c-csi-topology-sniffer-count — session 01KVZRDA8V0RAN269E4S0JAQFV (powered link-as-unit re-reduction, 2026-06-25)
This session re-reduces the same 12 runs as 01KVXT9B31TF088GH5TE539QZN with the unit of analysis switched from the floor (N=6, underpowered, area∼diameter confounded) to the link (N=40), executing the follow-on design the prior session called for.
Verdict: the wall-count-on-path operationalisation of layout-topology-fingerprint-discriminability is REFUTED in silico; floor area is the operative predictor. Strength demoted plausible → speculative.
Method
Reduction csi_topology_linkunit.py (monad-free, --selftest passes). Per-link topology metric = wall-count-on-path, the number of wall segments a link's Tx→Rx line-of-sight crosses, computed by 2D segment-intersection from each run's own inputs/scene.json (Tx/Rx coordinates + wall list) — no PostGIS, fully self-contained per run. Per-link discriminability = std (dB) of mean_amp_db across the 80 placement frames (how much a single moving body's position swings that link's amplitude). −240 dB "no path" sentinel clipped to −120 dB first. 3 Rx × 6 floors × 2 seeds = 40 link-observations.
Result
| test | value | reading |
|---|---|---|
| wall-count-on-path → D (pooled ρ) | +0.14 | no relation |
| wall-count-on-path → D (within-floor partial ρ) | +0.14 | null even after removing the area-confounded floor mean |
| floor (bbox) area → D (pooled ρ) | −0.54 | area is the stronger predictor; topology does not beat area |
| LOS-vs-NLOS defeater (wall-count 0 vs ≥1) | d=−0.13, p≈0.88 (n=4/36) | REFUTED |
| BLE/CSI per-link discriminability ratio | 2.49× | BLE more separable (reaffirms c-ble-csi-coregistration) |
Combined with the prior floor-level finding (required sniffer fraction k* best predicted by area, ρ=−0.74), the in-silico evidence now favours the area + AP-count baseline (adeel2019_ada0) over the topology hypothesis.
Honest bounds
- Limited wall-count spread: most Tx–Rx pairs cross exactly one wall in this corpus, so the area-dominance + within-floor null carry the verdict more than the wall-count axis on its own.
- One operationalisation: this tests wall-count-on-path + per-link amplitude std, not the full graph-metric set (the prior weak floor-level mean_degree ρ=+0.37 survives, unpowered).
- Still in-silico: single-body, uniform-material; the multipath-richness confound (c-csi-fidelity-material) and real-layout transfer are untouched. The literature analogy (
jung2025_3f2ebetweenness contamination,depatla2018_97a9shape>area,zhen2022_bb0badaptive placement) is what keeps the hypothesis fromrefutedrather thanspeculative; the deciding evidence is a small real-layout fingerprint study, folded into the IP-106 testbed.
A cheap, valuable negative: it simplifies the sniffer-count deployment story to area + AP count, and removes a speculative arm from the system-design chain. Recorded on layout-topology-fingerprint-discriminability. Reduction monad_knowledge/notebooks/python/csi_topology_linkunit.py; fig under the session prefix + _attachments/topology-linkunit/out/.