# Earth Rise: generated terrain skill and live update

This rehearsal generates two ordinary Luau package revisions from an agent-authored
recipe while the server and two native cameras remain running. It does not call
a language model during replay. V1 (package revision 2) adds one radius-2-metre
terrain sphere at the skill impact. V2 (package revision 3) replaces the same
skill with three overlapping spheres spaced 2.5 metres apart, producing an earth
wall. Existing terrain is retained. The wire skill ID remains `ground-pulse`.

```powershell
node scripts/authoring/record-terrain-skill.mjs `
  --bins target/whitepaper-recording-binaries `
  --character-cache target/structural-collapse-evidence/host/authoring/characters/meshy-hero `
  --out target/new-terrain-skill-run --seconds 120 --fps 6 `
  --ffmpeg <path-to-ffmpeg.exe>
```

Requires Windows, Node 22+, mkcert, installed edge dependencies, compatible
engine binaries and an already-downloaded character cache. No paid asset job is
submitted. Use a fresh output directory. See RECORDING_REHEARSAL.md for setup.

Narration: two clients join the shared world; the agent generates and publishes
Earth Rise V1; a player casts it and the ground grows; the agent generates V2
and publishes it through optimistic generation 2; another cast grows a wall.
Both clients remain connected throughout the two activations. Cast and impact
use gentle green upward motes, without explosion, smoke or blast impulse.

All publication uses AuthoringRuntime, admission, CAS, composite checkpoint and
tick activation. Casts use player socket messages. No admin world edits occur.
The test requires authoritative skill outcomes and both native clients presenting
cast/impact cues for both revisions. Native swapchain reports must progress from
0 to 1 to 4 terrain brushes and both cameras must agree on each stage's terrain
hash. The host must remain serving until recording finishes.

The edited clip is `whitepaper/construction-stacked-1.25x.mp4` (the shared
recorder's filename): aligned top/bottom native streams at 1.25x, without audio.
The recording demonstrates adding terrain and changing an existing skill live;
it does not establish player damage, movement over the mound, or live model
inference. Source and release receipts identify what was generated and admitted.