diff --git a/docs/src/gcode/g-code.adoc b/docs/src/gcode/g-code.adoc
index 6253e57afb2..19e519dc480 100644
--- a/docs/src/gcode/g-code.adoc
+++ b/docs/src/gcode/g-code.adoc
@@ -1073,14 +1073,19 @@ threads.
 
 [[gcode:g33-tech-info]]
 .Technical Info
-At the beginning of each G33 pass, LinuxCNC uses the spindle speed and
-the machine acceleration limits to calculate how long it will take Z to
-accelerate after the index pulse, and determines how many degrees the
-spindle will rotate during that time. It then adds that angle to the
-index position and computes the Z position using the corrected spindle
-angle. That means that Z will reach the correct position just as it
-finishes accelerating to the proper speed, and can immediately begin
-cutting a good thread.
+A G33 move waits for the spindle index, then accelerates from rest to the
+synchronized feed. The commanded position follows the spindle angle measured
+from that index, so a given thread is cut in the same place whatever the
+spindle speed.
+
+While the axis is still accelerating it falls behind that position, by roughly
+the pitch squared times the spindle speed squared over twice the axis
+acceleration limit, and it makes the distance up over the rest of the move.
+Allow enough lead-in for it to catch up before the cut begins, or the start of
+the thread is cut with the wrong lead. LinuxCNC reports 'lead-in too short to
+reach sync' when the move ends before the axis has caught up. A small lag
+proportional to speed remains for the rest of the move, which is why the
+spindle speed must not change while a thread is being cut.
 
 .HAL Connections
 The pin 'spindle.N.at-speed' must be set or driven true for the motion to
diff --git a/src/emc/tp/tp.c b/src/emc/tp/tp.c
index b193b76fb83..f7d58f50e47 100644
--- a/src/emc/tp/tp.c
+++ b/src/emc/tp/tp.c
@@ -3567,6 +3567,29 @@ STATIC void tpSyncVelocityMode(TP_STRUCT * const tp, TC_STRUCT * const tc, TC_ST
 }
 
 
+/**
+ * Warn when the segment is too short to absorb the lead-in lag.
+ * The tracking loop below closes an error e with v = v_spindle + sqrt(e * a),
+ * which takes 2 * sqrt(e / a) seconds.
+ */
+STATIC void tpCheckSyncLeadIn(TC_STRUCT const * const tc, double pos_error)
+{
+    double accel = tcGetTangentialMaxAccel(tc);
+    if (accel <= 0.0) {
+        return;
+    }
+    double err = fabs(pos_error);
+    double catchup = tc->currentvel * 2.0 * pmSqrt(err / accel) + err;
+    double remaining = tc->target - tc->progress;
+    if (catchup > remaining) {
+        rtapi_print_msg(RTAPI_MSG_ERR,
+                "spindle-synchronized move %d: lead-in too short to reach sync, "
+                "need %f more travel\n",
+                tc->id, catchup - remaining);
+    }
+}
+
+
 /**
  * Run position mode synchronization.
  * Updates requested velocity for a trajectory segment to track the spindle's position.
@@ -3603,10 +3626,13 @@ STATIC void tpSyncPositionMode(TP_STRUCT * const tp, TC_STRUCT * const tc,
         target_vel = spindle_vel * tc->uu_per_rev;
         if(tc->currentvel >= target_vel) {
             tc_debug_print("Hit accel target in pos sync\n");
-            // move target so as to drive pos_error to 0 next cycle
-            tp->spindle.offset = tp->spindle.revs - tc->progress / tc->uu_per_rev;
+            // Leave the sync origin on the spindle index. The lag built up
+            // while ramping to sync speed is a real position error, so hand it
+            // to the tracking loop below; folding it into the origin instead
+            // shifts the thread by an amount that grows with spindle speed.
             tc->sync_accel = 0;
             tc->target_vel = target_vel;
+            tpCheckSyncLeadIn(tc, pos_error);
         } else {
             tc_debug_print("accelerating in pos_sync\n");
             // beginning of move and we are behind: accel as fast as we can