Motivation
csi-static-occlusion showed, on one synthetic floor, that the variance signature (CV) rises monotonically with occupancy and that the blockage↔variance crossover is band-dependent. The open thesis question it cannot answer with a single floor: does that relationship transfer across environments? Cross-session / cross-environment drift is the explicit motivation for periodic BLE recalibration (ble-periodic-calibration); Choi/Ling (Wi-CaL) report a large within-session → leave-one-session-out MAE gap. This experiment reproduces that drift in silico across the ResPlan dataset of real apartment layouts — the cheapest possible test of geometry-invariance before any hardware campaign.
Design
- Simulation:
exp-csi-static, identical occupancy protocol per floor (n_agents ∈ {0,2,4,6,8,12},placement=random,n_placements=48, bands {2.4, 5.0} GHz). - Floors: ≥6 ResPlan apartments (e.g.
resplan-12439,-1374,-147440,-7421,-16157,-12419), each with a placed Tx/Rx link (prerequisite). - Reduction: per-floor CV(N) and mean|H|(N); then the cross-floor distribution of the CV slope and the crossover N* — the spread across floors is the drift metric.
Evaluation
- Per-floor monotonicity (Spearman) as the within-environment sanity gate.
- Cross-floor coefficient of variation of the CV-vs-N slope = geometry sensitivity. Large → recalibration required (supports the thesis); small → a robust proxy.
- A leave-one-floor-out toy predictor (fit CV→N on k−1 floors, test on the held-out) to quantify cross-geometry MAE inflation — the in-silico analogue of Wi-CaL's session gap.
Driven by c-csi-cross-geometry-resplan.
Results (6 sweeps / 60 runs, 2026-06-04 — criteria verdicts)
- Plumbing — MET. All six apartments staged (50–84 walls each; uniform Tx-largest-room +
3-Rx policy,
resplan-12439on its 5-Rx multiroom layout) and ran the identical protocol; 60/60 runs clean. - Per-floor monotonicity — MET. ρ(N, CV) ∈ [0.6, 1.0] for every floor × band cell (gate ≥ 0.5).
- Transfer — DRIFT DEMONSTRATED. The CV-vs-N slope varies across apartments with relative spread 0.38 (2.4 GHz) / 0.50 (5.0 GHz), and the leave-one-floor-out predictor's MAE inflates 2.37× at 2.4 GHz (1.27 → 3.01 people) and 3.09× at 5.0 GHz (1.22 → 3.77 people) — both far beyond the 1.5× criterion. A CV→occupancy mapping fit on five apartments mispredicts the sixth by ≈ 3 people.
Interpretation (sober). The occupancy→variance relationship is qualitatively universal
(monotone everywhere) but quantitatively geometry-specific — exactly the structure that
makes per-environment calibration necessary and periodic re-calibration valuable
(ble-periodic-calibration). The in-silico inflation (≈ 2.4× at 2.4 GHz) is of the same
order as Wi-CaL's real-world session gap (0.13–0.18 → 0.35–0.41 MAE, ≈ 2.3–2.7×) — an
encouraging, though metric- and scale-different, correspondence; it should be tested, not
assumed, by exp-csi-calibration against real captures.
Caveats: single seed; 24 placements/level; one floor pools 5 links vs 3 elsewhere; the
predictor is deliberately a per-band linear toy (a real counter would normalise per
environment — the very step whose necessity this measures); uncalibrated 5 dB body-loss
prior. Full analysis: notebooks/experiments/csi-cross-geometry-generalization/01_cross_geometry_analysis.ipynb;
figures in _attachments/experiments/csi-cross-geometry-generalization/.