{
 "name": "a room never seen: does physics with memory beat memory alone on configurations not yet measured",
 "registered": "2026-10-05",
 "why": "On Motus (results-motus.json, not met) physics added nothing over memory: one room, 830 furniture layouts, so memory alone already covered it. The claim physics can carry is the other regime: a room measured one configuration at a time, predicting each next configuration from the few measured so far. Developed on Arni only (already read; physics + residual beat memory in 8 of 8 orders, at every neighbour count 1 to 20 for memory). Scored here on a room no run has read.",
 "data": {
  "source": "dEchorate (Di Carlo, Tandeitnik, Foy, Bertin, Deleforge, Gannot; Zenodo 5562386, CC BY 4.0): one 5.705 x 5.965 x 2.355 m room whose surfaces are each reflective or covered by absorbing panels, 30 microphones, 7 loudspeakers, file dechorate_rir_processed.hdf5 (md5 02171edaaf2210cc4182a7e5b1f183a9), 48 kHz",
  "configurations": "the ten codes 000000 000001 000010 000100 001000 010000 011000 011100 011110 011111 (digit order floor, ceiling, west, south, east, north; 1 = reflective). 020002 is left out: it adds furniture the physics below does not describe.",
  "items": "an item is one (configuration, loudspeaker, microphone): 2,100 items. Its measured decay time per octave band 125-4000 Hz is T20 (-5 to -25 dB) by params.py (band, decay, fit), because the processed responses are 0.417 s long, too short for T30 in the reflective configurations. A band params.py cannot measure is left out of that item's score; an item with no measured band is did-not-run.",
  "checked before registration": "only the arrival sample of the direct sound for four pairs in 011111, to confirm 48 kHz. No decay was computed."
 },
 "order": "20 orders of the ten configurations, numpy default_rng(20261005).permutation per order. In each order the first two configurations are measured to start; each of the remaining eight is predicted from all configurations measured before it, then measured.",
 "arms": {
  "memory": "the mean of log decay time per band over the 5 nearest measured items (distance: number of differing surface digits, plus 1 if the loudspeaker differs, plus 1 if the microphone differs; ties in stored order)",
  "physics+memory": "per band, 0.161 V / T = c0 + c1 * open area, least squares over all measured items, where open area is the area of the surfaces with digit 0 (floor and ceiling 34.03 m2, west and east 14.05 m2, south and north 13.44 m2) and V = 80.15 m3; with one distinct open area measured, c1 = 0. Then corrected by exp(r * n / (n + 5)), r the mean log ratio (measured / physics) over the same 5 nearest items, n their number."
 },
 "check": "per item, the mean over measured bands of 1 / (1 + (|predicted - measured| / measured) / 0.05), as protocol.json",
 "outcomes": {
  "primary": "physics+memory's mean item score over the eight predicted configurations exceeds memory's in at least 18 of the 20 orders",
  "secondary": "pooled over orders: the share of predicted configurations where physics+memory's mean score is higher; share of items within 10% in every measured band per arm; the anytime betting wealth of order 1 (stake half, bet on the clipped score difference / 0.5)"
 },
 "publish": "results-dechorate.json whatever the outcome"
}