Gauntlet test report — anomaly review, v18 dev (2026-08-02)

A ramp, not a cliff

Same 9+0.1 control, same rig, a dominant top-30 field — and a loss that shows the clock running down move by move, right into the one moment it matters.

70.8%
score / 1260 games
2900.4
live elo (+17.4)
+38.8
elo margin over 2nd
3rd
version showing this signature

vs Halogen 16.0 — depth ramps down, right on schedule

Unlike the earlier reports' sudden cliffs, this one shows a genuine decline across four consecutive moves — the clock visibly running out, bottoming exactly at the fatal move.

+3.13
45...Rd8
depth 7
+3.13
46...Qh3+
depth 4
+4.00
47...Rf6
depth 6
+6.41
48...a5
depth 1
-0.73
49...Qe6
depth 3
Search depth declines 7 → 4 → 6 → 1 across four straight moves — a genuine ramp-down, not an isolated glitch. Black's own eval keeps climbing right through it (+3.13, +3.13, +4.00, then a spike to +6.41 on a single ply of search), completely disconnected from what's actually about to happen on the board.
+6.41
move 48, depth 1
-0.73
move 49, one ply later
Mate
forced, ~15 moves later

Same signature, three versions in a row

The 17.0 and 17.1 reports both showed sudden depth cliffs — healthy search one move, depth 1 the next. This game shows the ramp itself happening in real time, which strengthens the read: this looks like a genuine clock exhaustion, not a per-version search quirk.

Reproducing across 17.0 → 17.1 → 18-dev on the same hardware and time control points toward something structural in how time gets allocated at 9+0.1 on this rig, rather than a regression tied to any single build.

A dominant result, one flagged edge case

FIELD RESULT (top 30 opponents, 42 games each)
Points891.5 / 1260
Score70.8%
Live Elo2900.4
Start Elo2883.0
Δ+17.4
Margin over 2nd (Reckless)+38.8
RIG
Hardwarei5-4570, 8GB RAM
Time control9s+0.1s (5Pct of 3+2)
BuildStockfish 18 dev, 2026-08-02
One loss out of 1,260 games, against a field SF dominates by a wide margin. Flagged as a reproducible, version-independent edge case — not a strength concern.