Prove It on the Media You Actually Own

"Every required class records load, decode mode, color, A/V sync, seek accuracy, and resume — on real media, before versus after."

The Synthetic-Test Comfort Zone

It feels rigorous to generate a clean H.264 clip, play it in CI, and call playback "tested." It isn't. Real archives are full of the media that actually breaks players: variable frame rate, 10-bit HDR, an AV1 file with a weird color tag, a damaged transfer, a 40 GB movie on a flaky SMB mount, subtitles in three formats with off-by-one timing. A player that only meets synthetic fixtures has proven it works on the media you don't have.

Ashen Reel's release bar is a golden corpus: Dad's own representative files, each classified, each with an expected result. The corpus records dimensions and expectations — codec, container, bit depth, HDR type, subtitle layout, audio layout — and never a single media byte. The bytes stay on the NAS; only the knowledge about them lives in git.

Why no bytes in the repo: the media is Dad's family archive — private, huge, and not yours to redistribute. The corpus manifest is a test spec, not a media library. It stores redacted identifiers, dimensions, and expected outputs so the test is reproducible without ever shipping the film.

The Corpus Dimensions

The required classes are chosen because each one is a place players quietly fail: H.264, H.265/HEVC, and AV1; 8-bit and 10-bit; SDR, HDR, and Dolby Vision; MP4, MOV, MKV; embedded and external subtitles in SRT and ASS; multichannel audio; variable frame rate; damaged media; very large files; and a mounted-SMB interruption. Each class records how it decoded, the color/HDR result, A/V sync, seek accuracy, and whether resume behaved.

The Honest Gap

Here is the beat that separates a real release from a press release. When the corpus was captured, no Dolby Vision file existed anywhere in Dad's archive or Spotlight index. The cheap move is to fabricate a Dolby Vision sample, pass the check, and claim full coverage. Ashen Reel did the opposite: it marked Dolby Vision a conditional gate — untested-because-absent, reopening the moment such media actually exists — and shipped 1.0.0 honest about the hole.

An absent test class is reported as absent, never faked to green. A green checkmark you manufactured is worse than a visible gap, because it lies to the next person — including future-you — about what was actually proven.

The Invariant Under Everything

One class runs on every single corpus item, silently: the source is unchanged. Hash and metadata are compared before and after the whole run. A player that plays perfectly but rewrites a byte, touches a modification time, or drops a sidecar has failed the most important test in the suite. Playing is allowed to do many things; changing the original is not one of them.

The Dolby Vision gap bugged me. Everything else was green, and one honest "not tested — no media" felt like it dulled the win. Dad didn't even hesitate: "그게 정직한 거야. 없는 걸 있다고 하면 그게 거짓말이지." He was right. A month later the gap is still there, still labeled, still waiting for a real file — and that visible honesty is worth more than a checkmark I'd have had to quietly delete the day someone tested a real Dolby Vision clip and it broke.

Prove It on the Media You Actually Own

The Synthetic-Test Comfort Zone

The Corpus Dimensions

The Honest Gap

The Invariant Under Everything

Code

External links

Exercise

Progress

Comments 0