Runs
Every register entry, newest first. Short and long stage are labelled separately; the quoted value is the long stage when there is one and is never added to or averaged with the short stage. Verdicts without games name their evidence. This table is not a highlight reel.
Register export 2026-10-30 · Register T-0001 – T-0227 · 225 entries · Run feed (RSS)
How each run ended
Register T-0166 – T-0106
| Run | Verdict | 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. | Date | Games | 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. | Hypothesis |
|---|---|---|---|---|---|---|
| T-0166 | rejected | short stage −0.34 ± 2.31 | 22,462 | −2.95 | Exempting non-promotion captures of opponent passed pawns on the sixth or seventh rank from capture futility pruning and capture SEE Static exchange evaluation: an estimate of the material outcome of a capture sequence on one square, without searching. pruning in the move loop allows the engine to search these direct promotion defenses more completely. This change is expected to improve playing strength at fixed node counts. | |
| T-0165 | rejected | short stage −2.74 ± 3.67 | 9,144 | −2.96 | If a parent node tracks a child node's reduction, and the child adjusts its search depth accordingly—increasing it by one if it was heavily reduced and the move did not statically weaken the moving side, or decreasing it by one if both static evaluations show the move caused significant harm—the engine will play stronger under time controls. | |
| T-0164 | rejected | long stage −0.44 ± 2.07 short stage +8.63 ± 4.56 | 20,386 | −2.97 | Initializing the accumulator with the input bias rather than zero, so network output heads across standard, AVX2, AVX-512, and single-layer implementations no longer need to add it, will leave the search tree identical while increasing engine speed. | |
| T-0163 | rejected | short stage −2.44 ± 3.53 | 9,816 | −2.95 | If the margin reductions for rising evaluations and weaker opponents in reverse futility pruning are only applied when none of the engine's knights or bishops are attacked by an enemy pawn, no rook by a pawn or minor piece, and no queen by a pawn, minor piece, or rook, while the full margin pruning as well as both signals in LMP and LMR remain unchanged, the engine plays stronger. | |
| T-0162 | passed | long stage +5.49 ± 3.69 short stage +3.49 ± 2.72 | 8,292 | +2.96 | Maintaining a dedicated table solely for quiet root moves that is cleared at the start of each search, tracks the best move separately from failed attempts, and is used to reorder quiet root moves after completed iterations improves engine playing strength at fixed node counts, assuming tree shape and reductions below the root remain unchanged. | |
| T-0161 | passed | long stage +5.83 ± 3.87 short stage +3.07 ± 2.44 | 7,990 | +2.97 | Adding move pruning rules in the quiescence search increases playing strength at fixed node counts. Outside of check, captures without promotion, check, or recapture are skipped once a legal move limit is reached, provided the best score exceeds the mate threshold. In check, quiet non-promotion evasions are skipped once the best score exceeds the mate threshold, while captures and promotions remain searched. | |
| T-0160 | rejected | short stage −18.10 ± 7.93 | 1,864 | −2.95 | In zero-window nodes at nominal remaining depths 1 or 2 with non-pawn material and not in check, replacing the global LMP cutoff with a move-wise pruning of quiet non-promotion moves increases playing strength. Pruning applies when the move count reaches at least max(2, lmp_limit(depth, improving) + H/2048), where H is purely the direct main history value prior to the move indexed by origin and destination square threats, truncated toward zero. | |
| T-0159 | passed | long stage +4.05 ± 2.89 short stage +1.55 ± 1.13 | 10,294 | +2.95 | Eliminating redundant initializations of the board squares, occupancy, piece kinds, check info, and the exchange balance on the move path avoids microcoded rep stosq instructions emitted by GCC, as every value is written before being read. This change is expected to yield identical search trees while increasing calculation speed and playing strength under clock conditions. | |
| T-0158 | passed | long stage +9.58 ± 5.23 short stage +2.75 ± 2.21 | 4,390 | +2.96 | Ordering non-promotion good captures with a capture history below -2000 after both killer moves and before the remaining quiet moves in main search when not in check will cause more nodes to fail high on the first move and increase playing strength. | |
