{
 "name": "Crowds against real shoppers: six windows of one day in the ATC shopping centre, Osaka",
 "registered": "2026-10-10",
 "why": "The ETH/UCY test (protocol-peds.json, 9a1ca4c) missed and stays a miss (3 of 6 scenes, 5 needed). This is a new study on street-like crowds no run here has opened, with the ETH/UCY bar: all windows but one.",
 "data": "ATC pedestrian tracking dataset (Brščić, Kanda, Ikeda, Miyashita 2013; free for research use, citing Zanlungo, Brščić, Kanda 2015, Phys. Rev. E 91:062810), day 2012-11-14 for development (atc-20121114.7z, sha256 6036e361...) and day 2012-11-18 for the test: the first day after the development day in list_of_days.txt without a tracking-problem mark. The test day is extracted from atc-tracking-part1.7z (downloading to [host name redacted] holdout2 at registration; its sha256 goes into the run notes before extraction). Nothing from 2012-11-18 has been opened.",
 "conversion": "experiment/crowds/atc_convert.py at this commit: windows of 180 s starting 10:00, 11:30, 13:00, 14:30, 16:00 and 17:30 Japan time; one sample per 0.4 s tick (nearest sample within 0.2 s), people with fewer than 3 ticks dropped; walkable area by the Jülich rule from the window's own positions.",
 "crowds": {
  "jupedsim (candidate)": "experiment/crowds/crowdsim.py at this commit, JuPedSim 1.4.2, arguments: --model avm --radius 0.12 --radius-sparse 0.22 --rho0 0.5 --time-gap 0.45 --anticipation 0.75 --reaction 0.3 --strength 8 --range 0.2 --look 1.0 --vmax 1.5 --grace 10 --stand 0 --wall-slack 0.2 --dt 0.02 --seed 1 (holdout.sh). The anticipation velocity model; each person enters at their real first position and time and steers toward where their own timetable puts them one second ahead (the shortest walkable path from their real first to last position, walked at constant speed over their real time span); personal space 0.22 m radius where nobody is near, shrinking to 0.12 m as local density rises. Inputs per person are those of the straight walkers.",
  "socialforce": "livingworlds/render 398957f pedreplay.py (copied byte for byte as experiment/crowds/pedreplay_398957f.py, sha256 9db67325...) via opp.py: use_groups, spring, density_aware and walls on, walkable = the run's WKT, the configuration of protocol-peds-juelich.json; re-running it on [host name redacted] reproduces the recorded Jülich outputs exactly (checked on four runs, maximum difference 0)",
  "orca": "the same set-up with ORCA",
  "straight": "experiment/peds/peds_straight.py unchanged"
 },
 "measures": "experiment/peds/peds_score.py and peds_realism.py unchanged (sha256 d4b6fe32..., 21426a04...): rank sum over speed, nearest distance, straightness and contact against the real run; the jupedsim slot holds the candidate, which is a JuPedSim model",
 "primary": "the candidate's rank sum is lower than both socialforce's and straight's in at least 5 of the 6 windows (peds_score.py's own need for six scenes)",
 "failures": "a window in which any of the four crowds produces no output counts as not won by the candidate; nothing is re-run",
 "secondary": "candidate against orca; each measure",
 "development": "ATC 2012-11-14: 19 of 19 windows won (the 19:30 window has no social force crowd: pedreplay.py raised an error there); also the development of protocol-crowds-juelich.json (same crowd and arguments). dev-results.json.",
 "order": "after this commit and the download: experiment/crowds/holdout.sh atc on [host name redacted] as one capped unit, then results-crowds-atc.json is committed whatever it says",
 "publish": "results-crowds-atc.json whatever the outcome",
 "limits": "one indoor public space on one day, tracked by range sensors (body centres, not heads); shoppers wander, so first and last positions say less about the path than in a corridor"
}