Scrubbing needs the file to buffer first. Cloudflare Pages serves this asset without
byte-range support, so the seek bar does nothing until the whole file has loaded. Playing from
the start works immediately; the Download button lets you scrub freely in any local
player.
This page plays a compressed proxy, 720p at ~370 kbps, to fit Cloudflare's 25 MB
per-file limit. The master is 1920×1080, 311 MB, at
D:\AI\scam-defense-lab\video10\out\SDL10_master_music.mp4. Judge the captions, the
pacing and the read here; judge fine picture detail on the master.
Your four notes
"Flickering all the time." — reported three times, and I only measured it on the third.
The captions were built as one event per word, and each event only existed while that
word was being spoken. In the 50–90 ms pause between two words, nothing was drawn.
Measured on the cut you watched: the subtitle was absent from 44% of frames and switched
off 30 times in 20 seconds — roughly 800 blinks across the film.
My first two fixes found real defects (3-word groups, then 288 overlapping events) and
neither touched this, because I checked the caption track for overlap and never once
for gaps. The animation is now gone entirely: 121 plain cues, ~4.3s each, zero
gaps, zero overlaps, no colour changes, nothing moving. Same detector on the new file:
0 blinks.
"Change the font — like the other font we used." It was Georgia 68, which is this
film's card face, not its caption face. Back to what videos #1–#9 actually shipped:
Inter Black, uppercase, 46px, cream on a dark scrim.
"Subtitles should say the number, not letters." Fixed everywhere — and worth telling
you what it turned up. We already had a tool for this and I never called it. When I did, the
tool itself was broken: it dropped everything after the word "thousand", so
"one thousand six hundred dollars" would have printed "1,000 600", and "two thousand
one hundred" printed "2001 100". I rewrote its number grammar and every figure on screen is
now checked against the case file.
"The guy sounds rushed / sped up." Nothing sped it up — ElevenLabs simply returned a
faster read. Measured the same way on both files: #9 was 144 words per minute, #10 came
back 152. Stretched to 141.5, pitch untouched, and all 75 shots re-cut to the new
timing. That is why the runtime moved 8:16 → 8:52.
The music
You said the pick was mine, so I made one rather than hand it back. It is a
six-chapter bed following the plan already written for this film: thin for the cold
open, procedural through THE CLAIM, a pulse arriving at THE ROUTE where it becomes a
business, circular and unresolved through THE LOOP, thinning across THE JOB, almost nothing
from THE END on.
It sits about 30 dB under the narration and ducks further when he speaks. The three
scripted silence drops are in: before the title sting, before "they went back to work", and
before the closing line.
Generated with ElevenLabs, whose music output is cleared for commercial use including
monetised YouTube. If you'd rather have a YouTube Audio Library track, say so — the
master rebuilds without the bed in one command and you add yours at upload as before.
The one thing I cannot check: I can't hear this. Every level above is measured, not
judged — bed loudness, ducking depth, the silence windows. Whether the bed actually
suits the film is your ear, and it is the main thing I need from this pass.
Known and deliberate
Item
State
Loudness
−16.0 LUFS, true peak −1.9 dBTP. Short of the −14 target on purpose:
the narration peaks too close to full scale to lift further without limiting, which the spec
forbids. YouTube only turns audio down, so it will leave this alone.
One retimed shot
A6-05 ("both of them pleaded guilty") had 10.0s of clip for a
10.8s line once the read slowed, so that shot alone runs 7.7% slower. Measured against its own
source: the clip was already 32% near-static, the retimed version is 27%. No new judder.
Resolution
Master is lanczos 720→1080. Grok's own upscaler is better but needs
75 manual passes.
Sound effects
None. The spec calls for sparse diegetic only.
Verified on the finished file, not on the parts
All 12 cards and 4 animated graphics sampled from the assembled master at their own anchor
second and looked at — not inferred from the render exiting cleanly.
Caption blink-offs counted frame by frame on the finished file across four 20-second
windows: 0 in 1,920 frames — and the same detector run on the previous cut as a
control, where it correctly found the caption absent 44% of the time.
Duplicate-frame count measured in three 30-second windows: 0, 1 and 4 frames out of 720.
Every numeral on screen traced to the fact sheet: $179.99, $499.99, $1,600, $1,380, $10,528,
2,100, 580,000, 71 and 68 months, 2023, 2022.
Bed level sampled inside the real gaps in the read, and each silence window confirmed silent.
Picture and audio streams agree to 0.013s; the end card is intact at its full 10 seconds.