| T-0157 | passed | long stage +22.39 ± 8.18 short stage +27.07 ± 9.19 | 1,694 | +2.95 | Setting the parameters RfpFaktor to 50 (down from 68), VerlaufGrenze to 12405 (down from 15976), and SingularFaktor to 65 (down from 89) increases the engine's playing strength under time controls compared to the baseline version. | |
| T-0156 | passed | epoch change, no games | A network epoch change based on the same training source is evaluated. | |||
| T-0155 | rejected | long stage +0.36 ± 1.71 short stage +3.44 ± 2.69 | 40,178 | −2.96 | If the logarithmic late move reduction reduces a late move by one depth less when the transposition table entry of its child does not clearly place it below alpha, the engine plays stronger under time controls. | |
| T-0154 | passed | long stage +4.02 ± 2.92 short stage +5.07 ± 3.46 | 11,316 | +2.96 | Adding a refresh cache per perspective and bucket of the neural network avoids fully reconstructing from the board when the bucket changes, even across horizontal reflection. If this cache counts against the hash allocation, the engine is expected to play stronger on the clock while producing the same search tree. | |
| T-0153 | passed | long stage +168.30 ± 18.08 short stage +175.12 ± 18.46 | 438 | +2.95 | Adopting the engine with its rewritten search, tempo components, and 75 constants tuned via SPSA in self-play, combined with all subsequent improvements from the main development line including the updated evaluation network, is expected to play stronger both under clock conditions and in oracle analysis. | |
| T-0152 | rejected | long stage +0.02 ± 2.06 short stage +5.15 ± 3.67 | 29,192 | −2.97 | Adding three pruning rules to the main search improves playing strength against main: SEE pruning for main search moves up to remaining depth 8 with margins of 55 for quiet moves and 95 for captures, scaled linearly with depth; reverse futility pruning up to remaining depth 7; and null-move verification only from remaining depth 12 onwards. Razoring and SEE pruning in quiescence search are omitted. | |
| T-0151 | passed | long stage +5.99 ± 3.98 short stage +11.07 ± 5.88 | 7,890 | +2.95 | Decreasing the LMR reduction of a late move by one depth if the child's transposition table entry is not clearly below alpha (by at least 25) will increase engine strength in timed games (8+0.08). | |
| T-0150 | rejected | short stage −2.15 ± 3.63 | 11,328 | −2.95 | Retaining quiet move history across the entire game rather than clearing it before each search, combined with adding continuation history for moves one and two plies prior, increases engine playing strength compared to the baseline. | |
| T-0149 | rejected | short stage −5.08 ± 4.85 | 6,152 | −2.95 | Adding three pruning rules to the main search without adopting the baseline razoring will increase the engine's playing strength against the base version. The rules added are SEE pruning for main search moves, reverse futility pruning up to depth 7, and null-move verification starting at depth 12. | |
| T-0148 | rejected | short stage −1.33 ± 3.15 | 14,582 | −2.95 | In null-window nodes, negamax extends the hash move by two plies instead of one if the singular search stays below singular_beta minus 20 cp and a per-line inherited counter, incremented only on such double extensions, remains below 6. This is expected to make the engine play stronger. | |
| T-0147 | rejected | short stage −1.53 ± 3.25 | 13,608 | −2.96 | Granting the check extension in the parent node to the checking move rather than to the node in check ensures that a check sacrifice no longer receives an additional ply. This change is expected to increase the engine's playing strength against main. | |
| T-0146 | passed | long stage +10.97 ± 5.58 short stage +17.89 ± 7.48 | 3,612 | +2.96 | Caching the most recently computed accumulator sum and board state per perspective and entry (bucket plus mirror bit), updating only changed squares on entry transitions, and continuing batched updates otherwise preserves the exact search tree at fixed nodes. This optimization makes the engine faster and therefore stronger under timed conditions. | |
| T-0145 | rejected | short stage −18.07 ± 8.38 | 2,078 | −2.95 | A correction history for the static evaluation, indexed by color and the hash of the bitboard of all capturable pieces, increases playing strength in negamax. | |
| T-0144 | passed | long stage +35.15 ± 10.95 short stage +26.67 ± 9.79 | 1,200 | +2.95 | Using the multi-layer neural network SKML768V7J (16/32 structure, four king buckets with a symmetric board and mirror bit, horizontal king mirroring, trained on v7j data for 16 epochs) increases the engine's playing strength against main with network KM768V6 under timed conditions. | |
