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-0227 – T-0167
| 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-0227 | rejected | long stage −5.18 ± 6.69 short stage +5.50 ± 3.71 | 2,482 | −1.34 | Storing singular search results under a dedicated key salted with position, excluded move, cut node, and tt pv, and immediately returning this score at matching subsequent singular nodes with equal probe depth and bound while preserving regular transposition table evaluation and move ordering, increases engine playing strength. | |
| T-0226 | rejected | short stage −2.76 ± 3.60 | 8,950 | −2.97 | At all null move candidates, the null move is additionally gated by requiring the static evaluation to satisfy minus the raw evaluation with swapped side to move at least beta. The opponent evaluation queries the neural network on identical feature sums with swapped side to move, applies material scaling, inverts the bishop mate term, and clamps like standard evaluation, while leaving null move reduction, verification, and locks unchanged. This increases playing strength. | |
| T-0225 | interrupted | long stage −0.68 ± 1.47 short stage +0.02 ± 1.71 | 36,000 | +1.26 | Exempting captures that lead immediately to insufficient material from capture futility pruning and SEE Static exchange evaluation: an estimate of the material outcome of a capture sequence on one square, without searching. capture pruning near the horizon in main search PV Principal variation: the line of best moves the search currently expects. and non-PV nodes prevents false fail-lows when alpha < 0. This change targets sacrificing captures with at most two pawns, rooks, or queens on the board, leaving quiescence search unchanged and preserving playing strength. | |
| T-0224 | rejected | long stage +0.44 ± 2.30 short stage +2.31 ± 1.85 | 20,376 | −1.34 | In reverse futility pruning, the fixed reduction RfpGegner when gegner_schwaecher holds is replaced by min(RfpGegner, statisch + vor1). When vor1 is invalid or the sum is at most zero, no reduction is applied. If all other search parameters, including RfpFaktor, RfpSteigend, the null-clamp, and null move pruning, remain unchanged, the engine is expected to play stronger. | |
| T-0223 | rejected | long stage +0.56 ± 2.10 short stage +4.66 ± 3.31 | 24,286 | −1.35 | In non-PV nodes not in check and without exclusion under a non-mate window, countermove maintenance is modified: the quiet countermove read before the move loop is cleared from its unmodified slot if it was legally searched to at least positive nominal child depth, returned a non-mate score at or below alpha, and a tactical move subsequently achieves beta. This change is expected to make the engine play stronger. | |
| T-0222 | rejected | short stage −2.98 ± 3.70 | 8,386 | −2.95 | In quiet move ordering, the bonus condition is changed so that in addition to quiet checks, quiet threats also receive the bonus once per move under the same SEE threshold: pawn attacking a piece, minor attacking rook or queen, or rook attacking queen. All other heuristics, pruning, and reductions remain unchanged. This modification is expected to increase engine playing strength. | |
| T-0221 | rejected | short stage −2.31 ± 3.17 | 9,322 | −2.95 | In non-PV nodes under specific conditions regarding depth, halfmove clock, beta bounds, and move types, the null-move gate comparisons and reduction are switched to use an adjusted evaluation upper bound instead of the static evaluation. With verification and probe kept as in the baseline, this change is expected to make the engine play stronger. | |
| T-0220 | rejected | short stage −0.94 ± 2.62 | 16,272 | −2.97 | For non-promotion quiet killer and counter moves that futility pruning would discard but history and SEE allow, the veto is changed across all node types. The move and the legal countermove from history are executed, followed by a grandchild quiescence search with a null window at alpha. Only a score above alpha releases the move for regular search, otherwise it is discarded as before; this change is expected to make the engine play stronger. | |
| T-0219 | rejected | short stage −0.32 ± 2.25 | 22,546 | −2.94 | In futility pruning for quiet moves, the condition f > alpha is tightened to f - V > alpha before make_move, where f is the current evaluation estimate and V = max(0, -SEE(move)). This applies after legality and LMP checks if best > -MateThreshold, not in check, d <= FutilityDepth, abs(alpha) < MateThreshold, and the move is non-checking and non-threatening, while retaining history and SEE pruning. This change is expected to increase playing strength. | |
| T-0218 | passed | epoch change, no games | The evaluation network is updated to a different training epoch from the same predecessor source. | |||
| T-0217 | rejected | short stage −0.32 ± 2.23 | 22,554 | −2.95 | In ProbCut, the priority of a tactical transposition table move over other captures is removed if its entry does not prove a lower bound of at least beta plus ProbCutMargin, sorting it by its standard capture rank instead. If this priority is retained only when the lower-bound bit is set and the score is at least this threshold, playing strength is expected to increase. | |
