Quality

MotionForgeQuality measures defects in an animation clip regardless of who or what made it.

It knows nothing about providers, motion definitions or generation. A limb inside a torso is a property of a clip, and a retarget onto different proportions produces the same defect with no generative model involved anywhere.

It is an add-on. Removing it changes nothing about the core pipeline.

Why measurement, not inspection

The usual way to check a generated clip is to watch it. That does not scale past about twenty clips, and it does not give a pipeline anything to gate on.

So: one number a gate can test.

The five functions

All take asset paths, all Blueprint- and Python-callable.

CreateMeasurementProxy

Builds a physics asset dense enough to measure against, fitted to a character's own vertices, with a configurable minimum bone size — default 1 cm against the engine's own 20 cm default, which produces a palm with no fingers.

It is never assigned to the mesh. The character keeps its gameplay ragdoll; the proxy exists purely to be measured against.

Measured on our own character: 114 bodies, 100% surface coverage — versus the shipped gameplay ragdoll's 21 bodies and 38.8%.

MeasurePenetration

Measures how far one shape's axis is inside another's surface.

Not surface overlap — which scores a hand resting on a shoulder the same as a hand buried in a chest, and is therefore useless as a defect measure.

Calibrated against each body pair's own rest-pose overlap, so a hand-authored idle reads 2.4 cm instead of 11.9 cm. Without that calibration every clip fails, including correct ones.

• Joint-separation filter (default 3), so adjacent bones do not report themselves.
• The physics asset's own collision-disable table is ignored by default — honouring it tests nothing on a ragdoll that disables all 210 pairs.
• Reports SurfaceCoveredPercent first. Under 90%, everything else is advertised as unmeasured rather than clean.

That last rule is the discipline: a measurement that could not see most of the body reports that it could not see, instead of reporting a low number.

MeasurePenetrationDepth

The same measurement reduced to the deepest overlap in centimetres — "the one number a gate can test".

Returns UnmeasurableCm (9999) when nothing could be measured, so an unreadable clip fails a less than threshold rather than passing it.

ListMotionClips

Every animation in a content folder, one row each — path, name, skeleton — read from the asset registry rather than from motion definitions, so it finds clips nothing generated.

CorrectPenetration

Implemented, writes a new corrected clip, and refuses outright rather than writing anything when the correction does not improve the measurement.

A correction stamps physics.depenetrate in the clip's MotionForge provenance record with the before and after numbers, so a corrected clip is distinguishable from the clip it was corrected from.

As pipeline nodes

Four of the five appear in the pipeline graph with no extra registration, through the node library:

The proxy generator has no node.

Surfaces

No dedicated panel. The pipeline nodes are the only editor surface, and all five functions are reachable from Blueprint and Python. There is no bespoke window, menu entry, console command or details customisation.

The toolset exposes three of the five as MCP tools: MeasureMotionPenetration, CorrectMotionPenetration and CreateMotionMeasurementProxy. ListMotionClips and MeasurePenetrationDepth have no tool — an agent reaches them through Python or a pipeline.

Validating the oracle

One practice worth stealing: validate a measurement by running it on something you already know the answer for.

Running the penetration measure on a hand-authored idle — content a human made, approved and shipped — is what surfaced the need for rest-pose calibration. An oracle that fails known-good content is not detecting a defect; it is a defect.