c-csi-crowd-temporal — session 01KVWYPHMHZVGRYQBW5Z9ZYEQ7 (re-confirm + extend, 2026-06-24)
Verdict: Claim 1 ("WiFi feels the crowd") confirmed and strengthened. A coupled JuPedSim→Sionna-RT sweep of 30 runs — N∈{4,6,8,11,14} × band∈{2.4,5.0 GHz} × seed∈{0,1,2} — shows the motion-driven temporal CSI variance feature scales monotonically with occupancy at both carriers and across all three seeds. This extends the sealed seed-0 6-run baseline (01KT99FX2Z…) to a finer N-ladder and seed-replication, and the ρ-stability holds.
Criterion 1 — coupled chain + valid artefacts — ✅
All 30 runs chained walk-notebook (geometry: from_floor) → sionna-csi-runner on the same PostGIS floor resplan-12439-floor-0 / experiment csi-link-resplan-12439-multiroom, each emitting a frame-indexed links.parquet, csi.hdf5 [frame,link,…], and the carried-through trajectory.parquet. Every run gate-passed (exit 0); the reduction read all 30 with no NaN/Inf. (One non-blocking walkable-QC flag — 0.51 m² of sub-0.40 m two-abreast channel near (9.1, 16.2) — is a geometry note, not a run failure; it is the mechanism behind the high-N seated-window caveat below.)
Criterion 2 — temporal CV rises monotonically with N — ✅ (strongly)
Walking-window temporal CV (sliding-window σ/μ of per-frame amplitude, linear domain) vs occupancy, mean over seeds:
| N | walking CV @ 2.4 GHz | walking CV @ 5.0 GHz |
|---|---|---|
| 4 | 5.2e-5 | 2.6e-5 |
| 6 | 3.2e-3 | 3.9e-3 |
| 8 | 5.0e-3 | 5.9e-3 |
| 11 | 2.6e-2 | 2.7e-2 |
| 14 | 4.4e-2 | 4.2e-2 |
Strictly monotone → Spearman ρ(N, mean walking CV) = +1.00 at 2.4 GHz and +1.00 at 5.0 GHz (criterion ≥ +0.6). The two bands track each other closely (the variance feature is geometry/motion-driven, not strongly carrier-dependent here).
Criterion 3 — static↔temporal gap quantified — ✅ (with an honest caveat)
The seated window (trailing 20% of the frame axis, the matched-occupancy static proxy) collapses to the fading floor (~1e-6 CV) in the majority of runs while the walking window scales with N — a within-run walking/seated separation of up to ~5 orders of magnitude (dynamic_to_noise reaches 1e4–1e5 at N≥11). This is exactly the feature a frozen-placement static sweep cannot expose, and reproduces the sealed session's seated-collapse.
Caveat (reported, not hidden): seated_cv_max = 0.0336 and dynamic_to_noise_min = 0.77 are NOT failures. (a) The two sub-1 ratios are all low-N runs (N=4/8 specific seeds) where both windows sit at the ~1e-6 noise floor — a ratio of two near-zeros, not residual walking motion. (b) Three high-N runs (N8/s0, N11/s0, N14/s2) retain seated CV ~0.014–0.034 because congestion (the two-abreast jam channel) delays full parking, so the trailing window is not yet stationary. The seated collapse is therefore conditional on all agents having parked within the analysis window — a finding that reinforces, rather than weakens, the motivation for Claim 2: in a genuinely seated/static crowd the CSI temporal feature starves, and a BLE anchor must carry occupancy there.
Tie-back
H4 occupancy-csi-variance stays supported (clean re-run, no strength change). The result establishes the WiFi half of the dial-and-label mechanism: the in-room variance feature is real and motion-driven. Reduction artefacts (csi_temporal_dynamics.metrics.json, temporal_stats.csv, fig_csi_temporal.png) under the reduction run 01KVX0GN40QHH21EHZ9042BTXJ. 30 source runs attached.