docs(bench): profile the grain ladder against 86 real titles

[lusoris]

--grain N was an ffmpeg noise=alls= value with no physical meaning and no stated safe range. This profiles it against a real library so a value can be read as "about the grain of X", and marks where the measurement stops being trustworthy.

Method. 86 titles — 56 films + 30 TV series (one episode each) — stratified across the 1960s–2020s, 24 frames each at a fixed seek, measured two ways: a native centre crop=640:360 (true per-pixel grain) and scale=640:360 (what the corpus feeds the filters). Only aggregate statistics recorded; no content redistributed.

What real content measures (native crop, n=84)

percentile	grain_sigma
p10	0.00173
p25	0.00381
median	0.00666
p75	0.01020
p90	0.01352

Median by decade:

decade	film	TV
1960s	0.00823	—
1980s	0.00788	0.00700
1990s	0.00644	0.01045
2000s	0.00408	0.00997
2010s	0.00649	0.00822
2020s	0.00215	0.00275

Grain collapses in the 2020s — 0.0022 film, with median grain_flat of 1.000: the estimator finds the frame essentially entirely flat. That is consistent with the repeated honest negatives on clean content, and it's why the grain axis has to be injected rather than selected.

TV is grainier than film from the 1990s on (1990s TV 0.0105 vs film 0.0064) — film-sourced series keep grain that feature Blu-rays often have remastered out. Thanks for the nudge; folding TV in materially changed the picture.

Downscaling hides grain differences

	median	p25	p75
native crop	0.00666	0.00381	0.01020
scaled to 640x360	0.00698	0.00568	0.00945

Medians barely move but the interquartile range halves (0.0064 → 0.0038): downscaling lifts the floor with resampling residual and lowers the ceiling by averaging real grain away. Benchmarking at the corpus resolution therefore under-detects grain differences — a caveat on any BD-rate delta attributed to grain.

The actionable output

Use netflix-bar as the grain base, stay at or below --grain 16:

--grain	grain_sigma	Δ	grain_flat	trust
0	0.01243	—	0.600	yes
8	0.01380	+0.00137	0.572	yes
12	0.01510	+0.00267	0.535	yes
16	0.01745	+0.00502	0.438	marginal
20	0.01824	+0.00581	0.171	no

Do not use bbb — it starts at only 33% flat coverage and goes non-monotonic past 16: sigma peaks at 0.0244 then falls to 0.0181 by grain 24. A non-monotonic axis is not an axis.

Honest caveat

Both corpus clips read higher than nearly all real content (0.0124–0.0131 vs a 0.0067 median) despite being the clean sources — they're compressed low-resolution distribution encodes, and codec artifacts put high-frequency residual exactly where the estimator looks. So grain_sigma is high-frequency residual in flat areas, with grain one contributor and codec noise another. Fine for deciding whether removable impairment exists; not safe for cross-source absolute comparison unless encodes are comparable.

🤖 Generated with Claude Code