Pixels per artwork

Authoritative source, reproduced unchanged. Exact Markdown · Agent copy with resolved links · The files: data/pixels-per-artwork/. These instructions govern research runs, not checkpoint interviews. They keep the run's own words: a room is a community, a mark an artwork, the analyst the aggregator, the case lead the community coder and the global model the population model.

Every placement routed to the atlas artwork or artworks whose outline contains it, and a 30-minute panel per artwork. Built from two public sources, the pixel record and the atlas, by two commands. The study's copy is Layer C/C5_atlas_flow/ (its build kit, build/, is what this folder holds).

Build it (from data/; needs numpy and Pillow, pandas optional; about 1.5 GB free while it works):

1. python3 pixels-per-artwork/code/fetch_sources.py: downloads the record into pixels/place_tiles_sha1.csv if it is not there (894 MB, from archive.org), uses the atlas in atlas/atlas_2017.json, and writes pixels-per-artwork/inputs/. About 2 to 5 minutes.
2. bash pixels-per-artwork/run_all.sh: the five stages, into pixels-per-artwork/data/. About 2 minutes, 1.1 GB.

fetch_sources.py stops unless every count matches: 16,559,897 rows, 1,166,924 placers, 16,556,641 on the canvas, 1,473 artworks (1,236 naming a subreddit), 956,741 covered pixels, 1,319,701 pixel-artwork memberships, and the final canvas's fingerprint.

Files after it:

Read by: aggregator/handler/build_canvas.py (its c5_data path points here), which reads artworks_index.csv and the panel's placements.

Three things to know.

• Overlap. Outlines overlap on 30.5 per cent of covered pixels, so a placement appears once per containing artwork: 23,058,857 rows for 16,556,641 placements. is_primary == 1 keeps one row per placement, in the smallest containing outline (16,136,055 rows).
• Placer ids. placer_uid indexes inputs/placer_hashes.txt.gz, which fetch_sources.py writes. It numbers placers as the study did, by a 64-bit key of each hash, so ids equal the study's (all 1,166,924 checked). Join across builds on the hash all the same.
• Ties. Timestamps are whole seconds, so most rows tie with another, and the two releases of the record order ties differently. fetch_sources.py sorts the rows as the study's copy was sorted (as text: time, hashed name, the rest of the line), and build_artwork_flow.py breaks every remaining tie by colour and placer. Repaints, reverts and restorations depend on this order. Under this rule, on the primary partition, they are 4,370,182, 12,186,459 and 5,695,139; the study's C4 notes quote 4,369,968, 12,186,673 and 5,695,494 from before the rule. Placements, placers and pixels touched do not depend on order.

Tested 2026-09-26: from nothing, in a Linux workspace: the download, the parse and the five stages in about 4 minutes. Every array of pixel_record.npz equals the study's stored parse. The outputs equal C5_atlas_flow/data/ as rebuilt that day: artworks_index.csv byte for byte; the long table, the tidy panel, the placer table and the per-artwork files checked, on decompressed content.

Codebook

Flow rows: artwork_id (0 to 1472, the atlas's order); ts_ms; x, y; color; placer_uid (index into inputs/placer_hashes.txt.gz); n_artworks (outlines containing the pixel); is_primary (1 in the smallest of them).

artworks_index.csv: per artwork, its name, subreddits, area_px, center, placements, placements_primary, distinct_placers, first_ts_ms, last_ts_ms. placers.csv.gz: placer_uid and the hash. panel_30min.csv.gz: per artwork and 30-minute bin, placements, placers, pixels_touched, repaints, reverts, restorations, net_change_px, each also in a _primary version, with bin, bin_start_ms, hours_from_open, area_px and placers_total.