The clips we have been building are not the entry. They are the reference implementation that de-risks it. This page says what the video is for, what we ship and when — and puts the one expensive decision in front of you.
The competition asks students to author a branching scenario with two decision levels, and specifies that video clips are required at the branch points — to visualise the scenario "and the moments before each branch point". So each clip has exactly one job:
That framing sets the bar, because of how the rubric is weighted:
Content = organisation 10 · design of choices 20 · explanation 20 · language 10. Overall = understanding developed 10 + topic relevancy 10.
A no-login folder: MP4 clips for the intro and both branch moments (<60 s each), a PDF/Word design + rationale, and the mandatory AI-usage declaration. Up to 4 competences, at least 2 domains.
Where a student goes from an idea → framework-guided domain & competence choice → branching structure → storyboard → export, with a rubric checklist. Launches ~5 Oct.
These are the reference implementation: they prove the pipeline, and they become the platform's demo content. They are not the entry yet.
Design gates production. Do not render a clip before the decisions it frames are final — every design change costs a re-render. On the 3D route that cost is ~12–13 minutes per minute of video.
The trade is believability versus cost per revision — and the rubric caps how much believability is worth. Click a route.
Believability good · cost per revision high (~4 fps software render; 90 s ≈ 12–13 min)
Believability medium (deliberately stylised) · cost per revision low (seconds)
Believability highest · cheap to re-cut, expensive to re-shoot; consent & casting
Believability high · cost per revision low; needs editing skill
"Two girls with their arms up and waving are not engaged in authentic interaction." Two of the three causes were invisible until we measured the rig — we cannot see the render, so we calibrate numerically.
| Defect | How it was found | Status |
|---|---|---|
| Both avatars used the same model (tinted copies) | By inspection | fixed two distinct CC0 avatars — a female student and a male lecturer |
| Both avatars were facing away from the camera | Measured from the toe bones (toes point forward): the 180° VRM-0 orientation flip was being overwritten by the staging yaw | fixed every frame verified facing the camera |
| "Arms down" was inverted, and the non-gesturing arm was left in the bind T-pose | Measured by reading the hand world position back while sweeping the arm rotation | fixed both arms posed every frame; all four hands rest low |
| Listener nearly frozen; gestures were emblems, not conversational | Staging research (turn-taking, backchannels, gesture taxonomy) | fixed backchannel nods, brow raise, tilt; speaker glances away and returns |
Guard scripts now exist so these cannot silently regress: verify_pose.py, verify_interaction.py, calibrate_arms.py, check_facing.py, probe.py.
Bavelas, Coates & Johnson (2000): listeners are co-narrators — a dead listener degrades the speaker's own performance. More impactful than any render upgrade. Also: a wave is a speech-independent emblem (it means hello), not a conversational gesture — which is exactly how it read.
rhubarb-lip-sync (MIT, verified running on this M2) — audio-accurate mouth shapes instead of a viseme flap.vrm-mixamo-retargeter (MIT) — genuine idle / talk / listen cycles.| # | Decision | Our recommendation |
|---|---|---|
| 1 | Which production route — 3D, 2D graphics, live action, or hybrid? | B or D for the entry; keep the 3D avatars as the platform showcase. |
| 2 | Which theme for the scenario? | Academic integrity under pressure — highest recognition for students, maps cleanly to several competences. |
| 3 | Who is on the 4-person team, and who owns the design (the 60%)? | Assign a named design owner now — it is the highest-weighted work and it gates production. |
| 4 | Does the platform MVP generate clips, or stop at storyboard + export? | Stop at storyboard + export for the 5 Oct launch; clip generation is a heavy second phase. |
| 5 | Is the platform part of our entry story, or a separate side project? | Separate. Keep the entry clean and scoreable; use it to showcase the platform. |
Full reasoning: projects/Scenario/roadmap.md. Supporting research: interaction-research.md, interaction-tools.md, sources.md, plan.md.