What we tested
Criterion 1 (chain integrity + artefacts): Twelve design-of-record runs were executed on 2026-06-04 on floor resplan-12439-floor-0 with the csi-link-resplan-12439-multiroom link layout. Each run chained walk-notebook (geometry: from_floor) → sionna-csi-runner — the executed design, replacing the parametric jupedsim-runner that could not register against a real floor. Four runs spot-checked (01KT8WCGV8, 01KT8WJSZF, 01KT8X8N9R, 01KT8XGWFR) all report gate_passed: true, exit code 0, and carry links.parquet, csi.hdf5, and trajectory.parquet in S3; additionally the reduction notebook consumed all 12 links.parquet without NaN/Inf in the computed statistics. The 13th run (01KT8VXB6C, N=6, 100 frames) is attached for provenance only and excluded from quantitative claims due to a shorter frame axis.
Criterion 2 (monotonic CV vs crowd size): The reduction notebook (csi_temporal_dynamics.py, window_frames=10, walking window 5%–55%; mean — not median — sliding-window CV pooled over links, because link crossings are brief and the amplitude is piecewise-constant with sparse bursts) computed per-N walking-window CV and dynamic-to-noise over the 12 runs. Per-N means at 2.4 GHz: N=4 → CV 2.4e-6, N=8 → 7.3e-3, N=14 → 3.5e-2. At 5.0 GHz (seed 0): N=4 → 1.4e-6, N=8 → 1.7e-2, N=14 → 2.3e-2. Spearman ρ(N, walking CV) = +1.0 at both bands — the ρ ≥ +0.6 criterion is met decisively. Dynamic-to-noise follows the same order (N=4: ≤1.1; N=8: up to 3.3e5; N=14: up to 5.7e5). Caveats: N=4 sits at the fading floor on all seeds and bands (CV ~1e-6, dyn/noise ≤ 1.4) — with this 5-link layout and ~1.2 s frame sampling the temporal counting feature activates between N=4 and N=8, a sensitivity-floor finding, not a failure. Run 01KT8XGWFR (N=14, 2.4 GHz, seed 2) has elevated seated-window CV (2.4e-2) from run-end truncation of the seek-and-occupy walk; the other 11 runs show clean seated collapse (seated CV ≤ 3.7e-6).
Criterion 3 (static↔temporal gap): Requires a matched-occupancy comparison against the c-csi-crowd-occlusion static placement bundle. No matched static reduction was performed in this session. The direction is qualitatively clear — a static snapshot ensemble cannot produce the walking-window CV of 3.5e-2 observed at N=14 — but the quantitative matched-occupancy bundle comparison is unresolved, pending a dedicated reduction over the static occupancy runs at N ∈ {4, 8, 14} on the same floor. That reduction is the named follow-on for the next session.