# Opening-frame coding protocol v1

Frozen 8 October 2026 before the new visual coding pass. Descriptive editorial pilot,
not an independent consumer study or evaluation of SaliencyLab prediction accuracy.

Sample: every one of the nine video assets currently linked by the public case-study
views (03,05,06,07,08,09,10,11,12). Convenience sample selected by archive availability,
not random sampling. The archive and earlier model descriptions already exist;
this is not a blind or preregistered experiment. Do not use archived model labels
as the new observations.

Question: do readable brand names and explanatory text occur together in two fixed
opening samples? Inspect extracted video frames at 0.5 and 2.0 seconds at native
resolution. Record explicit readable brand name (including product packaging) and
readable explanatory/offer text beyond a brand name, separately for each sample.
Do not infer familiar brand colors or unlabelled shapes as explicit names. Ambiguous
or illegible is unknown. These two instants do not establish continuous first-three-
second coverage, a first-appearance timestamp, sound-off comprehension or recall.

Capture source URL, SHA-256 of the sampled video, duration, times, coding values and
short visible evidence. AI-assisted visual coding by Codex, no independent human
rater. No inference about audience response or campaign outcomes. Report all assets,
failed retrievals and ambiguous observations. Aggregate only completed observations.
Publish derived observations and source links; do not relicense third-party media.

Reproduction
1. Read each sourceUrl and videoSha256 in opening-samples-v1.json.
2. Download the public source video and compare its SHA-256 before sampling.
3. For each timeSeconds, run:
   ffmpeg -ss <seconds> -i <local-video.mp4> -frames:v 1 frame.png
4. Inspect the frame at native resolution using the coding rules above.
5. Keep disagreements and unknowns; do not overwrite the original dataset.

The archived frame hashes identify our PNG outputs; differing FFmpeg builds can
produce different PNG bytes even when decoded pixels are equivalent. Video hashes
identify the exact source bytes. A changed source requires a separate observation.

Coding clarification recorded with the results: brand/model names alone are not
explanatory copy; an explicit readable duration claim on packaging does count.
No inference is made from a familiar color, silhouette, car grille or swoosh alone.