Locating the blockage-dominant regime: through-wall, sparse-LOS CSI
Current assessment (2026-09-15)
The tested simulation reversed the predicted regime: in-room links showed blockage dominance and one-wall links showed variance dominance. No multi-wall class was produced.
The 52-run, three-seed comparison supports a conditional inversion, not a general conclusion about through-wall sensing. The simulated drywall palette produced a small 2.4 GHz loss, so the material model affects the interpretation. The absent multi-wall class and missing rendered figures remain evidence gaps.
Next evidence needed: Retain the negative and the missing class. Check material and receiver assumptions before extending the crossover to real links, and report uncertainty at the placement/seed unit.
Basis: Thesis Narrative, the September 10 synthesis, and the September 15 register review. See the review record for evidence sources and outstanding checks.
Earlier protocol and dated results
The methods and observations below retain their original scope. Read their readiness statements and proposed next steps as historical. The current assessment above records what has changed; frozen thresholds have not been retrospectively revised.
The problem, in plain words
Picture a hallway with a lamp at one end and your eye at the other. If one person walks through the beam, you see a clear shadow — that is blockage. Now imagine the lamp is in the next room, behind a frosted-glass wall: the light that reaches you is already scattered and soft, so a person walking around barely changes what you see. That is the difference between a line-of-sight link and a through-wall link, and it turns out to matter enormously for how a crowd imprints itself on a Wi-Fi signal.
The field has argued about this for a decade. Depatla and Mostofi showed you can count people through walls with Wi-Fi (depatla2018 ↗), and robust ambient-signal systems followed (guo2022 ↗). Other work counts people from the statistical richness the crowd adds to the channel (zhou2020 ↗). Both are right — in different geometries. The thesis needs both regimes exhibited on the same platform so the two schools reconcile as one continuum rather than two contradictions. The in-room study gave us the variance end. This experiment was the search for the blockage end.
What we are trying to prove
- Hypothesis (falsifiable): on a through-wall link (Tx and Rx in different rooms, ≥1 wall between), mean attenuation becomes the leading occupancy feature — the blockage↔variance crossover moves to N* > 0 and the attenuation slope is steeper than the in-room control's. If instead the through-wall link is variance-dominant from N ≈ 0 (crossover at zero), the hypothesis is refuted.
- What a null means: a null is not a wasted run. If blockage never leads even behind a wall, then the L4.5 ↔ L4.6 axis is not an in-room-vs-through-wall axis at all — it is something else (as it turned out: a strong-LOS-vs-weak-LOS axis). That reframing is itself a chapter result, and it bounds where Depatla-style through-wall counting can be expected to work.
- Supporting physics check: the N = 0 baseline path loss should increase with the number of walls between Tx and Rx, validating the simulator's ITU wall-material chain. If it does not, the finding is about the simulator, not the physics.
How the experiment works (plain method)
- Scene. A real apartment floor plan (
resplan-12439-floor-0) is loaded into the ray-tracing engine (hoydis2023 ↗). Each wall is resolved to an ITU-R P.2040 radio material, so through-wall attenuation is physical, not a fudge factor. - Links. A transmitter sits in
bedroom-0. Receivers are placed to give graded wall counts: one in the same room (in-room control), and others in adjacent rooms (meant to be 1-wall and ≥2-wall). - Occupancy sweep. Bodies (0, 1, 2, 3, 4, 6, 8, 12 of them, each a lossy absorber) are dropped at 24 random placements per cell, at two Wi-Fi bands (2.4 and 5.0 GHz), repeated over three random seeds. That is the 48-cell main grid, plus a 4-cell diagnostic arm that re-runs the empty scene at deeper ray-reflection settings (depth 4 and 6) to check the baseline is not a truncation artefact — 52 runs in all.
- Reduction.
csi_through_wall.pycounts the walls each link actually crosses geometrically (segment-crossing count — ground truth, not an assumption), builds the N = 0 path-loss ladder per band and depth, and computes each link class's attenuation and variance (CV) slope vs occupancy plus the blockage↔variance crossover, reported as mean ± spread across the three seeds.
