demo(skill-guard): GhostCommit detection video (Video 1)

[ernanibmurtinho]

Video 1 — the security detection demo. Demo-kit house style (80×20 → 1200×676 @ 30fps, 59s). The rendered MP4 + thumbnail are at demo/kit/skillguard.mp4 / .thumb.png (regenerable, not committed as binaries).

Every verdict is real

The screenplay shells out to the shipped gecko scan-doc / scan-image against the canary fixtures — what's on screen is what the CLI returned:

• scan-doc AGENTS.md → POISON (convention-text → follow_rendered_instructions + exfil_encoded_target)
• scan-image build-spec.png → POISON (ocr → … — metadata saw nothing; OCR read the rendered pixels)
• scan-image payload.png → POISON (png:tEXt(base64) → … — enhancement (a) decode+rescan)
• scan-image architecture.png → CLEAN (a real diagram; no false alarm)

Honesty (demo-kit rules = law)

• Real execution only, no mocked output. Leak-checked: 0 secret-shaped strings in the cast.
• No over-claim: no LSB/steganography-decode claim (that's the disclosed residual). Close is the positioning line + the determinism lesson, not a promise.

Render command

uv run --with pyte --with pillow python demo/kit/render_cast.py \
  demo/kit/skillguard.cast demo/kit/skillguard.mp4 \
  --scene "GhostCommit — hidden in a picture|Gecko Skill Guard" \
  --scene "Every channel, one engine|Gecko Skill Guard" \
  --scene "Security through determinism|npx @geckovision/gecko"

🤖 Generated with Claude Code