 0 / 0 / −3 / −1 — / — 0 −9
346 320 128 2026-10-03 0 / 0 / 0 / 0 — / — 0 −17
347 321 401 2026-10-15 0 / 0 / −4 / 0 — / — −3 −5
348 321 672 2026-10-15 0 / 0 / 0 / 0 — / — −1 −17
349 321 901 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 −17
350 325 781 2026-10-15 0 / 0 / 0 / 0 — / — +1 −6
351 326 486 2026-10-21 −1 / 0 / 0 / 0 — / — 0 +9
352 326 635 2026-10-21 0 / 0 / 0 / 0 — / — 0 −11
353 331 163 2026-10-09 −2 / 0 / 0 / +3 — / — −5 0
354 334 578 2026-10-03 0 / 0 / 0 / 0 — / — +9 −14
355 335 143 2026-09-21 0 / 0 / 0 / 0 0 / 0 +1 +11
356 335 547 2026-10-03 0 / 0 / 0 / 0 — / — 0 −17
357 335 959 2026-10-03 0 / 0 / 0 / 0 — / — 0 +2
358 389 96 2026-10-03 −4 / 0 / 0 / 0 — / — +4 +16
359 390 516 2026-10-09 +4 / 0 / 0 / 0 — / — +6 +12
360 391 119 2026-10-03 0 / 0 / 0 / 0 — / — +1 −2
361 391 547 2026-10-21 0 / 0 / 0 / 0 — / — −1 +11
362 396 423 2026-10-03 0 / 0 / 0 / 0 — / — −1 −7
363 398 466 2026-10-09 0 / 0 / 0 / +1 — / — −5 +15
364 407 755 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +7
365 411 26 2026-10-03 0 / 0 / 0 / 0 — / — −1 −5
366 412 44 2026-10-03 0 / 0 / 0 / 0 — / — 0 +3
367 414 415 2026-10-03 0 / 0 / 0 / 0 — / — +3 0
368 415 309 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +15
369 418 54 2026-10-21 +3 / 0 / 0 / 0 — / — 0 +8
370 424 401 2026-10-03 +1 / 0 / 0 / 0 — / — 0 +6
371 426 119 2026-10-21 0 / 0 / 0 / 0 — / — −3 −7
372 427 140 2026-10-21 −5 / 0 / 0 / 0 — / — +4 −16
373 429 180 2026-10-09 0 / 0 / +1 / 0 — / — +8 +10
374 429 754 2026-10-09 +3 / 0 / 0 / 0 — / — −7 −11
375 430 536 2026-10-03 0 / 0 / 0 / 0 — / — +6 −3
376 432 835 2026-10-03 0 / 0 / 0 / 0 — / — 0 −15
377 433 571 2026-10-09 0 / 0 / 0 / 0 — / — 0 +3
378 436 692 2026-10-21 0 / 0 / 0 / +5 — / — +1 +11
379 437 176 2026-10-21 0 / 0 / 0 / 0 — / — 0 +13
380 443 340 2026-10-03 +1 / 0 / −2 / 0 — / — 0 +7
381 444 637 2026-10-03 0 / 0 / 0 / 0 — / — 0 −14
382 446 422 2026-10-03 0 / 0 / 0 / 0 — / — 0 −3
383 447 4 2026-10-21 0 / 0 / 0 / 0 — / — 0 +10
384 449 19 2026-10-03 0 / 0 / 0 / 0 — / — −6 −9
385 451 844 2026-10-21 0 / 0 / 0 / 0 — / — 0 0
386 452 897 2026-10-03 +3 / 0 / 0 / 0 — / — −5 +1
387 453 242 2026-10-21 0 / 0 / 0 / 0 — / — +1 −9
388 453 396 2026-10-21 0 / 0 / +1 / 0 — / — 0 +2
389 454 318 2026-10-03 0 / 0 / 0 / 0 — / — +1 +8
390 455 900 2026-10-21 0 / 0 / 0 / 0 — / — 0 +6
391 456 241 2026-10-21 0 / +1 / +1 / +4 — / — −1 +16
392 457 505 2026-10-03 0 / 0 / 0 / 0 — / — −2 +9
393 458 335 2026-10-09 0 / 0 / 0 / 0 — / — 0 +14
394 459 719 2026-10-03 0 / +3 / +3 / 0 — / — +1 +16
395 460 707 2026-10-09 0 / 0 / 0 / −2 — / — +1 +2
396 461 241 2026-10-09 0 / 0 / 0 / 0 — / — 0 +3
397 461 268 2026-10-09 0 / 0 / +2 / 0 — / — +9 0
398 461 722 2026-10-03 0 / 0 / 0 / −2 — / — 0 +5
399 461 899 2026-10-21 +4 / 0 / 0 / 0 — / — 0 −6
400 468 49 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +8
401 468 164 2026-10-09 0 / 0 / 0 / +4 — / — −6 +1
402 469 514 2026-10-03 0 / 0 / 0 / 0 — / — −1 −3
403 470 887 2026-10-21 0 / 0 / 0 / 0 — / — +5 +1
404 474 175 2026-10-09 −5 / 0 / 0 / 0 — / — 0 −3
405 475 467 2026-10-21 0 / 0 / 0 / 0 — / — +1 +7
406 475 672 2026-10-21 +1 / 0 / 0 / 0 — / — 0 +2
407 478 54 2026-10-09 0 / 0 / 0 / 0 — / — 0 −1
408 479 675 2026-10-03 0 / 0 / 0 / 0 — / — 0 0
409 482 130 2026-10-21 +3 / 0 / 0 / −5 — / — 0 −6
410 483 387 2026-10-21 −3 / 0 / 0 / 0 — / — −3 +14
411 483 750 2026-10-09 0 / 0 / 0 / 0 — / — 0 +6
412 485 2 2026-10-21 −3 / 0 / 0 / 0 — / — 0 −2
413 485 768 2026-10-21 +2 / 0 / 0 / 0 — / — 0 +1
414 488 466 2026-10-03 −1 / 0 / 0 / 0 — / — 0 +8
415 488 640 2026-10-21 −2 / 0 / 0 / 0 — / — 0 +8
416 491 12 2026-10-09 0 / 0 / 0 / 0 — / — +5 +8
417 498 144 2026-09-21 0 / 0 / 0 / 0 0 / 0 0 +10
418 502 423 2026-10-03 0 / 0 / 0 / 0 — / — −2 −7
419 502 870 2026-10-21 0 / 0 / 0 / 0 — / — 0 −11
420 503 498 2026-10-03 +1 / 0 / 0 / −1 — / — 0 +16
421 503 754 2026-10-21 0 / 0 / 0 / 0 — / — −9 −2
422 504 753 2026-10-21 +5 / 0 / 0 / 0 — / — +3 −12
423 505 641 2026-10-21 0 / 0 / −5 / 0 — / — 0 +3
424 507 907 2026-10-09 +4 / 0 / 0 / 0 — / — 0 0
425 510 955 2026-10-03 0 / 0 / −1 / 0 — / — 0 −4
426 514 144 2026-10-21 −2 / 0 / +5 / +5 — / — −4 +7
427 514 226 2026-10-21 +2 / 0 / 0 / 0 — / — 0 0
428 516 720 2026-10-21 −4 / 0 / 0 / 0 — / — 0 +16
429 517 316 2026-10-21 0 / 0 / 0 / 0 — / — −8 +6
430 518 260 2026-10-09 0 / 0 / 0 / 0 — / — −1 +10
431 518 884 2026-10-09 0 / 0 / 0 / 0 — / — 0 +6
432 519 424 2026-10-03 0 / 0 / 0 / 0 — / — 0 −6
433 519 595 2026-09-21 0 / 0 / 0 / 0 0 / 0 −2 +2
434 522 622 2026-10-03 0 / 0 / −3 / 0 — / — 0 +3
435 524 739 2026-10-03 0 / 0 / 0 / 0 — / — −6 −17
436 526 591 2026-10-03 −4 / 0 / 0 / 0 — / — 0 −14