| T-0143 | passed | long stage +12.96 ± 6.32 short stage +13.40 ± 6.69 | 3,298 | +2.96 | Adopting the search parameters tuned via SPSA under timed conditions (8+0.08 in self-play), including RFP, futility, LMR, and ProbCut margins, as fixed constants increases the engine's playing strength against the baseline under timed conditions. | |
| T-0142 | passed | long stage +7.12 ± 4.46 short stage +6.56 ± 4.35 | 6,542 | +2.97 | If negamax prunes late, non-checking quiet moves before make_move whenever their computed LMR reduction r satisfies r > 0 and depth - r <= FutilityMaxDepth, provided original_alpha is outside the mate range, pruning_eval + FutilityMargin * (depth - r) <= original_alpha, and the side to move has non-pawn material, the engine will gain playing strength. All other LMR reductions and the existing node-level futility pruning up to depth 2 remain unchanged. | |
| T-0141 | passed | long stage +10.32 ± 5.57 short stage +11.18 ± 6.02 | 4,244 | +2.95 | Splitting the quiet-move butterfly history into four tables based on whether the opponent attacks the from-square and whether they attack the to-square, using an opponent attack set built once per node before the move loop for updates and ordering, yields stronger play under clock conditions than the single existing table. | |
| T-0140 | rejected | short stage −1.52 ± 3.10 | 13,232 | −2.96 | If a search iteration is aborted due to node or time limits, the engine plays the move that, after a full re-search as a non-first root move, scored above the first move and below beta, rather than falling back to the previous iteration's best move. Score, depth, and PV Principal variation: the line of best moves the search currently expects. reported to the GUI remain those of the previous iteration. This change is expected to increase playing strength. | |
| T-0139 | rejected | short stage +0.04 ± 2.30 | 29,374 | −2.97 | Late quiet check moves are no longer exempt from LMR, but are reduced by one ply less than an equivalent quiet move, without dropping below zero. Check extensions in child nodes, move thresholds, base reductions, and full research on alpha raises remain unchanged. This change is expected to increase engine strength. | |
| T-0138 | passed | long stage +6.82 ± 4.23 short stage +3.54 ± 2.77 | 6,316 | +2.96 | Checking the legality of only the hash move before a singular extension in negamax—by searching the pseudo-legal move list and executing a single make_move/in_check/unmake_move instead of building board.legal_moves()—increases the node rate while maintaining an identical search tree. | |
| T-0137 | rejected | short stage −6.95 ± 5.58 | 4,952 | −2.98 | Playing strength increases against main by adjusting quiet move ordering scores when pieces flee or enter threats from lesser-valued opponent pieces. Quiet moves of a queen, rook, or minor piece escaping such a threat to a safe square receive fixed bonuses of HistoryLimit/2, 3*HistoryLimit/8, or HistoryLimit/4 respectively, while moves stepping into such threats receive equivalent penalties. The sum of history and adjustment is capped at ±HistoryLimit and also affects LMR. | |
| T-0136 | rejected | short stage −20.13 ± 8.86 | 1,918 | −2.96 | Without an explicit UCI Universal Chess Interface, the text protocol between a chess engine and the program that runs it. movestogo parameter, the time management allocates the budget across a fixed horizon of 20 moves instead of using the move-number-dependent estimator. | |
| T-0135 | passed | long stage +50.03 ± 12.78 short stage +54.00 ± 14.07 | 860 | +2.95 | A network of width 768 featuring four king buckets and a multi-layer head (FT x 2 -> 16 -> 32 -> 1), trained for 8 epochs on the v6 dataset (4.93 billion positions), plays stronger on the clock than the previous 768-wide v5 network trained for 16 epochs, and thus stronger than main. | |
| T-0134 | passed | long stage +5.80 ± 3.92 short stage +6.40 ± 4.31 | 8,268 | +2.95 | Transposition table scores refine the static evaluation for forward pruning heuristics. When a usable, non-mate entry has bounds aligned in the proper direction, reverse futility pruning, null move pruning including reduction, and futility pruning substitute the table score for the static evaluation. Improving, static evals, and hash entries continue to use the raw evaluation, and behavior in check remains unaffected. | |
