[shashwat-deep]

Next step of the evidence half (docs/plan.md, "Your half", item 4): the measured-light breakdown, "the darkest tenth of faces by mean luma inside the box, against the rest" (protocol part B). Stacked on #13, which is stacked on #12, #10, #9 and #2; merge those first.

What is in it

• eval/face_luma.py measures each face's mean luma and writes eval/manifest/luma.csv, one row per face of faces.csv. Each mean is taken:

  • on the image as Chromium decodes it, through transforms.upright() from Eval: rotation and darkening transforms in the runner #12;
  • with Y′ = 0.299 R′ + 0.587 G′ + 0.114 B′ (BT.601, the luma JPEG stores), on 0 to 255;
  • over the pixels whose centres lie inside the face's box.

• eval/score.py reads luma.csv when it exists and adds missRateBy.luma: the darkest tenth (the ⌈n/10⌉ faces with the lowest mean luma among the n faces scored, ties broken by image and face ID) against the rest. Both come with bootstrap intervals, plus the cutoff luma. The per-face CSV gains luma and darkestTenth. A luma.csv that doesn't match faces.csv stops the scoring, with a message to rerun face_luma.py.

• eval/manifest/luma.csv for the current draw (1,797 faces, 12 s to make).

• 5 tests:

  • the pixel-centre rule;
  • BT.601 on a known patch;
  • a turned image;
  • the breakdown on 20 made-up faces where the two darkest are missed;
  • that the committed luma.csv matches faces.csv.

  77 eval tests pass in total.

On the current draw (before the review)

Among the 1,298 report-split faces that aren't depictions, the darkest tenth is 130 faces with mean luma at or below 56.2. The median face is at 109.5 and the darkest at 8.1. The review will change the draw a little, so rerun face_luma.py after it (the test catches a mismatch).

For both of us to agree

Part B of the protocol now defines luma (BT.601 on the decoded image), "inside the box" (pixel centres), and the darkest tenth (⌈n/10⌉ of the faces scored, ties by ID).

🤖 Generated with Claude Code