A medoid is the existing point in a set that minimizes the sum of
distances to every other point in that set — not a new point constructed from the
set, an existing member of it. Iris applies the same idea to a round's submissions:
the Leader Node takes every image the committee fetched for a given
Area of Interest and timestamp, and picks the one that sits closest, collectively, to
all the others.
"Closest" is measured concretely, not aesthetically. After orthorectification and
per-band normalization bring every submission into a comparable pixel space, the Leader Node scores each candidate against the rest using standard
image-distance metrics — mean absolute difference, mean squared error, and spectral
angle mapper — and the image with the lowest total distance to the group wins. There's
no machine learning model in this step and no embedding space: the metric is concrete,
and it's auditable by anyone willing to rerun it.