You are given ONE image-generation prompt, the REFERENCE image(s)
that were attached to it (labelled), and {count_word} candidate images,
labelled {label_list} in the order attached, all generated from that
exact prompt with those exact references.
Pick the ONE candidate that most faithfully realises the prompt.

LOOK BEFORE YOU SCORE. For every candidate, write out what you
actually see on four axes, in the `readings` field, BEFORE you give
it a number. These are the axes where scoring-without-looking has
failed in practice — a judge gave a candidate whose muzzle pointed
nowhere near its target the top score, and never mentioned the aim.

DIRECTION — where every gaze, weapon muzzle, pointed object and
travelling body is aimed, and whether that aim lands on the thing the
shot text says it lands on. Name the target you see. "He aims at the
man on the dock" is an observation; "composition is good" is not.

BUILT SPACE — when the shot is inside or around a made structure,
count and place its parts. Which fixed fittings does this space have,
how many of each are visible, and is that the right number? Where is
each person standing or seated, and does that position work with the
fittings around them? Do those fittings face the way this camera and
that position require? Is any reflection optically possible from this
camera position?

ENTITIES — for every object and person the prompt names, say whether
the thing in the picture IS that thing. A banknote has a country and
a denomination. A pendant has a shape. A person has an apparent
ethnicity, age and sex. Where the prompt or a reference fixes any of
those, check them one by one and say which ones match.

PHYSICS — for every body and object off the ground or in motion, say
what holds it up and whether the pose could be produced by the action
the shot text describes. Name the support you see: the foot still on
the ground, the hand gripping something, the impact that threw it, the
surface it rests on. A body in flight has a launch — a person mid-leap
shows the bend and thrust that produced it, and lands somewhere. A
held object has a hand on it. Say plainly when you find none: "nothing
supports it" is the reading. This axis is easy to skip because a
floating figure can look composed and well lit; measured on this
pipeline, three of four judges scored such a candidate highest and not
one of them mentioned that the man was lying horizontally in mid-air.
A shot text that says a body is in the air does NOT excuse this — it
makes the support the thing you must find. Note also that a pose
copied from a reference and merely rotated (a walking figure turned on
its side) is not a leap.

Then judge ONLY fidelity to the prompt and its reference
instructions, in this STRICT priority order — a higher item always
outweighs everything below it:
1. HARD VIOLATIONS (disqualify regardless of anything else):
   collage/panels, duplicated or extra bodies, invented people or
   objects, leaked markers/diagrams/text, physically impossible
   anatomy or staging. A duplicated fitting (two of something the
   space has one of), a person placed where the staging does not put
   them, and a reflection that cannot occur from the stated camera
   position are all HARD violations of this kind.
   So is a body or object your PHYSICS reading found nothing to
   support — a figure hanging in the air with no launch and no landing,
   an object floating with no hand on it. File that one whenever the
   reading says nothing supports it; do NOT file mere awkwardness or a
   pose you find ungainly.
   When the prompt states structure facts with a
   vertical (up/down) direction for a stair or route, a candidate
   that renders the opposite direction is also a HARD violation; and
   when the prompt commits the camera to a stated position on or in
   the structure, rendering that committed viewpoint from ground the
   structure does not have (for example turning a stated
   landing-level view into one from above the topmost level) is a
   HARD violation too. An aerial, crane or elevated exterior view
   that the prompt itself stages is NOT a violation.
   List each one you find in that candidate's `hard_violations`.
   `hard_violations` is for disqualifying faults only — things that make
   the frame unusable. A detail that merely differs from the reference
   (a roof outline not perfectly identical, a pendant shaped a little
   off, a background element shifted slightly) is NOT a hard violation:
   say it in that candidate's verdict instead. Measured on this
   pipeline, one judge filed 139 such minor differences as hard
   violations against another judge's 17, which drowns the real ones.
2. THE SHOT TEXT'S STAGING — the single highest positive axis:
   the exact framing and shot scale the shot text states (close-up,
   wide, insert, over-the-shoulder and the like), the camera axis
   and viewpoint, subject size and crop within the frame, and the
   exact moment/action. A candidate that realises the shot text's
   framing beats any candidate that shows MORE of the scene, MORE
   of a body, or MORE contract items by widening or reframing.
3. Placement and action relations between the people/objects the
   framing includes, time of day, and the LOCATION lock (the place
   must be the one the reference shows — same architecture,
   materials, fixtures).
4. Identity, wardrobe and props — judged ONLY for what is actually
   visible inside the chosen framing.
5. Pose/immobility and carried-state clauses — judged ONLY to the
   extent the framing shows those body parts or items. NEVER
   penalise a candidate because the framing (correctly) excludes a
   body part, garment or carried item; NEVER reward a candidate
   for widening the frame to show them.
Consistency with the attached reference images counts as prompt
fidelity within its priority level. Ignore generic aesthetic appeal.
Check every candidate point by point before deciding; base every
verdict only on what is visible.

A candidate that reads wrong on DIRECTION, BUILT SPACE or PHYSICS
cannot outscore one that reads right on them, however handsome it
looks. These three do not trade against each other: a candidate that
aims correctly but floats unsupported has not won the aiming point —
it has failed a different axis, and both go in your verdict.
If every candidate fails an axis, say so in `all_candidates_fail`
rather than pretending the least-bad one is good — the caller uses
that flag to decide whether the shot has to be taken again.

Also give every candidate an integer score 0-10 for that same
fidelity.
Output: winner, ranking best-to-worst, per-candidate readings on the
three axes plus any hard violations, and per candidate a score plus a
one-line Korean verdict citing the decisive prompt points.
Write every prose field in Korean.