Table A1. The published set: Stockfish evaluations along the ladder (cp, White's point of view), Berserk 14 at 400 M nodes, Lc0 0.32.1 at 100 k tree nodes.

Appendix B. Former pairs

White Black Failed 400 M / 800 M / 1.6 G / 4 G Berserk Lc0
7 676 C D 0 / 0 / 0 / 0 +16 +30
25 902 D 0 / 0 / 0 / 0 0 −19
33 930 A 0 / 0 / 0 / 0 0 +9
39 571 A 0 / 0 / 0 / 0 0 −1
41 928 D 0 / 0 / 0 / 0 +5 +21
61 620 A C 0 / 0 / 0 / 0 −17 −12
66 138 A B C D 0 / 0 / 0 / −15 −17 −24
73 705 A 0 / 0 / 0 / −4 0 +1
85 256 A 0 / 0 / 0 / +6 0 +12
106 955 D 0 / 0 / 0 / 0 0 +18
143 885 A D 0 / 0 / 0 / 0 +8 +24
188 896 A B 0 / 0 / 0 / −14 0 +3
196 480 C 0 / 0 / 0 / 0 −12 −2
198 819 D 0 / 0 / 0 / 0 −8 −24
223 767 D 0 / 0 / 0 / 0 0 −29
303 48 D 0 / 0 / 0 / 0 0 −20
303 823 A D 0 / 0 / 0 / 0 0 −26
317 934 A B C 0 / 0 / 0 / −12 −13 +1
328 184 A 0 / 0 / 0 / −5 0 +4
335 829 D 0 / 0 / 0 / 0 0 −20
394 672 A 0 / 0 / 0 / 0 0 +13
466 726 C 0 / 0 / 0 / 0 +12 +9
473 353 A 0 / 0 / 0 / +6 0 +6
495 851 A C 0 / 0 / 0 / −4 −15 +2

Table B1. Pairs of the first set removed by criterion v2 on 5 October 2026, with the conditions they fail.

Appendix C. Revision history

Publication on ch960.com

Published: 21 October 2026.

Permanent address: https://ch960.com/ar/asym/paper

PDF: https://ch960.com/asym/paper/balanced-dfrc-pairs-en.pdf

SHA-256 of the PDF: 23db2735d035f11da7157ba30b949259d026781552eaf22dca462d3846422151

DOI (the same files on Zenodo): 10.5281/zenodo.23536049

Same text in the other language: Russian version

CC BY 4.0 — the text, tables and figures may be quoted and shared with attribution to the author, Roman Morozov, and a link to this page.

How to cite: Morozov, Roman. Balanced start pairs in Double Fischer Random Chess: a computational search with multi-engine verification. Version 1.2. CH960, 2026. https://ch960.com/ar/asym/paper. DOI: 10.5281/zenodo.23536049

Copies in the Internet Archive: this page, PDF.

Every number in the text is computed by a script from the published data: the search journal, the verification runs, the engine passports and the composition with the reasons for every removal. The live state is on the encyclopedia and methodology pages.

موسوعة الأزواج المتوازنة · الأوضاع غير المتماثلة: المنهجية · كل الأوضاع غير المتماثلة

العب Chess960

في المتصفح أو في تطبيق Android. لا حاجة للتثبيت · مجاني، بلا إعلانات