You are checking a TOP-DOWN SITE PLAN that should carry staging
markers drawn on it. You are given the expected staging (each marker
described relative to landmarks visible on the plan, with coarse
normalized coordinates as secondary estimates) and TWO images in
order: first the ORIGINAL CLEAN SITE PLAN, then the ANNOTATED SITE
PLAN to check. Compare the annotated plan against the original for
every preservation judgment.

Verify ONLY these points, judging POSITION by each marker's
landmark-relative description (the placement authority — the coarse
coordinates are only a secondary cross-check):
1. A camera marker (labelled CAM or clearly the camera symbol) is
   present, placed where its landmark description says.
2. A field-of-view wedge opens out of the camera marker: two edge
   lines reaching the two described frame-edge spots, a sight line to
   the described look-target spot lying between them, and the area
   between the edges tinted while the plan under it stays readable.
3. One subject marker per expected slot is present, each placed where
   its landmark description says (small offsets are acceptable; a
   marker on the wrong feature, wrong side of its landmark, or
   swapped slots is not), each labelled with its slot code, and each
   carrying an arrow pointing toward the spot its facing description
   names (judge the arrow by direction, not length).
4. Compared with the ORIGINAL CLEAN SITE PLAN, the annotated plan is
   otherwise unaltered: same canvas and framing, every zone, landmark
   and label of the original still present and not redrawn into
   something else, and no extra symbols, scenery or text beyond the
   specified markers were added.
Judge tolerantly on small positional offsets, wedge width and arrow
length, strictly on presence/absence, on which landmark a marker
belongs to, on which way an arrow or wedge points, and on plan
preservation.
Output JSON: markers_ok (true only when all four points hold) and
details_ko (one short Korean line naming what is wrong, or "정상").