Reproducible demo videos

These silent videos visualize the frozen catalog.json. The gallery selects reviewed successful saved examples and explicitly labeled interface walkthroughs. Walkthroughs contain no invented model response. Failed and unverified videos are excluded from published media; original predictions, failed trajectories and metrics remain in the evaluation reports. This outcome-selected showcase is not a representative success-rate estimate. The catalog owns its current item counts and coverage.

Each selected item has an H.264 MP4, JPEG poster, English WebVTT captions and plain-text transcript. Technical views are 16 seconds; complete game episodes have their own durations. Playback timing is edited for readability and is not inference latency. No audio track is included. Older mixed-outcome technical montages are excluded from the selected publication.

Regeneration and selection

From the repository root, after rebuilding the source catalog and results:

python3 scripts/render_demo_videos.py --workers 3
python3 scripts/render_gameplay_videos.py --doom-capture /path/to/doom-10001
python3 scripts/attach_gameplay_catalog.py
python3 scripts/curate_demo_site.py
python3 scripts/check_site_assets.py --output reports/site/artifact-audit-showcase.json

The curation step reads reports/site/demo-outcome-review.json, keeps reviewed successes and interface walkthroughs, and removes excluded managed media. Rendering alone does not apply this publication selection. Publish with:

python3 scripts/export_site.py \
  --destination /path/to/Zefan-Cai.github.io/open-jev --prune-media

This also removes excluded files from the destination’s managed media.

Install Pillow, ffmpeg and ffprobe; use Python 3.10 or newer and a TrueType font such as Helvetica on macOS or DejaVu Sans on Linux. Font and encoder versions can affect file bytes. Rendering and replay verification are CPU only and make no model inference calls. Regenerating the source catalog needs its retained source predictions and frozen datasets; viewing committed media requires neither a model nor a service.

The current artifact audit under reports/site/ records every published video’s hash, dimensions and codec. Original generation audits remain there as historical evidence. The public site supplies native player controls, captions and transcripts.

Selected continuous game videos

gameplay-doom, gameplay-trex and gameplay-wiki show complete successful episodes with a visible goal and final outcome. gameplay-overview concatenates those episodes in that order. Its exact duration and chapter boundaries are recorded in gameplay-verification.json.

The runner and Wiki re-execute every saved action through their original local Python environments, checking observations, rewards and terminal state. Doom uses all 42 captured native ViZDoom frames from its 11-decision episode at 0.2× playback speed. The engine supplies no screen buffer after terminal, so the last available screen is held behind the result overlay. No saved steps or captured frames are omitted within a selected episode.

These episodes use fixed seed 10001 and the original 100-step 9B pilot. They were selected because they succeeded; failed Snake and platformer episodes remain in the underlying evaluation reports. No new policy evaluation is implied. gameplay-verification.json binds the source/native-frame hashes, engine matches, successful outcomes, codecs and complete overview chapters. Earlier broader-gallery audits are historical; see the report index.