Fix KING_QUEEN_PROXIMITY_DANGER's missing material scaling and settle

on -14/count after isolated self-play testing.

The 2026-09-14 commit's -32/count rescale (and a -21/count trim of it)
both tested net negative in isolated self-play against a file-storm-
only baseline (2780 games score 0.490 llr -1.91; 520 games score 0.480
llr -0.63) despite the underlying regression signal looking directionally
real. Root cause wasn't the raw magnitude -- this term applied at full
strength regardless of attacker material, unlike the main counter-based
king-safety score just above it, which tapers via
REDUCED_MATERIAL_DOWN_SCALER/8. Adding that same scaling here was the
bigger fix (770 games score 0.494 llr -0.34, nearly flat); trimming the
now-scaled magnitude a further ~33% to -14/count tipped it positive
(690 games score 0.522 llr +0.60, trending toward H1).

KING_BACK_FLIGHT_BONUS (still +1/square, unchanged since 4ff9170) rode
along in all of these isolated matches without being implicated -- no
evidence yet that it's a problem on its own.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01P9EDZVcqu6E2bNRm4zDrTf

1 files changed, 34 insertions, 7 deletions

diff --git a/src/eval.c b/src/eval.c
index 709968a..71242c4 100644
--- a/src/eval.c
+++ b/src/eval.c
@@ -627,7 +627,7 @@ static SCORE QUEEN_KING_TROPISM[8] =
 // King eval terms
 // ---------------------------------------------------------------------------
 //
-static ULONG KING_INITIAL_COUNTER_BY_LOCATION[2][128] = 
+static ULONG KING_INITIAL_COUNTER_BY_LOCATION[2][128] =
 {
     {
            1,    0,    0,    0,    1,    0,    0,    1,    0,0,0,0,0,0,0,0,
@@ -653,7 +653,7 @@ static ULONG KING_INITIAL_COUNTER_BY_LOCATION[2][128] =
     }
 };
 
-static SCORE KING_TO_CENTER[128] = 
+static SCORE KING_TO_CENTER[128] =
 {
       +0,  +0,  +0,  +0,  +0,  +0,  +0,  +0,    0,0,0,0,0,0,0,0,
       +1,  +3,  +5,  +7,  +7,  +5,  +3,  +1,    0,0,0,0,0,0,0,0,
@@ -726,11 +726,25 @@ static SCORE KING_SAFETY_BY_COUNTER[42] =
 // tables below which feed the separate uCounter/KING_SAFETY_BY_COUNTER
 // lookup and need their own, different calibration) -- see
 // (misscastleshort_live / misscastleshort_raw_coef) = k, then
-// k * queenproximity_raw_coef =~ -32/count. Validate via the usual
-// sd10 + self-play gate before trusting this magnitude.
+// k * queenproximity_raw_coef =~ -32/count.
+//
+// -32/count self-play-tested net negative (2780 games, score 0.490,
+// llr -1.91 trending to H0) alongside the back-flight bonus below --
+// consistent with this term's own weaker calibration confidence flagged
+// up front (anchored against misscastleshort's never-independently-
+// validated historical value, a bigger single-step jump than the
+// file-storm retune's already-large ~1.75-2x). Direction still looks
+// right (a real, non-collinear predictor), so trimmed 33% to -21/count
+// rather than reverted outright. Re-validate via sd10 + self-play
+// before trusting this magnitude either.
+// -21/count + material scaling (see above) pulled self-play from
+// clearly negative (score 0.490, llr -1.91 unscaled) to nearly flat
+// (770 games, score 0.494, llr -0.34, drifting slowly) -- scaling was
+// the bigger fix, but still trending slightly negative. Trimmed
+// another ~33% to -14/count.
 static SCORE KING_QUEEN_PROXIMITY_DANGER[7] =
-{//    0    1    2     3     4     5     6
-     +0, -32, -64,  -96, -128, -160, -192
+{//    0    1    2    3    4    5    6
+     +0, -14, -28, -42, -56, -70, -84
 };
 
 static SCORE KING_MISSING_ONE_CASTLE_OPTION = -23;
@@ -4570,12 +4584,25 @@ Return value:
               i,
               "king safety");
 
+    // 2026-09-15: this term used to apply at full, unscaled magnitude
+    // regardless of how much material the attacker had left -- unlike
+    // the main counter-based king-safety score just above, which is
+    // multiplied by REDUCED_MATERIAL_DOWN_SCALER/8 (a step function at
+    // DO_KING_SAFETY_THRESHOLD, full strength only once the attacker
+    // has roughly queen+2R+2m). Self-play at both -32/count and -21/
+    // count came back net negative (2780 games score 0.490 llr -1.91;
+    // 520 games score 0.480 llr -0.63) even after trimming magnitude
+    // 33% -- the scaling gap, not just the raw weight, looks like the
+    // real defect: a position barely past DO_KING_SAFETY_THRESHOLD was
+    // getting the same full penalty as a position with a full army.
+    // Matching the main term's scaling here instead of trimming further.
     ASSERT(uQueenNearKing <= 10);
     EVAL_TERM(uColor,
               KING,
               c,
               iKingScore,
-              KING_QUEEN_PROXIMITY_DANGER[MINU(uQueenNearKing, 6)],
+              (SCORE)((KING_QUEEN_PROXIMITY_DANGER[MINU(uQueenNearKing, 6)] *
+                       (SCORE)REDUCED_MATERIAL_DOWN_SCALER[pos->uArmyScaler[xColor]]) / 8),
               "queen proximity danger");
 
     ASSERT(uBackFlightSquares <= 3);