# Documented Manual Rubric

Use this rubric only when the evaluator report identifies a criterion that
cannot be measured reliably. Record reviewer identity and qualifications only
when a real reviewer supplies them; none are asserted here.

## Motion reproduction (maximum affected points shown in report)

Compare synchronized reference and fresh replay at normal speed and
frame-by-frame. Award proportional credit for the intended left step, right-arm
raise, two distinct waves, lowering, return, and head/torso response. Depth
differences that cannot be resolved from the single reference camera must not
be penalized unless they create visible contact, silhouette, or balance errors.
Do not require the reference's hand/shin brushing.

Ten of the 50 motion points are reserved for pose orientation and cannot be
awarded from centroid agreement. Compare matched evaluated-mesh principal axes
at the configured event frames. Use the piecewise-linear angular curves in
`config/evaluator_config.json`: body segments receive full credit through 12°
and zero at 45°; hands receive full credit through 20° and zero at 70°.
Reflection and per-frame realignment are forbidden. Record the awarded
orientation points and reviewer identity in the evaluation report.

## Visual and structural fidelity

For topology-changed or procedural geometry, compare silhouette, proportions,
all 14 recognizable parts, material regions, neutral appearance, and absence of
collapse from at least the reference view and evaluator side view. Different
mesh grouping, object order, names, and rig construction receive no penalty.

## Temporal/contact quality

Inspect support-foot planting, visible sliding, floor penetration/floating,
limb intersections, detached parts, snapping, and jitter. A controlled foot
roll is not sliding. Removing the reference's documented hand/shin contacts is
an improvement and must not reduce the score.

## Preview/replay consistency

The preview must be a faithful presentation of the submitted scene.
Antialiasing, compression, and codec differences are handled by the automated
SSIM tolerance. If the author deliberately used materially different lighting,
color management, background, or render engine for the preview, automated SSIM
is not reliable and this rubric must judge consistency of motion content,
timing, framing, and visible geometry. If no preview is submitted, this
component receives zero rather than being inferred from untrusted JSON.