U19 creates the GSAP-main baseline corpus and a per-item PSNR compare tool for the GSAP-to-anime.js migration. These scripts do not port or render anime.js content. They render the pinned GSAP fork point through the built HyperFrames CLI and compare later candidate videos against those baselines.

Baseline Directory

Default:

/Users/miguel07code/dev/hyperframes-animejs-eval/baselines

Override with either:

HYPERFRAMES_GSAP_BASELINE_DIR=/path/to/baselines
node --import tsx scripts/eval-gsap-vs-anime/render-baseline.ts

or:

node --import tsx scripts/eval-gsap-vs-anime/render-baseline.ts --baseline-dir /path/to/baselines

Layout:

baselines/
  index.json
  render-baseline.log
  17b852784bf3/
    block/<name>.mp4
    component/<name>.mp4
    example/<name>.mp4
  second-baselines/
    17b852784bf3/<item>-<timestamp>.mp4

The manifest is keyed by registry item key plus fork SHA and records item name, kind, relative video path, fork SHA, render timestamp, fps, duration, dimensions, status, and any render error.

Render Baselines

Fresh run, including the required clean build:

node --import tsx scripts/eval-gsap-vs-anime/render-baseline.ts

Resume after an interruption without rebuilding:

node --import tsx scripts/eval-gsap-vs-anime/render-baseline.ts --skip-build

Render one item:

node --import tsx scripts/eval-gsap-vs-anime/render-baseline.ts --skip-build --only block/data-chart

The renderer discovers 147 registry items: 109 blocks, 25 components, and 13 examples. It writes pending manifest entries up front, renders serially with the built CLI (node packages/cli/dist/cli.js render ...), updates the manifest after each item, logs to stdout and render-baseline.log, records failures, and continues.

Examples are rendered from index.html. Placeholder __VIDEO_DURATION__ tokens are patched from the registry duration, placeholder video/audio tokens are removed, and awkward deck-style examples use the authored or inferred root span. If no duration can be inferred, the documented fallback is 5 seconds.

Compare One Item

Compare explicit videos:

node --import tsx scripts/eval-gsap-vs-anime/compare-item.ts \
  --baseline /path/to/gsap.mp4 \
  --candidate /path/to/candidate.mp4

Compare a candidate against a manifest baseline:

node --import tsx scripts/eval-gsap-vs-anime/compare-item.ts \
  --item block/data-chart \
  --candidate /path/to/candidate.mp4

The compare CLI emits a JSON verdict to stdout and exits non-zero for "damaged". It samples 100 checkpoints by default, matching the existing regression harness pattern: one checkpoint per 1 percent of the shared video duration, with ffmpeg's psnr filter applied at the selected frame.

Damage Criterion

Hard failure: any evaluated checkpoint below 30 dB is damaged.
Low band: checkpoints from 30 dB inclusive to below 45 dB trigger a lazy second-baseline check.
Second-baseline waiver: if the second GSAP baseline is also below 45 dB at the same checkpoint, that checkpoint is waived as inherent render noise.
Screening flag: average checkpoint PSNR below 50 dB sets screening_flag: true, but this is informational and not a hard failure by itself.

When --item is present and a checkpoint enters the 30-45 dB band, the compare CLI lazily renders a second GSAP baseline through the built CLI. Disable that with --no-render-second-baseline, or provide an existing file with --second-baseline.