diff --git a/src/search.cpp b/src/search.cpp
index b9766ff..018e68f 100644
--- a/src/search.cpp
+++ b/src/search.cpp
@@ -24,8 +24,14 @@ namespace ironphoenix {
         constexpr int HISTORY_MOVE_PRUNE_MAX_DEPTH = 4;
         constexpr int HISTORY_MOVE_PRUNE_MARGIN_PER_DEPTH = 1'500;
 
+        // Post-LMR continuation history: reward a reduced quiet move that proves
+        // strong enough to beat alpha and require a full-depth re-search.
+        constexpr int POST_LMR_CONTINUATION_BONUS = 1'334;
+
         constexpr int NULL_MOVE_MIN_DEPTH = 2;
         constexpr int NULL_MOVE_DIVISOR = 190;
+        constexpr int REVERSE_FUTILITY_MAX_DEPTH = 4;
+        constexpr int REVERSE_FUTILITY_MARGIN_PER_DEPTH = 200;
         constexpr std::size_t CHECK_BIT_WORDS = (MAX_MOVES + 63u) / 64u;
 
         // Non-PV nodes normally cut off quickly. Selecting only the strongest
@@ -516,10 +522,43 @@ namespace ironphoenix {
         const bool inCheckNode = pos.inCheck();
         const auto& tuning = Tuning::search();
         const Color us = pos.sideToMove();
+        const Color partner = static_cast<Color>(static_cast<unsigned>(us) ^ 2u);
         const bool mateWindow = !(alpha > -MATE_THRESHOLD && beta < MATE_THRESHOLD);
 
-        // Null Move Pruning (NMP): keep expensive partner-check detection and
-        // static evaluation completely off nodes where NMP cannot run anyway.
+        // Static pruning shares one evaluator result between reverse futility
+        // pruning and null-move pruning. Keep partner-check detection lazy so PV,
+        // check and mate-window nodes do not pay for it.
+        int staticEval = 0;
+        bool staticEvalValid = false;
+        bool partnerInCheck = false;
+        bool partnerCheckKnown = false;
+
+        // Reverse Futility Pruning (RFP): at shallow non-PV nodes, if the static
+        // evaluation already exceeds beta by a depth-scaled safety margin, the
+        // position is unlikely to fall below beta after a normal move. Avoid RFP
+        // while either teammate is in check and never return a synthetic mate.
+        if (!pvNode
+            && ply > 0
+            && depth <= REVERSE_FUTILITY_MAX_DEPTH
+            && !inCheckNode
+            && !mateWindow
+            && hasNonPawnMaterial(pos, us)) {
+
+            partnerInCheck = pos.inCheck(partner);
+            partnerCheckKnown = true;
+
+            if (!partnerInCheck) {
+                staticEval = evaluate(pos);
+                staticEvalValid = true;
+
+                const int margin = REVERSE_FUTILITY_MARGIN_PER_DEPTH * depth;
+                if (staticEval < MATE_THRESHOLD && staticEval - margin >= beta)
+                    return staticEval;
+            }
+        }
+
+        // Null Move Pruning (NMP): reuse the static evaluation and partner-check
+        // result when RFP already computed them at shallow nodes.
         if (allowNull
             && !pvNode
             && ply > 0
@@ -528,9 +567,16 @@ namespace ironphoenix {
             && !mateWindow
             && hasNonPawnMaterial(pos, us)) {
 
-            const Color partner = static_cast<Color>(static_cast<unsigned>(us) ^ 2u);
-            if (!pos.inCheck(partner)) {
-                const int staticEval = evaluate(pos);
+            if (!partnerCheckKnown) {
+                partnerInCheck = pos.inCheck(partner);
+                partnerCheckKnown = true;
+            }
+
+            if (!partnerInCheck) {
+                if (!staticEvalValid) {
+                    staticEval = evaluate(pos);
+                    staticEvalValid = true;
+                }
 
                 if (staticEval >= beta) {
                     StateInfo nullState;
@@ -716,8 +762,32 @@ namespace ironphoenix {
                             const int reducedDepth = depth - 1 - reduction;
                             score = -negamax(pos, reducedDepth, -alpha - 1, -alpha, ply + 1, false);
 
-                            if (!shouldStop() && score > alpha)
+                            if (!shouldStop() && score > alpha) {
+                                // Post-LMR continuation-history update: the move was
+                                // reduced because it looked unpromising, but the
+                                // reduced search still beat alpha. Reward the
+                                // continuation contexts before the full-depth retry.
+                                if (ply > 0 && searchStack_[ply].valid()) {
+                                    history_.updateContinuation(
+                                        ContinuationKind::PreviousPly,
+                                        searchStack_[ply],
+                                        movingPiece,
+                                        move.to(),
+                                        POST_LMR_CONTINUATION_BONUS);
+                                }
+
+                                if (ply >= 4 && searchStack_[ply - 3].valid()
+                                    && searchStack_[ply - 3].mover == us) {
+                                    history_.updateContinuation(
+                                        ContinuationKind::SamePlayer,
+                                        searchStack_[ply - 3],
+                                        movingPiece,
+                                        move.to(),
+                                        (POST_LMR_CONTINUATION_BONUS * 3) / 4);
+                                }
+
                                 score = -negamax(pos, depth - 1, -alpha - 1, -alpha, ply + 1, false);
+                            }
                         }
                         else {
                             score = -negamax(pos, depth - 1, -alpha - 1, -alpha, ply + 1, false);