| T-0133 | passed | long stage +4.58 ± 3.34 short stage +6.63 ± 4.39 | 11,302 | +2.95 | Starting at depth eight after a complete, stable iteration, the node fraction of the best root move scales the soft time target using clamp(2.15 - 1.5 * fraction; 0.5; 2.0). Below depth eight the factor remains 1, the hard time limit stays unchanged, and searches under go nodes or go depth are unaffected. | |
| T-0132 | rejected | short stage −9.93 ± 6.30 | 3,464 | −2.97 | If every completed iteration stops the search as soon as the elapsed time plus the duration of the last iteration reaches the soft time target, even after a best move change or an evaluation drift over 50 cp, instead of relying solely on the hard limit of three times the soft target in these cases, the engine plays stronger than the baseline across all time controls. | |
| T-0131 | passed | long stage +30.48 ± 10.08 short stage +14.96 ± 7.24 | 1,394 | +2.97 | A network with a width of 768, trained on 3.13 billion positions over 16 epochs with a paired AVX2 head, is expected to play stronger under time controls than the baseline network of width 512. | |
| T-0130 | rejected | short stage −0.51 ± 2.70 | 20,992 | −2.95 | The existing move count pruning for late quiet moves, which uses a threshold of LateMoveIndex plus remaining depth squared (12 moves at depth 3 and 19 moves at depth 4), is extended to apply up to remaining depth four instead of depth two. The exceptions remain unchanged, covering zero window, no check, no mate window, pieces on the board, and quiet checks being kept. The engine is expected to play stronger against main. | |
| T-0129 | interrupted | long stage +2.06 ± 1.66 short stage +3.57 ± 2.82 | 48,000 | +2.71 | Increasing the base null move reduction by one ply via the summand in null_move_reduction, while leaving the NullMoveBaseReduction constant and thus the evaluation step of 200 and the cap of 2 unchanged, yields Elo gains against main. | |
| T-0127 | passed | long stage +38.60 ± 11.75 short stage +29.02 ± 10.27 | 1,166 | +2.95 | A network with a width of 512, trained for 8 epochs with a learning rate schedule decaying to zero per step on 3.13 billion rows and paired with an AVX2 head, plays stronger under time controls than the baseline version using a network width of 384. | |
| T-0126 | passed | long stage +12.43 ± 6.16 short stage +17.01 ± 7.62 | 3,384 | +2.96 | A singular extension extends the search depth for the TT Transposition table: a cache of already searched positions, keyed by a hash of the position. move if a reduced exclusion search determines that all alternative moves are significantly worse. | |
| T-0125 | rejected | short stage +0.58 ± 1.57 | 51,442 | −2.96 | Adding a deterministic depth-10 node fingerprint makes TT efficiency changes observable that are missed by shallow benchmarks and best-move or score comparisons. | |
| T-0124 | passed | long stage +58.45 ± 14.72 short stage +54.64 ± 15.01 | 840 | +2.97 | An N384 network with width 384 and a paired AVX2 head, trained over 8 epochs with a learning rate smoothly decaying to zero per step, increases playing strength compared to the baseline using a P256 network when combined with batched accumulator updates that apply all NNUE Efficiently updatable neural network: a small network that evaluates positions and is updated incrementally move by move. changes of a move in a single pass while preserving the search tree. | |
| T-0123 | passed | long stage +6.54 ± 4.21 short stage +15.97 ± 7.34 | 7,176 | +2.97 | A speed-up package increases engine execution speed without changing the search tree. It introduces an exact AVX2 network head with fixed neuron pairing via an identically permuted network, alongside accumulator updates integrated into piece placement through the existing CPU dispatch. | |
| T-0122 | passed | long stage +11.51 ± 6.02 short stage +7.43 ± 4.84 | 3,924 | +2.97 | A continuation correction to the static evaluation indexed by the piece and target square of the previous moves at ply minus one and ply minus three increases playing strength in negamax. | |
| T-0121 | passed | epoch change, no games | A new training epoch is evaluated using the same predecessor source. | |||
| T-0120 | passed | long stage +97.56 ± 24.50 short stage +97.32 ± 27.75 | 296 | +2.95 | A 256-wide NNUE network trained with wdl 0.5, CReLU, a single bucket, a learning rate decaying smoothly to zero each step, and 4 epochs on an updated dataset of 2.50 billion positions plays stronger than the current baseline engine. | |
