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.

How each run ended

One mark per run in register order, final log-likelihood ratio of the quoted stage. Every SPRT run starts at 0; it passes at the upper bound and fails at the lower one. Hollow marks passed without games (identity, correction, epoch change, adoption) and sit outside the bounds, not as an SPRT exit.

Register T-0012 – T-0001

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 (about +2.94) and fails at the lower one (about −2.94).	Hypothesis
T-0012	passed	long stage +610.64 ± 66.39 short stage +476.43 ± 57.99	2026-08-29	502	+2.96	Ranking captures that a static exchange evaluation shows to lose material behind every quiet move, instead of ahead of them as MVV-LVA alone does, should raise the alpha-beta cut-off rate and make each search depth cheaper.
T-0011	passed	long stage +616.20 ± 70.58 short stage +476.43 ± 63.14	2026-08-29	500	+2.95	Searching every sibling move with a null window and re-searching only when it raises alpha should shrink the search tree at equal depth, because the existing move ordering makes the first move best often enough that the exact value of the others is never needed.
T-0010	passed	long stage +604.58 ± 65.00 short stage +476.43 ± 61.46	2026-08-29	502	+2.95	From depth 4 on, each iterative-deepening iteration searches inside a 50-centipawn window around the previous iteration's score and re-searches only the side that fails, so tighter alpha-beta bounds should prune subtrees that the full window searches.
T-0009	passed	long stage +593.01 ± 62.45 short stage +471.62 ± 65.07	2026-08-28	502	+2.95	Evaluation gains handwritten pawn-structure terms (doubled, isolated and passed pawns) and a king-safety term (pawn shield plus open files beside the king, scaled down with the opponent's remaining piece material), so that quiet positions are judged by more than material and piece-square tables, at a measured cost of 3.5 percent in nodes per second.
T-0008	passed	long stage +554.24 ± 57.63 short stage +394.50 ± 47.85	2026-08-28	506	+2.95	Reducing quiet moves ordered fourth or later by one ply, with a full-depth re-search whenever the reduced null-window probe raises alpha, should trade a small risk of overlooking a good quiet move for a roughly 2.5-fold smaller tree at equal depth, and so reach greater depth in the same thinking time.
T-0007	passed	long stage +490.05 ± 47.06 short stage +368.72 ± 44.52	2026-08-28	516	+2.95	Null-move forward pruning with a zugzwang guard — the null move is only tried when the side to move still has a piece besides king and pawns — should cut roughly half the search tree at equal depth and so reach deeper in the same time, without the endgame blindness an unguarded null move causes.
T-0006	passed	long stage +450.27 ± 41.56 short stage +357.77 ± 47.56	2026-08-28	524	+2.95	Coherent should play stronger because killer moves and a history heuristic order the quiet moves that MVV-LVA leaves untouched, letting beta cutoffs happen earlier so the same depth costs fewer nodes.
T-0005	passed	long stage +386.01 ± 37.16 short stage +316.26 ± 41.89	2026-08-28	542	+2.95	Searching the previous iteration's best root move first in iterative deepening should shrink the search tree, because the root never probes the transposition table and otherwise re-searches every depth in the same static MVV-LVA order.
T-0004	passed	long stage +368.09 ± 34.43 short stage +281.93 ± 37.25	2026-08-28	550	+2.96	A Zobrist hashing A way of computing a position hash incrementally: each piece on each square has a random number, and the hash is their combination.-keyed transposition table (always replace, cutoffs only when the stored entry is deep enough and its bound settles the question, stored move searched first) cuts the fixed-depth node count to 72 % at depth 4 and 53 % at depth 5 on the project's 500 random test positions without changing a single score, so the same time budget should reach deeper.
T-0003	passed	long stage +299.70 ± 28.40 short stage +241.11 ± 33.11	2026-08-28	582	+2.95	Sorting captures by MVV-LVA ahead of quiet moves at every node makes alpha-beta cut off far earlier - the same fixed-depth scores are reached with 45x to 381x fewer nodes across twelve positions at depths 1 to 3 - so within a fixed time budget the engine should search materially deeper.
T-0002	passed	long stage +125.62 ± 18.67 short stage +103.02 ± 22.80	2026-08-28	906	+2.95	Resolving captures at the leaves with a quiescence search capped at six plies should stop the engine from misjudging positions where the recapture lies just beyond the fixed depth, at the price of a larger tree; the bench fingerprint depth was lowered from 4 to 3 to keep a bench run under two seconds.
T-0001	failed	short stage −0.50 ± 1.71	2026-08-28	17,984	−2.97	Negative control: unchanged source code of reference v0. No difference is expected, therefore a fail at the lower bound. Tests the measurement chain, not the engine.