What we've found so far (honest)
The campaign session (01KT9BJ5172M9HHXXBE67DYS11, 2026-06-04, 52 runs, all attached and gate-passed) resolved 1 of 3 success criteria. The supervisor's own verdict: "Hypothesis inverted, not confirmed."
| Criterion | Verdict | What actually happened |
|---|---|---|
Layout stages 3 link classes; clean links.parquet, no NaN/Inf |
FAILED (layout finding) | Geometric wall counts came out {in-room: 0 walls; all four other links: 1 wall each}. No ≥2-wall link exists — the far-bedroom path slips through a doorway-gap alignment. So every class comparison below is in-room (1 link) vs 1-wall (4 links), not the intended three-way ladder. |
| N=0 path loss increases with wall count (ITU chain physical) | MET | The empty-scene path-loss ladder is monotone at all depths {3, 4, 6} on both bands — the earlier (single-seed) non-monotone 2.4 GHz ladder did not reproduce. The 1-wall step is 11.0 dB at 5 GHz but only ~0.5 dB at 2.4 GHz (an interior drywall is nearly transparent at 2.4 GHz). This alone explains the old "<1 dB through-wall" result as seed-robust physics, not noise. |
| Through-wall link is blockage-dominant (crossover N*>0, steeper mean slope than in-room) | FAILED, and inverted | Through-wall links: crossover N* ≈ 0 on every seed and band (variance-dominant from the first occupant); attenuation slope only −0.050 ± 0.019 dB/person at 2.4 GHz, −0.064 ± 0.024 at 5 GHz. The in-room control is where blockage dominates at low occupancy: crossover N* ≈ 2.0–6.8 at 2.4 GHz (noisy) and ≈ 1.3–3.6 at 5 GHz. |
Interpretation (sober). The geometry-continuum idea survives, but the axis is relabelled. It is not "in-room = variance, through-wall = blockage." It is "strong-LOS = blockage-first, weak-LOS = variance-only." Blockage needs a line-of-sight strong enough to shadow; that lives on the in-room link and is already absent behind a wall, where the field is diffuse. Depatla-style through-wall counting presupposes a strong, short-range through-wall LOS — a regime these drywall-apartment links at 2.4 GHz (0.5 dB/wall) simply do not produce. That is a real, publishable boundary on the method, not merely a failed prediction.
Caveats that bound all of the above: one floor, one layout, only two link classes (the ≥2-wall class was never staged), a uniform-drywall material palette that a sibling campaign (c-csi-fidelity-material) has just shown is load-bearing — at 11 dB/wall the inversion may not survive heavier (concrete) walls — ray depth ≤ 6, and 24 placements per cell. Converting the inversion into a quantitative claim needs: (1) a re-staged layout with a genuine ≥2-wall link, (2) a concrete-palette variant, and (3) a short-range strong-LOS through-wall link for a faithful Depatla recreation. No figures were shipped with this session (the requested csi_through_wall panels were not rendered), so there is nothing to read visually yet — the numbers above are the whole result.
Key references
- depatla2018 ↗ — the through-wall blockage school this experiment set out to reproduce (and inverted).
- guo2022 ↗ — a working through-wall counter; leans on ambient/statistical signal, consistent with our "no clean shadow behind a wall" finding.
- zhou2020 ↗ — the variance/statistical-richness counting school (the regime that won).
- hoydis2023 ↗ — the differentiable ray tracer that produced the CSI.
- zhang2024_df4a ↗ — site-specific 3D material segmentation; the fidelity bar the uniform-drywall palette misses.
- zhu2024 ↗ — ray-launching solvers disagree; grounds the red-team's fidelity caution.
- rampa2022 ↗ — physical EM multi-body blockage models; the mathematician's identifiability/CRB reference.
- santos2024 ↗ — environment/link identity as a variance driver; the statistician's exchangeability caution.
- guarino2026 ↗ — the reproducibility bar the SWE voice invokes.
- huang2025 ↗ — a public real-CSI anchor for the red-team's cross-check demand.
- wang2026 ↗ — the field-wide generalisation crisis the sim result must eventually clear.
- ling2024 ↗, zhevnenko2026 ↗ — commodity single-link deployment baselines behind the practitioner's siting rule.
- shahbazian2023 ↗ — the non-expert on-ramp for the science-communicator voice.