Conference Paper Video

Use for a narrated conference video derived from an academic paper. The source is an editable HyperFrames HTML timeline; the MP4 is a validated delivery, not the authoring source.

Stage: enhance

Expand the paper into a 10-14 scene English narrative lasting 300-600 seconds. Choose the target duration from the paper's complexity, evidence density, and explanation needs; preserve an already-selected target during repair or resume. Preserve the paper's claims, terminology, figures, tables, and provenance.

Stage: plan

Create ordered html_artifact scene frames with explicit durations and English speaker_notes. Use 1920x1080 at 30 fps. Select male or female; the runtime records the corresponding deterministic Kokoro voice id. Read delivery_contract before authoring. Write local-only HyperFrames HTML with native text and local source figures. Use exactly one data-composition-id root; static scenes must set data-no-timeline, while animated work must register a seekable window.__timelines entry. Every timed scene and narration audio element must have literal class="clip"; data-hf-clip alone is invalid. Do not use requestAnimationFrame, remote scripts, styles, fonts, images, media, iframes, data URLs, or network APIs. Keep all animation deterministic and seekable.

Keep all 10-14 scene sections directly in project/index.html; do not move them into data-composition-src files to address file-size or track-density warnings. Those lint warnings are advisory, while the ordered root scenes are part of the strict delivery contract. Use renderer-supported system fonts, or provide a local @font-face source for every custom font.

Stage: critique

Reject placeholders, external assets, missing narration, fewer than 10 or more than 14 scenes, timelines outside 300-600 seconds, lint errors, render errors, stale MP4 files, and incomplete media probes.

Stage: repair

Repair every blocking contract failure without restructuring valid root scenes in response to advisory lint warnings. Then require index.html, English transcript, English SRT and VTT, Kokoro voice metadata, a fresh MP4, ffprobe evidence, and delivery_manifest.json. Whisper may compare rendered speech with the canonical transcript as optional QA, but delivery cannot depend on Whisper.