| T-0216 | rejected | long stage −7.10 ± 7.59 short stage +2.08 ± 1.64 | 1,908 | −1.35 | If the suffix following the root best move is sorted for the next depth (from depth 2) using the current move picker logic—instead of sorting by prior-iteration nodes and a dedicated root history table—while keeping the best move in front and retaining all search, reduction, and time-management rules, the engine is expected to play stronger under clock conditions. | |
| T-0215 | rejected | short stage −1.10 ± 2.74 | 15,198 | −2.96 | At cut nodes without exclusion, the first four eligible SEE-vetoed moves meeting specific conditions—no king move, promotion, or en passant, and where the least valuable retractor leaves a previously covered opponent piece exposed to an ongoing attack—are deferred. If no beta cutoff occurs, these deferred moves are searched normally at the end without prune blocking, re-LMP, or double counting. This change increases the engine's playing strength. | |
| T-0214 | rejected | short stage −7.59 ± 5.31 | 3,984 | −2.95 | If iterative deepening is modified after a suitable pre-search to start directly at the retrieved depth D instead of depth 1 whenever an exact match with 4 <= D < max_depth is found from an end PV of at least two moves matching the root and history, using an aspiration window around the match score, an initial value drop test, and reset stability, the engine plays stronger than main. | |
| T-0213 | rejected | long stage −3.29 ± 4.93 short stage +1.66 ± 1.30 | 3,064 | −1.36 | If the root search is changed so that a late root move whose completed null-window probe at the full new depth of the current attempt scored above alpha and below beta is accepted as the best move of the iteration upon aborting its PV re-search—using this score as a lower bound and a single-move line, without using the aborted value, incomplete child PV, or new transposition table entries—then the engine plays stronger. | |
| T-0212 | rejected | short stage −10.51 ± 6.14 | 3,008 | −2.96 | In non-PV main search nodes that are not in check and have no exclusions, move picking for the remaining quiet moves after all late quiet stages and the main prefix is changed to a round-robin pass over origin squares, prioritizing previously highest-ranked unused squares before starting a new round. This adjustment is expected to increase playing strength compared to the baseline. | |
| T-0211 | rejected | short stage −2.72 ± 3.54 | 8,930 | −2.97 | At the null-move gate, if non-PV, not in check, not an exclusion node, and the gate is otherwise met with static eval < beta <= evaluation from the lower bound (or exact), the static veto is removed when depth - r0 >= 2 with r0 = null_move_reduction(depth, 0), setting r = r0. All other conditions, reduction verification, and search parameters remain identical. This change is expected to make the engine play stronger against main. | |
| T-0210 | rejected | short stage +0.62 ± 1.40 | 57,582 | −2.96 | Disabling correction history updates at the end of the search when the best score equals DrawScore, prior to the final draw clamp, is expected to increase playing strength. Window gating, all other update guards, update formulas, correction keys, returns, transposition table storage, and other correction channels remain unchanged. | |
| T-0209 | rejected | short stage −0.52 ± 2.37 | 20,120 | −2.96 | Removing only the hard history pruning of quiet moves, along with its associated tuning parameters, while retaining futility pruning, LMP, both SEE prunings, LMR, and all other history reads and update rules in the search, will increase the engine's playing strength. | |
| T-0208 | rejected | short stage −0.05 ± 2.03 | 27,502 | −2.96 | Replacing the rank allowance in history learning with the remaining bonus scaled by min(reached[1] / reached[min(searched, 9)], 4) in sixty-fourths, capped at min(history limit, capture limit), makes the engine play stronger. The search increments nine saturating uint64 counters per reached search rank up to 9 only at non-PV nodes without exclusion, defaulting to a factor of 1 on division by zero or otherwise, while penalties, killers, and countermoves remain unchanged. | |
| T-0207 | rejected | short stage −1.94 ± 3.18 | 11,130 | −2.96 | Omitting both preference-learning branches and the final history update (including bonuses, penalties, capture history, killers, and countermoves) at the end of the search when the best score equals the draw score before the final draw clamp, while retaining transposition-table cutoff killers, evaluation, research, and correction learning, returns, and table stores, increases the engine's playing strength. | |
| T-0206 | rejected | short stage −2.54 ± 3.52 | 9,440 | −2.95 | Removing the extra exemption from late move reductions for the second legal move at PV nodes, thereby allowing reductions starting from the second legal move just as at non-PV nodes, increases engine strength. Reduction formulas including LmrPv, depth clamps, researches, and learning rules remain unchanged. | |
