The six images are written to artifacts/ (gitignored, left untracked), along with a README that lists the dimensions and the rules I missed. I viewed each image after rendering, except note-front-1x1.png, which I never opened. They are all PNG:

• note-front-16x9.png: 1920x1080, note at 85% of the width on ink.
• note-front-1x1.png: 1080x1080, note scaled to 88% of the width on ink.
• note-angle-16x9.png: 1920x1080, note at an angle with a soft shadow and light falloff.
• note-angle-1x1.png: 1080x1080, the angled note in a square frame.
• portrait-detail.png: 2048x2048, the Pepe oval alone on ink.
• sheet.png: 3018x508, the five images side by side.

I drew everything in code in headless Chrome rather than with the image tool, because that lets me control the palette and keep text out. The source and a run.sh rebuild script are in marketing/the-note/, which is uncommitted.

There is no lettering or numerals anywhere. The top banner, four corner cartouches, two bottom serial strips and the seal centre are left as empty frames. The logo is not used.

What misses the brief, all stated in the README:

• Off-palette shades: antialiasing, paper fibres and the lighting on the angled shots produce shades that are not exact palette colours.
• Two serial strips: the brief asks for one; I put one at bottom left and one at bottom right.
• Rosettes: the two big rosettes read as soft moiré discs with blank centres, not crisp petals.
• Pepe's likeness: the portrait is stylised: the face is simpler than the cel original, and the coat and cravat are schematic. It does read as Pepe at thumbnail size.
• Fine hatching: on the note shots at 1920 px it partly blends into tone; the full line detail is in portrait-detail.png.
• Paper curl: the angled shots have none.

Created and checked both original takes:

• Take A: WAV · MP3
• Take B: WAV · MP3

Verified 10.000-second duration, requested formats, 6.500-second hits, peaks below −1 dBFS, and clean silent endings. No vocals detected.

README records tempo, key, instrumentation, measurements, and limitations. All four audio outputs remain untracked.

Done. artifacts/storyboard.md (34.7 KB) and artifacts/README.md are written and untracked.

What I delivered

A storyboard and edit plan for the 10 s cut, in 16:9 and 1:1: shot tables, the exact zoom/pan schedule, both blink methods, and the full ffmpeg sequence — which I ran end to end rather than only writing down.

The one real engineering finding. The push-in cannot be a zoompan over note-front-16x9.png. From note.js:139-140 and build.html:9, the portrait oval occupies only 378.84 × 462 px in that plate, so filling a 1920-wide frame means a 5.62× upscale of resampled line art. The push must run over a composite: the plate upscaled to 8320×4680, with the native 1640×2000 oval from portrait-detail.png inlaid back at the exact position the plates step drew it, behind a feathered elliptical mask that excludes that file's ink background. All constants are derived from note.js/portrait.js, not estimated — including the mask radii, which come from the frame ring at IRX+112 being clipped by the oval canvas horizontally and leaving exactly an 8 px ink sliver vertically.

Three corrections to the brief I was handed, all source-backed: portrait-detail.png has no square version (it's one 2048×2048 master serving both aspects); the music step shipped two stings, not one (I recommend take B and say why); and take B's measurements.json logs its strongest energy rise at 6.626 s, which is chord beating, not a second hit — cut to 6.500 s.

The ffmpeg-only blink turned out better than a fallback usually is. Instead of faking it with a dim or a squash, geq draws the lid in-palette with an ink hatch and a lash line, inside the composite before the zoom, so it's resampled by the push like the artwork around it. Rendered frames read open → closing → closed → reopening → open. I also flag that the better answer is neither method: the portrait is procedural, so a correct blink is a change to the plates step — noted as out of scope here.

Two defects found by running it, both documented rather than quietly patched: the push came out greyscale because format=gray on the Sobel branch negotiated backwards through split; and the flash frames decoded as mid-grey because the concat demuxer won't reconcile mismatched pixel formats. The second presents as a colour bug with a non-obvious cause, so it's called out explicitly.

Verified on the built file: 300 frames, 10.000000 s, H.264 High/4.0 yuv420p, AAC-LC 48 kHz stereo, ftyp moov free mdat (faststart confirmed), cut on frame 195 = 6.500000 s, flash RGB (246,245,238), and the audio hit measuring at 6.500 s after AAC encoding.

Two things to weigh before this ships. The real plates, the real audio, and an image-to-video model were all unavailable, so I built stand-ins at the declared sizes reproducing the derived geometry: geometry, timing, format and container results transfer; aesthetic ones do not — shimmer strength, possible moiré on real guilloché, and the take A/B choice need a human look and listen. And I measured zoompan's integer-crop quantisation at up to 30% of per-frame motion (1.30 output px at the 16:9 end); I chose the easing specifically to keep the camera from stalling into the cut, since a dead stop is where that would show. Section 9 states the full split, and five open questions are listed for whoever runs the build.