Exaggeration: Estimated, Not Penalized

"A dead-literal drawing is often a dead drawing. A grader that punishes the push is grading toward the corpse."

The Boundary of Procrustes

The last lesson's Procrustes trick is beautiful, but it has a precise edge. It removes position, size, and tilt — the similarity transform — and no more. A deliberate stretch survives it: if you pushed a gesture past literal, exaggerated the reach or the tilt to give the pose life, your construction still differs from the reference after alignment. A naive grader reads that leftover difference as error and marks you down. But it isn't error. It's the exact thing that separates a living drawing from a timid copy. Procrustes found the boundary; now the grader has to answer for what lies past it.

Estimate, Don't Penalize

The answer is a quiet act of respect. Instead of docking you for the stretch, the grader estimates how much you exaggerated — a single factor, call it k-hat — treats it as a parameter of your drawing, and factors it out. It fits your exaggeration, reconstructs what you were actually reaching for (the reference, pushed by your k-hat), and grades your construction against that. You are scored on whether you saw the solid consistently under your own push, not on whether you drew it flat and literal. The exaggeration is reported back to you as information — never subtracted as a penalty.

Why This Is Humane, and Also Correct

This isn't just kindness; it's better modeling. A grader that penalized exaggeration would systematically reward the mediocre — the cautious, literal copy would out-score the bold, alive one, and the tool would be quietly training timidity. By estimating the push instead, the grader aligns its incentive with what good drawing actually is: it rewards a strong, consistent gesture as much as a strong, consistent literal read. Humane and correct turn out to be the same choice here, because punishing a legitimate stylistic dimension was never accurate in the first place.

Model What They Did, Don't Dock the Deviation

Underneath is a principle that reaches far past drawing: when a deviation from your reference is a meaningful choice rather than a mistake, model it as a parameter instead of penalizing it as error. It's the difference between a rigid rubric and an intelligent one; between measuring a fixed distance from an ideal and fitting the richer model the person was actually working in. Loomis performs a three-way separation — align away the pose you don't care about, estimate the exaggeration you shouldn't punish, and grade the construction accuracy that's left. Getting that decomposition right is what lets an exact answer key live comfortably alongside expressive art.

Code

Estimate the style as a parameter; grade under it·python

# Procrustes removed position/size/tilt. A deliberate STRETCH survives it.
# Naive: treat the leftover as error. Humane AND correct: estimate it, factor it out.
def grade_with_exaggeration(yours, reference):
    aligned = procrustes_similarity(yours, reference)    # remove POSE (a nuisance)
    k_hat   = estimate_exaggeration(aligned, reference)  # fit the STYLE as a factor
    modeled = apply_exaggeration(reference, k_hat)       # what they were reaching for
    return construction_score(aligned, modeled)          # grade under THEIR push

# Scored on seeing the solid CONSISTENTLY, not on drawing it dead-literal.
# k_hat is reported back as information -- never subtracted as a penalty.
# Three-way split: pose (aligned away), exaggeration (estimated), construction (graded).

Exercise

Find an evaluation system you know that penalizes a legitimate stylistic or strategic choice as if it were an error — a writing rubric that docks any deviation from a template, a code linter that flags a valid idiom, a fitness score that assumes one 'correct' form. Redesign it: what parameter would you estimate to model the choice, so the evaluation grades skill under that choice instead of punishing the choice itself?

Hint

The tell is a rule that measures distance from a single ideal and calls all distance 'wrong,' even when some of that distance is a defensible decision. The fix mirrors Loomis: identify the dimension that's actually a choice (voice, idiom, style, strategy), estimate where the person landed on it, and score their execution relative to that — not relative to the one blessed default.