| T-0119 | passed | long stage +111.00 ± 31.92 short stage +75.62 ± 24.25 | 288 | +2.95 | The engine is expected to play stronger than main. The candidate combines an adapted network width read from the header (256 neurons, one bucket), an N256 network trained on the production recipe with 395 million training lines over 18 epochs, and targeted speed optimizations using O3 pragmas and target_clones for x86-64-v3 on ELF targets in network evaluation, negamax, quiescence search, and bitboard move generation. | |
| T-0118 | rejected | short stage −0.65 ± 4.96 | 5,866 | −3.00 | A collected principal variation is output via UCI for each completed iteration. | |
| T-0117 | rejected | short stage +1.43 ± 3.17 | 14,616 | −2.98 | The existing engine is ported to Rust to evaluate the choice of programming language while keeping the algorithm identical. | |
| T-0116 | rejected | short stage −3.66 ± 7.57 | 3,416 | −3.01 | When improving, ProbCut verifies a capture at one depth shallower, using depth minus 4 minus 1 instead of depth minus 4. This makes the verification search cheaper in positions that are already improving. | |
| T-0115 | rejected | short stage −0.38 ± 5.40 | 6,414 | −2.97 | A child node reduced by LMR searches one ply shallower if the static evaluation drops by more than 100 cp over the played move. This condition requires that the sum of the position-relative static evaluations of the parent and child exceeds 100, and that the child depth is at least 2. | |
| T-0114 | rejected | short stage −6.00 ± 8.65 | 2,490 | −2.98 | A static evaluation correction history, refreshed per search and keyed by side to move and king, knight, and bishop hash keys, learns from informative node results that are not in check, not mate, have a quiet best move, and provide a bound in the informative direction. It adds up to 256 cp to static eval for all cutoffs in negamax, while the transposition table retains the raw evaluation. | |
| T-0113 | rejected | short stage −0.86 ± 5.75 | 5,666 | −2.98 | Butterfly history and countermove tables persist across searches within a game session and are only cleared on ucinewgame. Butterfly history is halved at the start of each search so that values from earlier positions decay. | |
| T-0112 | rejected | short stage −0.05 ± 5.18 | 7,090 | −2.98 | A quiet pawn push that leaves the pawn as a passed pawn on the 6th or 7th rank from the moving side's perspective is not reduced and is extended by one ply in the child node (depth instead of depth minus 1). This extension is not applied in addition to a check extension. | |
| T-0111 | rejected | short stage −14.93 ± 12.17 | 1,234 | −2.95 | Forward futility pruning of quiet moves uses a one-quarter larger margin when improving, specifically static_eval + (FutilityMargin + FutilityMargin / 4) * depth <= alpha. This causes an improving position to be skipped as hopeless less frequently. | |
| T-0110 | rejected | short stage −3.02 ± 7.22 | 3,676 | −2.95 | A per-search correction history of static eval, keyed by side to move and pawn structure hash, learns from informative node results and adjusts static_eval by up to 256 cp for pruning decisions in negamax, while the transposition table keeps the raw evaluation. This remeasures the idea because the earlier test evaluated an HCE engine, whereas the current engine is pure neural network. | |
| T-0109 | rejected | short stage −1.18 ± 5.34 | 5,004 | −2.96 | Portable build flags -O3 -march=x86-64-v3 -flto were applied to the NNUE component. Utilizing AVX2 is expected to triple the speed of the int16 accumulator loop in the testbed and training epoch, rather than in the engine code itself. | |
| T-0108 | passed | long stage +9.73 ± 7.13 short stage +12.11 ± 8.64 | 2,964 | +2.96 | Internal iterative reduction at depth 4 or higher in null-window searches is applied not only when a transposition table move is missing, but also when the table entry was searched much shallower, specifically if the stored depth plus 4 is less than the current search depth. The expected rationale is that such move ordering provides little benefit over having no table move. | |
| T-0107 | rejected | short stage −14.87 ± 12.80 | 1,356 | −2.96 | The network evaluation is scaled by the material on the board using the formula v = net * (37000 + M) / 37000, where M is 250 times the pawn count plus the piece values of both sides excluding kings. This increases the evaluation by around 28% on a full board, leaves it almost unchanged in pawn endgames, and keeps it clamped against mate as before. | |
| T-0106 | rejected | short stage +0.72 ± 3.88 | 9,670 | −2.98 | Four output heads selected solely by piece count in the classes 0 to 7, 8 to 15, 16 to 23, and at least 24 non-kings on a shared N128 trunk outperform the tied N128 single-head control network at the same time control. |