{
 "name": "four rooms never seen: does the learned field beat the calibrated formula with memory at positions not measured",
 "registered": "2026-10-10",
 "why": "The method that met twice (results-extrapolate-arni.json, results-dechorate.json) is a calibrated formula corrected by the mean residual of the nearest measured items. Here a room is measured at 12 positions and every other position is predicted. In development on new real rooms (BUT ReverbDB, leave one room out, 12 of 31 positions measured) the learned field scored 0.481 against 0.463 for formula+memory and was higher in 8 of 8 rooms (dev-but-field.json). On dEchorate, whose microphones sit in clusters of five, it was level (0.348 against 0.350, dev-field.json). A network trained on simulated rooms to correct the formula lowered the score on new real rooms (0.452, and 0.428-0.454 trained on what the field leaves: dev-but-stack.json) and is not part of the candidate.",
 "code": {
  "engine commit": "29f5eca",
  "runner": "experiment/acoustics-model/run-haa.py, sha256 052fa375964dd31cc41217f8bb960c39311d8052f79b8d65da9311717592c51b",
  "checked": "smoke-haa.py ran the runner end to end on made-up responses in two of the rooms' geometry before registration"
 },
 "model": {
  "sha256": {"kernel.json": "d8e6d7874688e9f0354fb69351c4728756bf5bb1004a318ad8da8c81f9975e98"},
  "what": "experiment/acoustics-model/model/kernel.json: per parameter, the field's length (m), signal-to-noise ratio and log-distance weight, learned by train_field.py on all eight BUT rooms and every third dEchorate task (t: 0.5, 3, 0; edt: 1, 3, 1.5; c50: 1, 3, 1.5)"
 },
 "data": {
  "source": "Hearing Anything Anywhere (Wang et al., CVPR 2024; Zenodo 11195833, CC BY 4.0): the four base configurations classroomBase, dampenedBase, hallwayBase, complexBase; monaural responses RIRs.npy at 48 kHz and positions xyzs.npy",
  "scenes": "haa-scenes.json (sha256 0edc75ac5731dd25df57eb50a189f2633b7273f4f8e51581fb495ac9af8d856c), made by haa_geometry.py from the authors' room files (github.com/maswang32/hearinganythinganywhere, commit 9fb146487a82b9cbcf6334320ff5794e688d4a03, no licence file): surfaces, a material class from each surface's name, the loudspeaker position, the authors' 12 measured positions (train indices) and test positions (the complement of train and validation)",
  "read before registration": "the room files above; positions xyzs.npy of the classroom, dampened and complex rooms, for the distance from each test position to its nearest measured one (median 0.96, 0.74, 1.26 m; BUT 1.1 m). No response was opened.",
  "items": "an item is one test position of one room: 463 + 198 + 423 + 198. Its T (T20), EDT and C50 per octave band 125 Hz-4 kHz by common.analyse. A band that cannot be read is left out of the item's score; an item with nothing readable is did-not-run."
 },
 "measured": "the authors' 12 train positions per room; every arm sees only these",
 "check": "per item, the mean over the three parameters of the mean over readable bands of 1 / (1 + error / JND), JND 5% for T and EDT, 1 dB for C50 (ISO 3382-1 annex A)",
 "arms": {
  "memory": "the mean of the 5 nearest measured positions (log for decay times)",
  "formula+memory": "the method that met: per band T = geometric mean of the measured T; C50 and EDT by the revised theory (Barron and Lee 1988) with its reflected-energy scale fitted to the measured C50; corrected by the mean residual of the 5 nearest measured positions, shrunk by n / (n + 5)",
  "ism+memory": "image sources to third order with the absorption calibrated to the measured T, plus an Eyring tail; the same memory (no learning)",
  "formula+field (candidate)": "the same calibrated formula; its residuals at the 12 measured positions carried to each test position by the learned field (fewshot.gp_predict with kernel.json)"
 },
 "baseline": "formula+memory",
 "seed": 20261010,
 "bootstrap": 2000,
 "roomsNeeded": 3,
 "outcomes": {
  "primary": "formula+field against formula+memory, paired by item over the four rooms: one-sided sign test p < 0.05 AND the 95% bootstrap interval of the mean score difference (2,000 resamples of items, seed 20261010) above 0 AND a higher mean in at least 3 of the 4 rooms",
  "also": "every arm per room and pooled; the published few-shot numbers on these rooms (FLAC, DiffRIR) are not comparable to this score and are not compared"
 },
 "publish": "results-acoustics-model.json whatever the outcome"
}