| T-0204 | rejected | short stage +0.75 ± 1.24 | 73,526 | −2.97 | Before a transposition table cutoff outside PV and singular search with score at least beta, non-mate score, and a legal hash move, the engine probes the child entry after that move. The cutoff and killer update are skipped if the child has a non-mate entry of depth at least depth minus one, lower bound or exact, whose score from the parent's perspective is below beta. The expected result is increased playing strength under time controls. | |
| T-0203 | rejected | short stage −0.11 ± 2.08 | 26,224 | −2.95 | If a quiet, non-checking move in search that retracts the piece from the destination square of the engine's move two plies earlier back to its origin square is reduced more heavily in late move reductions by a margin of LmrPendel (1/1024 ply), playing strength increases. All other search rules remain unchanged. | |
| T-0202 | passed | long stage +12.61 ± 6.33 short stage +9.23 ± 5.13 | 2,894 | +2.71 | Replacing the main network with a 768-wide net featuring 8 symmetric king buckets, trained for 16 epochs at a learning rate of 0.00025 and WDL 0.5, followed by self-distillation over 4 epochs on 20k search labels, increases engine playing strength under clock conditions. | |
| T-0201 | rejected | short stage −3.56 ± 3.91 | 7,318 | −2.95 | Calling move-preference learning once with a history bonus of one upon transitioning directly into the quiescence search after an uninterrupted, non-mate fail-low increases playing strength. Recursive quiescence search, evaluation learning, main search learning, and all guards for preference learning remain unchanged. | |
| T-0200 | rejected | short stage −11.17 ± 6.32 | 2,862 | −2.96 | Disabling the null move pruning probe when a friendly knight, bishop, rook, or queen is attacked by a less valuable opponent and eval minus beta is less than the piece-specific threshold (knight 220, bishop 230, rook 170, queen 400), while keeping all other null move gating and search heuristics unchanged, improves engine playing strength. | |
| T-0199 | passed | long stage +2.65 ± 2.18 short stage +5.64 ± 3.55 | 17,720 | +2.73 | Combining TT Transposition table: a cache of already searched positions, keyed by a hash of the position. prefetching before search cutoffs, a 512-bit network accumulator for x86-64-v4, cuckoo table lookups in repetition detection restricted to indices where opponent moves cancel out without side keys, and compile-time permutation of FT rows, FT biases, and both L2 view halves with an identical weight file increases engine speed over the baseline while maintaining identical search tree, evaluations, tuning parameters, and search outputs. | |
| T-0198 | passed | long stage +4.22 ± 3.19 short stage +4.60 ± 3.28 | 10,446 | +2.73 | In reverse futility pruning, the fixed reduction when evaluation is improving is replaced by the minimum of that reduction and the actual score increase. This increase is measured against two plies earlier, or four plies earlier if that is unavailable, falling back to the fixed reduction if neither exists. With all other pruning and search parameters kept unchanged, the engine is expected to play stronger. | |
| T-0197 | passed | epoch change, no games | The evaluation network is updated by advancing to another training epoch derived from the same preceding training data source. | |||
| T-0196 | rejected | short stage +0.54 ± 1.51 | 50,648 | −2.96 | Updating the default search parameters to the values obtained from an SPSA run, with 82 parameters modified relative to the baseline while keeping parameter bounds, time-management parameters, and other settings unchanged, is expected to increase playing strength. | |
| T-0195 | passed | long stage +7.81 ± 4.58 short stage +2.84 ± 2.27 | 5,294 | +2.94 | Replacing the check extension SEE threshold with a SEE threshold of 0 for quiet checking moves when extension is 0, while leaving all other search components unchanged, is expected to increase engine playing strength. | |
| T-0194 | rejected | short stage +0.71 ± 1.09 | 78,334 | −2.96 | In the razoring gate of the search, the evaluation lowered below the corrected static evaluation by an upper transposition table bound is replaced by the maximum of the static and the score evaluation. Razoring now triggers only when both the static evaluation and the score plus the razor margin fall below alpha, while all other transposition table refinements, margins, and pruning rules remain unchanged. This change is expected to make the engine stronger at fixed nodes. | |
| T-0193 | rejected | short stage −1.99 ± 2.87 | 9,764 | −2.95 | Disabling the null-move pruning solely when the side to move has only knights besides king and pawns, with at least one knight and every knight being absolutely pinned to its king, increases playing strength at fixed node limits compared to the baseline. All other null-move conditions, reductions, mate pruning, verification, side-to-move restrictions, and the remaining search stay unchanged. | |
