diff --git a/public/js/exhibit.js b/public/js/exhibit.js
index a91d552..a8c3db5 100644
--- a/public/js/exhibit.js
+++ b/public/js/exhibit.js
@@ -69,17 +69,15 @@ function centredCorners(m, w, h) {
 }
 
 // Build the marker->camera matrix (POS-IT), converting js-aruco axes to three.js.
-// POS-IT frame is right-handed (X right, Y up, Z toward viewer). To bring the
-// full RIGID transform into three.js we must CONJUGATE by F = diag(1,-1,-1):
-//   M_three = F * M_pose * F
-// Premultiplying by F alone flips the translation but leaves the rotation's
-// pitch inverted -> "camera up/down moves the ghost the wrong way". The second
-// (right) multiply by F fixes the rotation handedness while keeping translation
-// correct (F*M*F's translation column is F*t, i.e. flipped exactly once).
+// js-aruco POS-IT gives the marker pose in a camera frame with X right, Y up,
+// Z toward viewer. Three.js's camera looks down -Z with Y up, so we convert by
+// premultiplying F = diag(1,-1,-1) (flip Y and Z). This is the SAME conversion
+// the working ar-head.js uses, and it is correct for pitch AND roll: tilting the
+// phone up moves an anchored ghost down-screen, and rolling the phone rotates the
+// ghost the same way. (An earlier F*M*F "fix" inverted both — don't reintroduce it.)
 const flipYZ = new THREE.Matrix4().makeScale(1, -1, -1);
 
-// Convert one POS-IT (R,t) solution into a three.js marker->camera matrix via
-// the F*M*F conjugation (see note above).
+// Convert one POS-IT (R,t) solution into a three.js marker->camera matrix.
 function poseToMatrix(R, t) {
   const M = new THREE.Matrix4().set(
     R[0][0], R[0][1], R[0][2], t[0],
@@ -87,8 +85,7 @@ function poseToMatrix(R, t) {
     R[2][0], R[2][1], R[2][2], t[2],
     0, 0, 0, 1
   );
-  M.premultiply(flipYZ);  // F * M
-  M.multiply(flipYZ);     // F * M * F
+  M.premultiply(flipYZ);  // F * M  (flip Y & Z: POS-IT frame -> three.js frame)
   return M;
 }