c-csi-topology-sniffer-count — session 01KVXQTTMGCKCGFVB3TN2G8WXD (2026-06-24)
Framing discipline (per the brief): in-silico, hypothesis-generating on a self-authored ray-traced propagation prior (uniform-material). N is small (6 ResPlan floors, 2 seeds), and the multipath-richness confound is not isolated. The defeater is a real possible outcome; the hypothesis strength moves only on it.
Verdict: the machinery works end-to-end; the science is mixed and underpowered — layout-topology-fingerprint-discriminability stays plausible (neither upgraded nor refuted). This is the first in-silico instantiation of the hypothesis's own settling experiment, run locally as a deterministic supervisor (12 exp-csi-static runs, no LLM fan-out).
What we tested
Six ResPlan apartments carrying a placed csi-link-* CSI experiment (12419, 16157, 12439-multiroom, 147440, 7421, 1374), single body (n_agents=1) swept over 80 random placements × seeds {0,1} at 2.4 GHz with ble.enabled — 12 runs, all gate_passed=True, every run carrying links.parquet + co-registered ble_links.parquet from one solved CFR. Each placement frame is a location; the standardized frame×link amplitude matrix is the location-fingerprint set (the runner records no body XY, so this — not per-room binning — is the artefact-supported fingerprint). NRG room-adjacency metrics (diameter 2–5, mean betweenness, mean degree) were computed from PostGIS states+transitions and injected via the reduction config; floor areas span 135–584 m².
What we found (C1–C4 ✓ as deliverables)
- Per-link discriminability D weakly favours topology over area. Spearman ρ(predictor, D): mean_degree +0.37, betweenness +0.26, diameter +0.19, area +0.14. Topology metrics edge out area, but every correlation is weak at n=6 (none significant). D is normalised to per-link RMS so the one 5-link floor (12439) is comparable to the 3-link floors — the "hold receiver count fixed" control, applied at the metric level.
- The operational payoff — required sniffer fraction — is best predicted by AREA, not topology. ρ(predictor, k*/n_links): area −0.74, diameter −0.63, betweenness +0.52, mean_degree +0.09. Bigger floors need a proportionally smaller fraction of their links — the area / AP-count baseline (
adeel2019_ada0) holds here, against the topology claim's operational form. - Defeater not triggered, but underpowered. Rank-median diameter split gives Cohen's d(hi−lo)=−1.44, p≈0.078 — not refuted (|d|>0.2), yet the sign is negative (high-diameter floors slightly less separable) and the test rests on 3 floors per stratum with three of six tied at diameter 3. Area and diameter are themselves confounded in this sample (the largest floor, 1374, is also the highest-diameter), so neither the positive D-trend nor the area-dominated k* can be cleanly attributed.
- Secondary finding (BLE thread): the co-registered BLE location-fingerprints are ~2× more separable than CSI's on every floor (D_BLE 1.06–1.32 vs D_CSI 0.56–0.76). In the single-body static regime, the narrowband BLE anchor carries more clean spatial structure than the CSI amplitude feature — consistent with c-ble-csi-coregistration (BLE resolves presence/location robustly) and a positive signal for the anchor-placement thread.
Honest scope and what this changes
- Hypothesis: layout-topology-fingerprint-discriminability stays plausible. The result is suggestive for per-link discriminability but the operational sniffer-count is area-dominated, and the design cannot separate topology from area at n=6. No upgrade; no refutation.
- The real deliverable is a sharpened follow-on design: (1) stratify a larger ResPlan slice to decorrelate floor area from graph diameter (the present confound); (2) add the per-link wall-count-on-the-Tx–Rx-path metric the hypothesis note flags as the cheap topology variable (links, not floors, as the unit of analysis — far more statistical power); (3) more floors for a real stratum test. Until then the area+AP-count baseline is the honest default for sniffer-count budgeting.
- In-silico only; uniform-material; the multipath-richness confound (c-csi-fidelity-material) and real-layout transfer remain open. Reduction notebook
monad_knowledge/notebooks/python/csi_topology_discriminability.py(new, this session); artefacts under the session prefix. 12 source runs attached.