Light-box protocol: T10 secuplug-t10-jxh42, 2026-11-07

The open stack passes 28 of 31 compared checks against the vendor stack; all 3 failed checks are explained or handled (none open).

Key numbers, open vs vendor

Where a direction is defined, open is better in 14, equal (within 3 %) in 7 and worse in 1 rows. This direction chip is the raw open-vs-vendor comparison; pass/fail uses the tolerance of each row.

1. At a glance
2. Summary (full text)
3. Findings and follow-up (3 rows)
4. Deviations and warnings (4)
5. Checks (pass/fail) (28 pass, 3 fail)
6. Protocol (camera, stacks, lamps)
7. A1 P1 defaults (default AE, comp 128)
8. A2 Matched exposure + NV12 noise (P2)
9. A3 Colour (dE00) and grey row
10. A4 Sharpness (slanted edge at the chart edges)
11. Pictures (16)
12. Method

One row per failed or deviating (warn) check and how it is handled. Issue states were read from GitHub at generation time. The check name links to its row in the checks table.

• box thermometer value was stale (watcher not running): ambient temperature not known for this run
• P2 point a (2x): the vendor stack is 30 % darker than at 4x/8x at the same exposure x gain; open is linear over 2x/4x/8x (vendor quirk at the 2x step).
• Night P2n: at identical exposure the vendor is ~9 % brighter (night tone curve); in P1n defaults open is brighter (longer IT), open noise lower. Low priority.
• 3000 K fair pass: the chart was off-centre (top row and right column cut, PTZ re-homes on every reboot); identical framing on both stacks, so the open/vendor comparison holds. The 5000 K colour pass was centred and checked on a snapshot before the start.

Cells: open / vendor value, Δ = open - vendor (relative to vendor), then the direction chip (better / equal / worse, where the row defines one) and the pass/fail verdict with its rule. Rows are grouped by test section.

Same integration time (lines) and gains (x1024) forced on both stacks through the streamer, read back from the ISP (verify). NV12: 8 frames of channel 1, centre half, temporal std per pixel.

chart_de.py (audit/issue230/colour) per phase, cells = open / vendor. Lower dE00 is better.

Suite colour-cct, P1 day

audit/chart/spyder.py: sRGB -> Lab D50, CIEDE2000, reference = Datacolor default values, D50 assumed, not batch specific.

dE00 per patch, P1 day

Grey row A1..A6

Outer edge of the two middle patches on every chart side (patch vs black chart frame), 8 frames averaged, edge ESF with sub-pixel registration from the small chart tilt (lbmetrics.edge_sharpness): 10-90 % rise in pixels (lower = sharper) and MTF50 in cycles/pixel (higher = sharper).

Downscaled to 800 px, JPEG q75. Chart and grey card only. Click a picture to see it large (Esc or click closes).

A1 p1_day

A1 p1_night

A2 p2_a

A2 p2_night

A3 dE00 overlay p1_day (chart_de.py)

A3 dE00 overlay p2_a (chart_de.py)

A4 MTF crops

• Same camera, same chart position, same lamps; open stack first, then the vendor ISP module + vendor libimp are loaded from /tmp (nothing flashed); the camera returns to the flashed image at the end (module reload or reboot).
• P1 = camera defaults (AE, AWB, comp 128). P2 = matched exposure: the same integration time and gains on both stacks, set inside the streamer and read back from the ISP; rows whose read-back differs are marked verify=FAIL and not compared.
• A/A: point p2_a is measured twice (p2_a, p2_a2); the relative difference is the repeatability of the method. A difference between the stacks counts only beyond 2 x that spread (floor 5 %).
• NV12 noise is measured on raw NV12 frames from the streamer (no JPEG in the path): temporal standard deviation per pixel over the burst, mean over Y 32..224, and U/V.
• Colour: SpyderCheckr 24 detected automatically, sRGB -> Lab (D50), CIEDE2000 against the Datacolor default values (D50 assumed, not batch specific); 'grey-normalised' removes the white-balance cast first.
• Sharpness: slanted-edge method on the chart's outer patch edges, frames averaged against noise; JPEG snapshot path.
• SSIM/PSNR (when run): both stacks record the same static chart scene at the same bitrate; each stream is decoded and compared against a high-bitrate capture of the SAME stack (reference), so the number is the encoder loss of that stack; a cross-stack comparison on static frames is listed separately and includes the ISP difference.
• Encoder (A5): both stacks get the same streamer settings (codec, rc_mode, qp / bitrate, gop) through a /tmp conf and a streamer restart; the recordings come over RTSP (ffmpeg -c copy). The reference is a mean image, because single NV12 frames are not frame-aligned with the stream; the scene is static, so every decoded frame can be compared with it. A second reference at the end of the phase shows whether the scene or the AE moved in between.
• Lamp switches are made by a person on a marker (NEED_* file); the times above are the marker times.