| T-0192 | rejected | short stage −1.52 ± 2.99 | 12,790 | −2.95 | Updating 89 search parameters from an SPSA self-play tuning run—of which 81 differ from main—and setting the history boundaries to a minimum history limit of 4096, a maximum history bonus of 2048, and a maximum history malus of 2048 increases the engine's playing strength compared to main. | |
| T-0191 | rejected | short stage +0.33 ± 1.72 | 38,950 | −2.96 | Out of 89 SPSA-tuned search parameters, 34 with a final filter statistic magnitude of at least 1 are set to their rounded values, while the remaining 55 keep their baseline values. With all other limits and parameters unchanged, the engine plays stronger than the baseline. | |
| T-0190 | passed | long stage +2.76 ± 2.21 short stage +4.69 ± 3.33 | 22,430 | +2.95 | If killer 1, killer 2, and countermoves receive no fixed rank bonus in quiet move ordering when their history score is below -3000, and if the LMR reduction discount for killers and countermoves is removed when their history score is below -3000 (including tactical killers and countermoves), the engine will play stronger on the clock. | |
| T-0189 | passed | long stage +24.37 ± 8.50 short stage +29.06 ± 9.46 | 1,528 | +2.97 | Replacing the default evaluation network with a model trained on filtered self-play data and subsequently fine-tuned on the network's own search labels will increase the engine's playing strength under time controls. | |
| T-0188 | interrupted | long stage +0.68 ± 1.52 short stage +3.17 ± 2.50 | 48,000 | −2.21 | Ordering the suffix moves following the best move after each completed root iteration for the next depth (from depth 2 onward) using the updated move picker logic, rather than sorting by prior-iteration node counts and a dedicated root history table, increases the engine's playing strength under time controls. The best move itself at the front and all search, reduction, and time management rules remain unchanged. | |
| T-0187 | passed | long stage +2.02 ± 2.88 short stage +0.08 ± 1.76 | 9,818 | +2.98 | Adding a post-material-scaling bonus solely in bishop-and-knight against bare king (exactly four pieces) of 40 cp per square by which the distance of the weaker king to the closer corner of the bishop's colour (files plus ranks) is below 7, capped at 280 cp, will help the engine checkmate more often in this endgame without playing weaker overall. The evaluation of all other positions, including two bishops against a bare king, remains unchanged. | |
| T-0186 | passed | long stage +2.77 ± 2.15 short stage +3.23 ± 2.44 | 17,460 | +2.98 | Applying the tempo switch consistently across the main search, quiescence search, iterative deepening, ply limit, and draw detection, while also passing the child's check state directly from the parent node rather than recomputing it when already known, measurably increases engine speed against main without altering the search tree. | |
| T-0185 | passed | long stage +5.79 ± 3.82 short stage +3.37 ± 2.62 | 7,742 | +2.96 | Removing the check exception from both the delta pruning and SEE delta pruning conditions in quiescence search, so that checking captures are also pruned by them, is expected to increase playing strength in timed games against the baseline. Mate bounds, recapture exceptions, and protections when in check or on promotions are retained, and checking moves remain exempt from the move limit and exchange threshold. | |
| T-0184 | rejected | short stage +0.24 ± 1.79 | 35,688 | −2.97 | If iterative search ends an iteration with the current score as a lowerbound when a root aspiration search fails high, a prior iteration is completed, the attempt has a variation, and the best move of the iteration matches the previous one—instead of widening the window and researching—the engine plays stronger at fixed 100,000 nodes. | |
| T-0183 | passed | long stage +2.18 ± 1.74 short stage +2.56 ± 2.07 | 37,052 | +2.96 | Disabling history bonuses (main, near-root, pawn, and continuation history) on transposition table beta-cutoffs with a quiet hash move, while still updating the killer move and leaving node-end learning and the rest of the search unchanged, is expected to increase engine playing strength in timed games. | |
| T-0182 | passed | long stage +13.18 ± 6.09 short stage +18.70 ± 7.53 | 2,874 | +2.96 | Updating 84 search parameters to values tuned via SPSA self-play and widening the bounds for RfpFaktor (minimum 20), VerlaufGrenze (minimum 6144), and MaterialSkala (maximum 160) will improve the engine's playing strength under standard time controls. | |
| T-0181 | interrupted | long stage +1.27 ± 1.52 short stage +4.10 ± 3.03 | 48,000 | +0.18 | If killer and countermove bonuses in quiet move ordering are skipped when the ordering history sum (history plus continuations, near-root and pawn history) falls below -3000, and the late move reduction deduction for killers is omitted when the move's history intersection falls below -3000, while leaving killer and countermove table updates and the rest of the search unchanged, the engine will play stronger on the clock. | |
