{
  "name": "a second room: does physics with memory beat memory alone where memory starts empty, and does a gated self keep improving there",
  "registered": "2026-10-04",
  "claim": "docs/LIVES.md, the curve as revised: within a room as good as memory; better only under transfer to a room its memory has never seen",
  "data": {
    "source": "Motus (Götz, Schlecht, Pulkki; Zenodo 4923187, CC BY 4.0): one 4.9 x 4.4 x 2.9 m room in 830 furniture configurations, 4 loudspeaker positions, one Eigenmike; the omnidirectional channel",
    "items": "an item is one (configuration, source): its measured T30 per octave band 125-4000 Hz (params.py, ISO 3382-1, read from the archive in place) and its documented geometry and absorption (motus_room.py: volume, surface, equivalent absorption area per band from the mesh and the per-surface coefficients Motus states as assumed)",
    "order": "configurations drawn by the seed without replacement, each configuration's four sources together; nothing about the order depends on a measurement",
    "seed": 20261006
  },
  "check": "as protocol.json: the measurement, scored per item as the mean over measured bands of 1 / (1 + (|predicted - measured| / measured) / 0.05)",
  "arms": {
    "physics": "Sabine T30 = 0.161 V / A_b from the configuration's documented absorption, fixed, no learning and no memory",
    "memory": "starts empty in this room: each prediction is the mean measured T30 per band of the 10 nearest configurations already checked (distance: Euclidean on log equivalent absorption area per band, plus 1 if the source differs); with no neighbour yet, the mean of everything checked; with nothing checked, 0.5 s",
    "physics+memory": "the physics prediction corrected by the mean log ratio (measured / physics) over the same 10 nearest checked items, per band (with none, the physics prediction)",
    "gated": "a self whose procedure starts as physics+memory and whose practice is a per-band scale on the documented absorption (one factor per band, start 1.0) and the formula (sabine or eyring); it may change only through the engine's graded gate (alpha 0.05, spend epoch, trialsNeeded 20, budget 300, 4 racing); proposals each night: a CMA-ES fit of the six factors and the formula to its last 14 days of checked items (their derivedFrom) and three random edits (each factor times a log-uniform draw in [1/1.3, 1.3] with probability 0.3, formula switched with probability 0.1); the memory correction applies to every version, as on every arm",
    "book": "the same start procedure and memory, practice locked"
  },
  "time": { "itemsPerDay": 20, "days": 42 },
  "outcomes": {
    "primary": "physics+memory against memory alone, paired by item in the order checked, as an anytime-valid test (the engine's graded betting on score difference): the claim 'physics with memory beats memory alone in a new room' is supported if its wealth reaches 20 (alpha 0.05) within the first 200 items",
    "secondary": "the curve in the second room: gated weekly mean score rising (Mann-Kendall on six weeks, one-sided p < 0.05) while book's does not, and gated beating book in week 6 (one-sided sign test on paired items, p < 0.05)",
    "also": "per arm weekly mean score and share within two JND; the physics arm's per-band bias (median log ratio measured / predicted); adoptions and their diffs",
    "decision": "primary and secondary are reported separately; each is met or not as stated, whatever happens"
  }
}