diff --git a/public/js/exhibit.js b/public/js/exhibit.js
index 7169a8b..b8ab5a9 100644
--- a/public/js/exhibit.js
+++ b/public/js/exhibit.js
@@ -91,7 +91,7 @@ function dbgStatus() {
   const ws = netRef && netRef.ws ? ['conn','open','closing','closed'][netRef.ws.readyState] : 'none';
   return `ws ${ws}  rx ${rxLog.join(',') || '-'}\n` +
          `scene ${sceneLoaded ? anchorCount + ' anchors' : 'NOT LOADED'}  ghost recs ${activeGhosts.size}\n` +
-         `${fps} fps  detect ${DETECT_HZ}Hz\n`;
+         `${fps} fps  detect ${DETECT_HZ}Hz  fov ${camera3 ? camera3.fov.toFixed(0) : '?'}\n`;
 }
 function paintDbg() { if (dbgEl) dbgEl.textContent = dbgStatus() + dbgPose + dbgAim; }
 
@@ -99,9 +99,11 @@ function paintDbg() { if (dbgEl) dbgEl.textContent = dbgStatus() + dbgPose + dbg
  * camera frame sits where the solved pose projects it. Compares the marker's
  * measured pixel centre against its projected position, and reports the ghost's
  * on-screen NDC. |ndc| < 1 means the ghost is inside the frustum.
- *   px  = where the crest actually is in frame   (-1..1, 0 = centre)
+ *   px  = where the crest actually is on screen   (-1..1, 0 = centre)
  *   proj= where the pose says it should be
- * Those two disagreeing is a wrong frame convention, not tracking noise. */
+ * Those two disagreeing is a camera-model or frame-convention error, not
+ * tracking noise: a sign flip means a mirrored axis, a constant ratio means the
+ * projection does not match the lens. */
 let dbgAim = '';
 const _aimV = new THREE.Vector3();
 function updateAim(markers, W, H) {
@@ -112,8 +114,10 @@ function updateAim(markers, W, H) {
     const a = cvSolver.anchors.get(m0.id);
     let cxp = 0, cyp = 0;
     for (const c of m0.corners) { cxp += c.x; cyp += c.y; }
-    cxp = (cxp / 4) / W * 2 - 1;
-    cyp = -((cyp / 4) / H * 2 - 1);
+    // full-frame NDC, then rescaled into viewport NDC so it is directly
+    // comparable with proj (the cropped sides are not on screen)
+    cxp = ((cxp / 4) / W * 2 - 1) / _crop.x;
+    cyp = -((cyp / 4) / H * 2 - 1) / _crop.y;
     s += `m${m0.id} px(${cxp.toFixed(2)},${cyp.toFixed(2)})`;
     if (a) {
       _aimV.setFromMatrixPosition(a.mat).project(camera3);
@@ -225,6 +229,9 @@ async function start() {
   });
   video.srcObject = stream;
   await video.play();
+  // dimensions can land after play() resolves on iOS; recompute when they do
+  video.addEventListener('loadedmetadata', updateCameraIntrinsics);
+  video.addEventListener('resize', updateCameraIntrinsics);
 
   // detection canvas
   canvas2d = document.createElement('canvas');
@@ -327,7 +334,37 @@ async function start() {
 
 function onResize() {
   renderer.setSize(innerWidth, innerHeight);
+  updateCameraIntrinsics();
+}
+
+/* Match the virtual camera to the video the user can actually SEE.
+ *
+ * Two things were wrong before: the camera was hardcoded to a 60 degree
+ * VERTICAL fov, while a 1280x720 feed at ~60 degrees HORIZONTAL is only ~36
+ * degrees vertically; and #cam is `object-fit: cover`, so on a portrait screen
+ * the sides of the frame are cropped away and never seen. Both make world
+ * geometry project too far from centre — measured at 1.7x on device, which
+ * reads as ghosts sliding the wrong way when you tilt.
+ *
+ * So: take the focal length implied by HFOV_DEG over the full video width, work
+ * out how much of the frame survives the cover-crop, and set the fov from that
+ * visible height. Cover crops symmetrically, so the principal point stays
+ * centred and no lens shift is needed. */
+const _crop = { x: 1, y: 1 };   // fraction of the video width/height still visible
+function updateCameraIntrinsics() {
+  if (!camera3) return;
   camera3.aspect = innerWidth / innerHeight;
+  if (!video || !video.videoWidth) { camera3.updateProjectionMatrix(); return; }
+
+  const vw = video.videoWidth, vh = video.videoHeight;
+  const focal = (vw / 2) / Math.tan(THREE.MathUtils.degToRad(HFOV_DEG) / 2);
+  const scale = Math.max(innerWidth / vw, innerHeight / vh);   // object-fit: cover
+  const visW = Math.min(vw, innerWidth / scale);
+  const visH = Math.min(vh, innerHeight / scale);
+  _crop.x = visW / vw;
+  _crop.y = visH / vh;
+
+  camera3.fov = 2 * Math.atan((visH / 2) / focal) * 180 / Math.PI;
   camera3.updateProjectionMatrix();
 }
 
@@ -433,7 +470,7 @@ function loop(t) {
       tracking.lastSeen = performance.now();
 
       if (engine === 'opencv' && cvState === 'ready') {
-        // v3 path: one joint solve over every visible crest (runs in the worker)
+        // v3 path: one joint solve over every visible crest
         const fused = cvSolver.solve(markers, W, H, focal);
         if (fused) {
           camera3.position.copy(fused.position);