# The device kit
Nine ways for scroll to change the page. Each one is a different answer to "what
does the visitor's hand actually do here."
Pick per beat, never per page. The variety law from SKILL.md Step 2 applies:
four or more families, never the same one twice in a row.
Every act publishes `--sc-p` (0 to 1) on its own element, so anything you want
to drive that the kit does not cover, you can drive from CSS with `calc()`
against that variable. Reach for that before asking for a new device.
---
## 1. `scrub`: the wheel is a scrubber
The anchor device. A pre-rendered camera move plays under the reader's hand,
one frame per notch. This is the thing people screenshot and send to each other,
so spend it on the open.
```html
Your morning shouldn't need two drinks.
```
- `data-sc-span` is the act's scroll length in viewport-heights. 2.2 to 3.0 for
a hero. Below 1.8 the clip flies past; above 3.5 the reader starts wondering
whether the page is broken.
- `data-sc-dwell` (0 to 0.6) remaps time so the camera settles mid-act, exactly
where the copy peaks, and moves quicker at the edges. It is the difference
between a clip that plays and a shot that lands. Keep it at or below 0.6.
- `data-sc-src` (not `src`) is deliberate: the engine fetches the clip as a Blob
so it seeks without needing HTTP range support, and skips the fetch entirely
under reduced motion.
- The poster is a live frame-holder. It stays up until a real video frame has
painted, because iOS keeps a seeked-but-never-played muted video blank and
hiding the poster on metadata alone flashes an empty stage.
**At most two scrub acts per page.** The third one is no longer a surprise, and
it is the heaviest thing on the page.
### Clip time is not cue time
The single most damaging bug this device has, and it is invisible in every
screenshot taken one at a time.
A pinned stage is on screen for **one viewport before** its pinned travel begins,
sliding up into view, and **one viewport after** it ends, sliding off the top.
The act's progress `p` is 0 through the whole entry and 1 through the whole exit.
So a clip driven by `p` sits frozen on its first frame while it slides in, and
frozen on its last frame while it slides out. The reader has been scrubbing a
film with their hand, the film stops, and then the whole page slides a still
photograph past them. It reads as the site breaking, and it is the fastest way to
make an expensive page feel cheap.
The engine therefore maps the clip across the stage's **entire visible life**, not
across its pinned travel, and this is the **default**. Both ends are clamped to
scroll that actually exists, so a hero at the top of the document still starts on
frame one and an act near the bottom still reaches its last frame. Cues keep
using `p`, because cues belong to the pin.
**Pair it with `data-sc-dwell`.** Dwell moves quickly at the edges and settles in
the middle, which is exactly the shape this mapping wants: the fast motion lands
on the two slides, and the settle lands inside the pin where the copy is. The two
were built for each other.
`data-sc-clip-map="travel"` restores the old pinned-travel mapping. There is
almost no reason to reach for it, and reaching for it reintroduces the freeze.
The harness checks this now (see [verify.md](verify.md)), so a frozen clip fails
verification instead of shipping. Do not rely on noticing it by eye: every
individual frame of a frozen clip looks completely correct.
### The playhead is lerped
Scroll never writes `currentTime`. It writes a target, and a standalone rAF loop
walks the clip toward that target at a fixed fraction per frame. Wheel events do
not arrive at a constant rate, so a 1:1 write reproduces every gap in them and
the clip reads as a stutter rather than a glide. Three mechanisms, all on by
default:
- **Lerp 0.18 per frame.** `data-sc-lerp` overrides it, on the mount root for the
whole page or on one `