Locating the blockage-dominant regime — through-wall, sparse-LOS CSI

Motivation

csi-static-occlusion found variance (L4.5) dominant across the whole occupancy range in an in-room, multi-link corridor geometry — there was no low-occupancy blockage-dominant regime, because bodies rarely fell in the thin LOS sliver. The csi-sensing chain's two schools only reconcile if both regimes are exhibited; Depatla's blockage-dominant regime needs sparse-LOS / through-wall geometry. This experiment supplies it on a real apartment: a deliberately wall-obstructed link should make mean attenuation the leading feature.

Design

• Simulation: exp-csi-static, resplan-12439-floor-0, occupancy n_agents ∈ {0,1,2,3,4,6,8,12}, n_placements=48, bands {2.4, 5.0} GHz; max_depth ≥ 4 (more reflections matter through walls).
• Link layout (one experiment, two link classes): Tx in bedroom-0; Rx#1 in bedroom-0 (in-room control), Rx#2 in adjacent bedroom-2 (1 wall), Rx#3 in far bedroom-3 (≥2 walls).
• Body model: wired per-body loss (5 dB prior); through-wall ITU attenuation handled by the scene's wall RadioMaterials.

Evaluation

• Per-link-class mean-attenuation slope (dB/person) and crossover N*: expect through-wall links to show a positive crossover (blockage leads) vs the in-room control's N*≈0.
• Baseline through-wall path loss (N=0) per wall count → validates the ITU material chain.
• Contrast with csi-static-occlusion to map the geometry axis of the L4.5↔L4.6 tension.

Driven by c-csi-through-wall-blockage.

Results (sweep 01KT88GCHB6DBV63GB7V9FF18E; multi-agent analysed + adversarially verified)

First pass: 10 of the 16 planned cells (n_agents {0,2,4,8,12} × {2.4, 5.0} GHz, seed 0, 24 placements × 5 links, depth 3 — not the brief's depth 4). Criteria verdicts:

• Plumbing — MET. All runs return the scalar floor + clean links.parquet; the three link classes stage as designed.
• Physical ordering — PARTIALLY MET. The N=0 path-loss ladder is not monotone in wall count at 2.4 GHz (spread only 2.8 dB; far-bed0 sits 0.8 dB above the in-room reference); at 5.0 GHz the in-room→through-wall step is sharp and sensible (9.0–12.7 dB) but within-class ordering stays weak. Worth a look at the wall-material chain / depth before trusting absolute levels at 2.4 GHz.
• Blockage regime — DIRECTIONALLY MET, underpowered. The sign flips between link classes at 2.4 GHz: the in-room control gains +1.64 dB with 0→12 occupants (ρ(n, amp)=+0.90 — constructive multipath; the crowd signal is carried by CV, ρ(n, CV)=+0.80), while the far through-wall links lose signal (+0.55 dB net; +0.76 dB net and strictly monotone at 5.0 GHz). Loss-slope gap far-minus-in-room = +0.19 dB/person at 2.4 GHz; the mid link has the single strongest blockage slope (+0.144 dB/person at 5 GHz, net −1.7 dB). Caveats: net effects are < 1 dB; the far-class loss is driven mostly by bed3; single seed; only 1 of 8 class/band cells strictly monotone in mean.

Interpretation (sober). The geometry-continuum reconciliation of the two occlusion schools is supported in sign: through-wall links convert crowd into mean attenuation where the in-room link converts it into variance (and even constructive gain). It is a directional, not quantitative, result — converting it into a powered claim needs ≥3 seeds, the missing {1,3,6}-person cells, and the depth-4 variant. Analysis: multi-agent workflow csi-campaign-synthesis (3 analysts → adversarial verification → synthesis), 2026-06-04.