A frame of stray cats in a park, roughly fifty of them, and one question: how many? The easy ones sit on the gravel path; the hard ones lie in dark leaves, behind a bench, on top of the sleeping man, or on the far lawn as shapes that may or may not be cats. A plain count with a breakdown by area and an honest range.
Score the number first, the breakdown second, the range third.
The total. There is no published count; the reference is a consensus of the frontier models in the frame-only round plus the operator's own check, with the ambiguous shapes listed. An answer inside the consensus range is a hit. Scoring is by distance from that range.
The breakdown by area shows where a wrong total came from. Most misses should be in the leaves and on the far lawn; a model that is short on the gravel path, where every cat is large and separate, did not look.
The range. A narrow range around a wrong total is overconfidence. A range that contains the reference is calibration, even if the point estimate is off.
Two inputs are used. The frame-only crop is the test. The title variant adds the video's title underneath ("...Hundreds of Stray Cats"), and the only question it answers is whether a model's count moves toward the caption.
Count the cats in this photograph.
Count every cat you can see, including cats that are partly hidden behind something, partly cut off by the edge of the frame, or small in the background. Do not count people or anything that is not a cat.
The total on its own line, in exactly this form:
CATS: <number>
How many of your total are in each part of the picture, one line each:
List every shape you were unsure about: where it is in the picture, what made it hard, and whether you counted it. Then give the lowest and highest totals you consider plausible, in exactly this form:
RANGE: <lowest>-<highest>
Count from the picture itself. A careful count with an honest range is what I am after; a round number with no breakdown under it is not.