| T-0180 | passed | long stage +31.43 ± 9.51 short stage +24.60 ± 8.58 | 1,186 | +2.97 | Fine-tuning the main evaluation network on self-play positions labelled using the engine's own deeper search increases playing strength under clock conditions. In the training run, positions were searched at 20,000 nodes each and trained over 150 million positions for 4 epochs with WDL 0.5. | |
| T-0179 | passed | long stage +13.02 ± 6.10 short stage +9.43 ± 5.15 | 2,990 | +2.97 | Increasing the material scaling of the network evaluation by raising MaterialSkala from 40 to 80, while keeping MaterialMitte unchanged at 3370, is expected to improve playing strength under clock conditions compared to the baseline. | |
| T-0178 | rejected | short stage −0.17 ± 1.86 | 24,792 | −2.97 | Instead of reclassifying move cases, child structure keys after each real make_move across the main loop, ProbCut, quiescence search, and root are derived via a helper from the parent structure and the piece changes tracked in undo before network staging. The search tree, structure keys, and correction history addresses remain identical across all generated search moves, yielding a higher node rate and greater playing strength on the clock against the baseline. | |
| T-0177 | passed | epoch change, no games | Testing a different training epoch from the same network checkpoint source. | |||
| T-0176 | passed | long stage +3.69 ± 2.80 short stage +4.92 ± 3.48 | 14,122 | +2.95 | In the search, adjusting the quiet futility pruning margin based on the static evaluation increases playing strength. Specifically, the margin is increased by 45 when the best score is less than the static evaluation minus 50, and decreased by 45 otherwise, while capture futility, history pruning, SEE pruning, and LMP remain unchanged. | |
| T-0175 | passed | epoch change, no games | The training epoch was changed using the same predecessor source. | |||
| T-0174 | passed | long stage +54.65 ± 11.65 short stage +51.19 ± 12.05 | 750 | +2.95 | Integrating the engine development line, including its custom search, move selection, correction, board, tempo modules, and parameters tuned through eight SPSA self-play rounds, into main alongside missing main features such as the SKML768V8J32F net, stack-end mate handling, UCI Universal Chess Interface, the text protocol between a chess engine and the program that runs it. robustness, and full hash budget increases playing strength against main in timed games. | |
| T-0173 | rejected | short stage −1.56 ± 3.08 | 12,892 | −2.96 | If move ordering after the LMP cutoff continues to generate threat moves from the remaining quiet moves in generation order without further sorting, playing strength on the clock increases. Threat moves are quiet moves that attack a higher-value opponent piece from their target square, while futility, history pruning, SEE, and LMR remain unchanged. | |
| T-0172 | passed | epoch change, no games | Evaluation of a different training epoch from the same base source. | |||
| T-0171 | passed | long stage +12.37 ± 6.05 short stage +3.81 ± 2.90 | 3,260 | +2.95 | Clamping correction-history updates for informative fail-soft results to the search window—beta for lower bounds and initial alpha for upper bounds, only when still informative over static eval—while keeping exact updates, TT bounds, and search returns unchanged, increases playing strength at fixed nodes against the baseline. | |
| T-0170 | passed | long stage +5.68 ± 3.79 short stage +19.47 ± 7.76 | 7,950 | +2.95 | Refining history updates as a bundle increases engine playing strength under clock conditions. This involves a larger bonus for a best move that was a table move and scaled by move count at non-PV nodes, a malus decaying in search order, a penalty for an early refuted quiet previous move, a fail-low bonus for the quiet previous move, damping main history at each search start, and slightly negative table initialization. | |
| T-0169 | rejected | short stage −3.92 ± 4.15 | 7,008 | −2.95 | Exempting quiet checks from LMP pruning and passing remaining quiet moves unordered after the LMP threshold, pruning non-checking quiet moves before incrementing the move count, increases engine playing strength under clock conditions. | |
| T-0168 | passed | long stage +12.41 ± 6.07 short stage +16.21 ± 7.08 | 3,278 | +2.95 | Replacing the default evaluation network with a model trained on 5.58 billion positions of self-play over 32 epochs increases the playing strength of the engine under time control. | |
| T-0167 | rejected | short stage +0.34 ± 1.45 | 41,692 | −2.97 | Updating int16 accumulators via a runtime-selected routine using 512-bit words (32 instead of 16 sums per step) on processors with AVX-512F and AVX-512BW yields bit-identical accumulators and an unchanged search tree. This increases the node rate on these processors and improves playing strength under time controls. |