Verdict
A co-registered BLE/RSSI calibration anchor is robust to a commodity-NIC impairment where the CSI phase feature is not --- a positive, physically-grounded reason to anchor count-without-identify calibration on BLE rather than on a second CSI feature. In-silico, under a self-authored impairment prior (IP-101 Q3=A); design input, not measured fidelity.
What we tested
A 3 × 5 grid --- n_agents ∈ {4, 8, 12} × impairment snr_db ∈ {5, 10, 20, 30, 40} dB --- on resplan-12439-floor-0 / csi-link-resplan-12439-multiroom (n_placements=24), each cell with ble.enabled and a fixed phase-impairment base (CFO 10 kHz, 0.3 dB / 3° IQ, 1° phase noise, 10-bit). 14/15 cells gate-passed and carry both the clean (csi.hdf5) and impaired (csi_impaired.hdf5) CFR; one cell (N8 @ 40 dB) raced the session attach and is excluded. The reduction (csi_ble_impairment.py) recomputes the narrowband BLE RSSI from both CFR tensors and compares its occupancy rank-correlation against the CSI Rician-K slope from the scalar metrics.
What we found
- BLE anchor ordering invariant (C2, C3 ✓). The BLE RSSI occupancy rank-correlation |ρ| ≈ 0.43 is preserved clean→impaired to within 0.01 at every SNR from 5 to 40 dB; the half-ρ floor sits below the tested range. The CSI Rician-K occupancy slope, a phase/variance feature, is distorted ×1.2–×3.1 over the same impairment (peaking near 20 dB --- the AWGN-shadow non-monotonicity inherited from
c-csi-impairment-sim-to-real). - Scale compresses, ordering does not. The BLE RSSI slope (absolute dB/person) is AWGN-compressed at low SNR (clean −25.9 → impaired −6.7 dB/person at 5 dB, recovering to ≈ −21 by 40 dB). The anchor therefore needs its scale periodically re-fit --- which is exactly what calibration does --- while its occupancy ranking never degrades. This strengthens rather than weakens the periodic-calibration case.
- Physics (C1 ✓). RSSI is a narrowband power read, immune to the carrier-offset / phase-noise terms that corrupt CSI phase; only AWGN touches it, and only its scale.
Honest scope (C4 ✓)
Self-authored impairment prior --- a commodity-NIC design input, not measured fidelity (IP-101 Q3=A). No estimator-accuracy claim. synthetic-csi-sim-to-real-transfer and ble-periodic-calibration stay plausible; the defeater is multi-week hardware capture on ≥2 sites. Bounded by N=6 ResPlan / single seed / one floor; the prior single-profile "×20" CSI distortion was a near-zero-baseline artifact, corrected here to the pooled ×2–3.