Gauntlet test report — detailed anomaly review, v17.1

Ahead, then mated

Same 9+0.1 control, same aging rig, a 25-opponent field of 2,500 games. Five losses share a search-depth collapse signature — this is the full breakdown, game by game.

65.4%
combined score / 2500 games
891–119
W–L, 1490 draws
9+0.1
time control (5Pct)
5 / 5
reviewed losses fit the pattern

vs Motor 0.7.0 — mated while still counting itself ahead

Stockfish holds a real advantage most of the game. In the final three moves its search collapses to a single ply — and it never notices the mate arriving.

47. Kg3
+1.99
depth 1
48. Kh2
+1.56
depth 1
49. Qxf4
+1.64
depth 1
...Qg1#
Mate
delivered on the board
Not an adjudicated collapse — an actual checkmate, with the shortest window of any case reviewed (3 plies), and the eval never dips below "comfortably ahead" until the mating move lands.

Search depth across all five losses

Sage bars are healthy search (≥10 ply); brick-red bars are the collapse zone (<10 ply). All five losses reviewed show at least one such stretch — including one I initially misread.

vs Motor 0.7.0 · White (SF), moves 9–49fits pattern
move 9move 49 (mate)
3-move crash straight into mate — the sharpest of the five.
vs RubiChess 20230918 · White (SF), moves 9–67fits pattern
move 9move 67 (adjudicated)
Two separate crash windows: a brief one near move 47–50, a sustained one from move 56 onward.
vs RubiChess 20240817 · Black (SF), moves 7–87fits pattern (corrected)
move 7move 87 (mate)
Correction from the earlier draft: I first checked this game's White (RubiChess) depths and called it a "steady grind" with no clear cliff. Stockfish is actually Black in this game — checking Black's own depths shows the same late collapse as the others, concentrated around moves 77–87, right before the final mating sequence.
vs Berserk 13.0 · White (SF), moves 8–53fits pattern — most extended
move 8move 53 (adjudicated)
Moves 36–39 report the identical eval (+4.75) four times running, all at depth 1, before a hard reversal — the most unusual single detail of the five games.
vs PlentyChess 4.0.1 · White (SF), moves 8–51two separate issues
move 8move 51 (mate)
Two distinct failures in one game: a depth-24 miscalculation at move 26 (genuine deep search, not a clock issue), followed independently by a standard depth-1 crash from move 45 onward.

All 25 opponents, 100 games each

#OpponentWLDScore

Still comfortably dominant overall

FIELD RESULT
Combined record891W – 119L – 1490D
Combined score65.4%
Games2,500
RIG
Hardwarei5-4570, 8GB RAM
Time control9s+0.1s (5Pct of 3+2)
Hash / TB16MB / Syzygy
Five losses out of 2,500 games is a small fraction — Stockfish 17.1 is still the clear engine to beat in this field, and this matches the 17.0 gauntlet result closely. These are flagged as a reproducible edge case worth understanding, not evidence of a broad weakness.