Three Sub-Scores, Not One Grade

"A single grade ranks you. A breakdown teaches you. The grader is built to teach."

Three Axes of Seeing

Once Procrustes has aligned away the global pose, three things about your construction can still be wrong — and they are genuinely independent skills, so the grader scores each one separately. Angle: are the tilts and orientations between parts correct — the set of the jaw against the brow, the turn of the box? Proportion: are the relative sizes right — cranium to face, upper arm to forearm? Placement: are the internal parts positioned correctly relative to each other — do the eyes land on the right third, does the far edge sit where it should? Three faculties, measured on their own axes.

Why Not One Number

Collapse those three into a single grade and you've thrown away everything useful. '72%' tells you that you were off; it does not tell you where, and 'where' is the only part you can act on. Two students can score an identical 72 while being wrong in completely different ways — one nails proportion but skews every angle, the other has perfect angles and mangled proportions — and they need opposite practice. A single number hides that difference; the sub-scores expose it. The breakdown is what turns a grade from a verdict into a diagnosis.

The Composite Is Advisory

The grader does compute a composite score, but it labels it advisory on purpose. It's a rough roll-up for a glance — a way to sense the overall attempt at a distance — explicitly not the real signal. That labeling is a design decision with teeth: it steers you away from optimizing the one number (which teaches nothing) and toward the sub-scores (where the learning is). An advisory composite says, in effect, 'here's the summary if you want it, but the three numbers under it are the ones that will make you better.'

Feedback That Says Where

Underneath the drawing-specific detail is a universal truth about feedback: good feedback localizes. 'Wrong' is weak. 'Wrong here, in this specific way' is strong, because it points at the next action. It's the same reason a per-landmark reprojection error beats one global number, the same reason a stack trace beats 'it crashed,' the same reason a code review that names lines beats a thumbs-down. The three sub-scores are Loomis localizing the error for you — refusing to let the grade be a mood, insisting it be a map.

Code

Three independent faculties, plus an advisory roll-up·python

# After Procrustes aligns away global pose, three independent faculties remain.
def grade(yours, reference):
    a = procrustes_align(yours, reference)
    sub = {
        "angle":      angle_score(a, reference),       # tilts BETWEEN parts
        "proportion": proportion_score(a, reference),  # relative sizes of parts
        "placement":  placement_score(a, reference),   # internal positions of parts
    }
    # The composite is a GLANCE summary -- explicitly advisory, not the signal.
    sub["composite"] = advisory(weighted_mean(sub))
    return sub

# Two attempts can share a composite of 72 yet need opposite practice:
#   {angle: 0.95, proportion: 0.49}  vs  {angle: 0.49, proportion: 0.95}
# The sub-scores tell them apart; the single number never could.

Exercise

Pick a single-number score you currently optimize — a test grade, a code-quality score, a fitness metric, a KPI. Design three sub-scores that would turn it from a ranking into a diagnosis: what independent faculties or dimensions is that one number secretly averaging together? Then decide which sub-score, if it were the weakest, would most change what you'd practice next.

Hint

Almost every single number is a blend of independent things — a test grade averages topics you know cold with topics you don't; a KPI blends acquisition with retention. Splitting them out is what tells you where to spend effort. If two very different profiles can produce the same composite, that composite is hiding the diagnosis, and sub-scores are how you get it back.