/vox-explainer

Vox Collage Motion

Two speeds exist: ELEMENTS move in 12fps micro-holds over smooth eases; CAMERA moves smoothly. Hand-authored sparse tl.set holds (3-10Hz) read as chop — that is the failure this skill prevents.

The stepEase helper (elements)

const stepEase = (base, dur, fps = 12) => {
  const e = gsap.parseEase(base);
  const n = Math.max(1, Math.round(dur * fps));
  return (p) => e(Math.floor(p * n) / n);
};

Pure function ⇒ deterministic, seek-safe, render-safe. Card pops: fromTo(scale 1.10→1.0, 0.3s, stepEase("back.out(1.6)"), immediateRender: false) after an autoAlpha set. Block settles: one tween 1.24→1.0 over the beat, stepEase("power2.out").

Camera moves (always smooth)

• Zoom-isolation push: power4.in, ~1.2s, scale to the measured overflow (4.5-7×), translate x/y in the SAME tween so travel shares the ease.
• Inverse zoom-through (arrival): expo.out ~1.1s from scale ~3.2 with the target cell pinned — the previous beat's subject docks as a grid cell.
• Plate pushes: linear 3-4% over the beat (camera intent, not idle).
• Drive-pasts (viewer is moving): power3/4.in, blur ramps with speed, and the object VANISHES the frame it fills the screen — never hold a blurred wall.

Seam mechanics (zoom-isolation swap)

1. End the push tween ON the last pre-swap frame — a tween ending between frame samples never renders its final state.
2. Align EMPIRICALLY: render last-pre and first-post frames, mask-detect the carrier in both, and correct the tween's x/y by the measured delta (1:1). Use unclipped edges when the carrier overflows frame. Repeat until centers agree within ~30px and sizes within ~10%. Never trust origin math — helpers may override transformOrigin (script order wins).
3. The incoming cutout enters at 0.90-0.93 still GROWING (scale-velocity sign carried), decelerates in ~0.5s, then creeps ~1%/s (camera intent).

Choreography rules

• Show-before-tell: when VO is slow to the point, the visual may lead by ~1-1.5s (tags, morse taps, anchors). Reactions start ~100ms BEFORE their cause completes; never after.
• Word-synced events land ON word timestamps from the transcript. Decoupled sequences (promoted text staircases) run their own fast cadence instead of waiting for VO.
• Stillness before climax: 0.3-0.75s dead hold before a reveal is deliberate; longer than ~3s is a planning bug.

Gates

• Dead-time: MAD every consecutive frame pair of the render; no still run

  3s. Fill with staged reveals/camera intent — NEVER idle wobble (motion-continuity.md still governs). Measure at >=640x360 with TWO metrics — mean abs diff AND a changed-pixel count (pixels moving >12 levels). A frame pair is still only if BOTH are negligible (mean <0.25, changed <150px). At 320x180 mean-only, a 30px text reveal vanishes into the average and the sweep invents dead time that is not there — verify any reported run by snapshotting inside it before "fixing" it. Slow camera creeps (<=2% over a beat) also score ~0: they are legitimate intent, not motion the gate can see.

• Seam measurement as above, per zoom swap.
• Keyframe sheet: one rendered frame per beat, reviewed by eye — MAD alone misses blank-but-white failures.

Event-density gate (timeline-side; the pixel sweep cannot replace it)

The frame-difference (MAD) sweep is blind to creep-only stretches: a slow camera drift defeats the threshold while reading as dead air. So verify density on the TIMELINE, not pixels: no 3.0s window of any beat without a discrete authored event — a pop, a swap, a highlight igniting, a zoom starting, a caption promoting. Walk the timeline's event times per beat and log the largest gap. A beat that holds one composition for its whole duration with only camera drift is a planning bug: add a staged reveal.