Everything from yesterday still applies — plus a fourth dimension that makes video the hardest, costliest data you'll ever lead
Day 22 of 66
Don't throw away yesterday. A video frame is still judged on prompt adherence, aesthetics, technical artifacts, text, and a safety gate. Everything you named for images carries over. Video simply adds a dimension that has no image equivalent — the relationship between frames over time — and that single addition is what makes video labeling slow, expensive, and the place a text-native lead is most out of their depth. It's where your background pays off most.
For a still, you judge what is in the frame. For a clip, you also judge what happens across frames: does the subject stay the same person, does motion look natural, does anything flicker or melt between one frame and the next. A perfect frame inside a broken sequence is still a broken clip.
Across the clip, does the same character keep the same face, hair, clothing? Does the room keep the same layout, the same number of windows? Models drift: a coat changes color mid-shot, a background object appears and vanishes. Consistency failures are obvious to a trained eye and invisible in any single frame — you only see them over time.
Does movement obey physics and intention? Smooth, weighted, plausible motion versus stutter, teleporting limbs, sliding feet, unnatural speed-ups. This is the dimension a still image literally cannot have, and it's deeply tacit — you feel "wrong motion" before you can articulate it. Naming it is the lead move.
The temporal cousins of image artifacts: textures that shimmer frame to frame, faces that morph between identities, objects that warp as the camera moves, lighting that pulses. These live in the transitions, so frame-by-frame stills can hide them entirely.
A clip may be three shots. One shot can be flawless and another broken. Decide up front whether a label applies to a single shot or the whole clip — and make raters tag which moment failed. "The clip is a 3" is useless; "shot 2 has a morphing-hands failure at 0:04" is actionable.
A still is one judgment. A 5-second clip at 24fps is 120 frames — and the failures you care about live between frames, so you can't just sample one and call it. Watching costs real time: a rater spends the clip's full duration (often several passes) per label, versus a glance for an image. That's an order-of-magnitude cost difference, and it compounds across thousands of items.
You cannot label every frame — it's infeasible and it's wasteful. But you also cannot judge a clip from one frame, because temporal failures are invisible in a still. A lead designs a middle path that catches temporal failures without exploding cost. That design is the deliverable of this lesson.
Here's the workable middle path — the thing you'd actually put in a guideline so raters don't drown:
This is the same decomposition logic from your image rubric, with sampling layered on top so cost stays sane. The safety gate still applies to the whole clip — one violating frame fails it. And the underlying principle (does the output honor a stated standard, frame after frame) is the multimodal generalization of the Constitutional AI idea you met in Week 2.
You'll fold these temporal dimensions into the finalized multimodal rubric on Day 25.
An evaluator watches a clip and feels the motion is wrong. A lead designs the sampling strategy that lets a hundred raters catch temporal failures without watching every frame — turning a 10x cost problem into a written procedure that holds coverage while keeping the budget sane.
Say this in an interview: "Video labeling is an order of magnitude costlier than images because the failures live between frames, so I don't let raters label every frame or judge from one. I write a watch-once-plus-sample procedure that catches temporal breaks and still scales to thousands of clips."