The variation harness

Point it at an interaction and a range of variation. It plays the interaction across the range and reports a matrix: every configuration against every goal, with the worst cell selectable.

Knowing when a clip is finished. The honest cost of constraints is marking up a library of clips, and what makes that cost bearable is not authoring faster — it is not having to re-check everything by hand every time an artist changes a body. An author with a matrix knows which clips are done.

It is not a renderer and it is not a playtest. Nothing it measures involves looking at anything, which is what lets the same run be a build step.

Five, because five different things go wrong:

A hand that snaps shows in the jerk and nowhere else: the residual is small on both sides of a snap, which is exactly what makes a snap invisible in a residual plot.

⚠ A goal is judged on how far it missed when it was being asked for in full. Every tag eases in, and a goal at half weight is supposed to be half satisfied — measuring either as error would flag every clip ever marked up. The run warms up for one loop before anything is recorded, and only the samples at or near a goal's own peak weight count towards its residual.

⚠ Reach and limits are separate columns, because they are separate fixes. A straight arm that is still short means the contact is somewhere this body cannot get to — answered by moving the contact. A joint sitting at the end of its range means the pose the solver wanted is one this body may not adopt — answered by widening the limit or bending elsewhere. A rig that declares no limits reports nothing in the second column, which is not the same as passing it.

The clip is the authored AnimationClipContent and not a baked AnimationClip, because a baked clip's goals hold joint indices resolved against one skeleton — varying the body means baking again, and a harness handed a baked clip would pose every body with the first one's joint numbering.

Axes multiply. Two axes are a grid and not two runs, because the failures this exists to find are at the corners: a short body and a wide prop.

Rows are configurations, columns are goals, and a cell carries the worst of each measurement and when it happened. That last part is what makes the report a tool rather than a verdict — a number saying a clip is wrong somewhere is not actionable; the same number with a configuration and a moment attached drops you on the frame.

VariationHarness.Rebuild turns a row back into the body, the prop and the ground it was, so the editor can put that on screen. The panel's Worst Cell button jumps straight there.

⚠ A goal that never resolved is a third state, not a large number. It is the most common way a variation actually breaks — the hand cannot reach the shape on the small body, the binding resolves to nothing — and it outranks any amount of error. An author who reads it as a number spends the afternoon tuning a goal that is not running.

Declare the run in the project so the thresholds are somewhere the person authoring the clip can see them and change them. Numbers written into a test are numbers nobody believes.

The importer refuses a plan with nothing to play or fewer than two samples, and warns about two things a plan invites: thresholds all left at zero — a green build that means nothing — and a run of more configurations than anybody meant to start.

A .vxharness opens in the editor too: the panel shows the declaration, Run fills the matrix, and a run that cannot start says why rather than throwing — a plan whose clip has not been imported yet is the ordinary state of one somebody just wrote.

⚠ There is no vixen animation check verb. HarnessPlanContent.Resolve turns the declaration into a run given the clip, the rig and the shapes, and in the editor all three are found for you — EditorAnimation reads the rig out of the model's own source file using the import settings on its sidecar, so the joints are the joints the import will produce. What the command line would need is that resolver on the tool side of the fence, which is not built. Today the gate is three lines in a test project, which is enough to fail a build and is honest about what it is.

IVariationSource is an axis: a name, some values, and a way to apply one to a subject.

Subjects are mutable and handed to every source in turn, so two axes compose without knowing about each other.

The build gate the section above describes, in full. A .vxharness says what to play and what to hold it to; the test is the three lines that read it and one assertion:

verdict.Summary is the message and not a count, so a failing build says which body missed which goal by how much — an assertion that printed "false" would send whoever reads it back to the editor to find out what broke.

Picking the row to look at is the same call the panel's Worst Cell button makes:

• Pose constraints — what the harness is measuring
• Move sets — the other half of doc 34