Dataข้อมูล

Each wall's rows: left.json · right.json · thing.json · face.json · the math: math.json.

Fields: s picture sha1 · d date · u source · a account or site · h headline or post text · t thing label · ts share of CLIP mass on thing labels · r face match to RFK Jr. · doom · e likes+reposts+quotes (Bluesky) · p perceptual hash.

Method วิธี

Find. The things, not his name: outlets' own sitemaps (every article, its date and lead picture) filtered to slugs naming sugar, soda, dyes, recalls, the FDA, vaccines, Tylenol, fluoride and the rest; Bluesky search, day by day, for the same words, every attached picture and link-card picture; GDELT (a free index of world online news) for articles mentioning him. A first pass searched Bluesky by his name; those rows stay in the walls and out of the thing-story shares.
Face. OpenCV's YuNet finds faces; SFace turns each into 128 numbers; a face counts as his when its mean cosine similarity to his five nearest reference faces, taken from Wikimedia Commons photographs, clears the threshold.
Side. His face left of the picture's centre files it left; right of centre, right.
Thing. CLIP (ViT-B/32, trained on LAION-2B captioned images) scores the part of the picture beyond his face against 180-odd written labels (sugar, soda, food dye, Tylenol, vaccines… and scene labels like podium, crowd, tweet screenshot); a thing counts when thing labels take at least half the mass and the best one clears a cosine of 0.24.
Seam. A straight vertical cut, where one exists, is recorded; it does not decide the class.
Doom. A weighted word list over the headline or post text.

Misses and false matches both happen: a face turned away, a composite with no hard seam, a plate of food CLIP calls a podium. Scores ship with every row so a reader can set a stricter cut.

Window 2023-09-26 to 2026-09-26. Built 2026-09-26T14:40+00:00.