01KW04F1KYXHMM429A336Q16PW

CSI x BLE fusion under motion — corrected synthesis

Chain (C1, met). 6 coupled runs over {resplan-12439, 1374, 16157} x {2.4, 5.0 GHz}, walk-notebook(from_floor) -> sionna with ble.enabled + doppler.enabled. All gate-passed; each carries the co-registered links.parquet + ble_links.parquet (one solved CFR/frame) + csi.hdf5 + trajectory.parquet. (Loose wording caveat: walk metrics carry by-design NaN group fields; all three floors raised a two-abreast walkable-QC warning — valid runs.)

Reduction (C2, met). fusion_count builds, per (floor, band): ground-truth occupancy(t); a CSI feature (per-coarse-frame intra-frame Doppler-fading CV averaged over links); a BLE aggregate-attenuation feature; and the CSI-only / BLE-anchored / fused estimators, with the per-floor timeline.

Cross-floor result (C3, reported but the drift-bounding CLAIM is confounded). A CSI->count map calibrated on resplan-12439 and applied to the targets gives ~3.5-person mean MAE; periodic BLE recalibration on the target cuts it to ~1.0 (the figures: 1374 3.38->1.13, 16157 3.61->0.92; bands within 0.04). This is a mechanism illustration, not a 70.7% recovery rate, for two reasons the critic established: (a) leakage — the BLE "anchor" is an affine of RSSI attenuation fit to the same run's ground truth (per-person slope -14.3 dB on the source vs -73.0 on 1374, a 5x difference), so the fused path indirectly consumed target-floor truth; BLE here behaves as a proxy on-site recount, not the independent device-counter ble-ground-truth-sufficiency posits. (b) confound — source occ<=10 (seated) vs targets occ<=6 (through-traffic): a linear map fit high and applied low inflates from occupancy-range extrapolation, not purely geometry. Within-floor corroborates: BLE recal helps only the high-count furnished floor (12439 3.56->2.23); on the low-count floors the change (1.14->1.12, 0.93->0.92) is within single-seed noise. Band-dependence is negligible (a negative control, not a second evidence axis); single seed throughout, so every delta is a point estimate.

What it means (C4, framing held). The honest deliverable is the mechanism shape — a target-side absolute anchor can bound a relative CSI signal's cross-floor drift — which gives the Hybrid-Fusion chapter (weight 0.6) its first in-silico empirical and motivates ble-periodic-calibration / recalibration-trigger-from-drift. It does not establish the magnitude: the next iteration must (1) make BLE an independent device-counter (not an RSSI->count slope fit on truth), (2) match source/target occupancy ranges to isolate geometry from extrapolation, (3) add seeds for error bars. The real-data LOEO counterpart recovers ~42%; the synthetic overshoot is exactly the synthetic-to-real gap IP-106 exists to close. No detection-rate claim.

Critic (severity: high, disagreed with the original verdict — corrected above)

SC1/SC2 supported. SC3 borderline; SC4 only holds if the synthesis inherits the discipline (it now does).

Load-bearing findings (integrated, not hedged):

1. The 70.7% is partly tautological. The fused estimator re-fits CSI->count to a BLE-implied count, and that BLE count was fit to the same run's ground truth via the per-floor RSSI slope (-14.3 dB/person source vs -73.0 target 1374). "BLE as independent geometry-robust anchor" is false as executed — it consumed target truth. It is a proxy for an on-site recount.
2. Cross-floor drift confounded with occupancy-range extrapolation. Source 12439 occ_mean 6.17/max 10 (seated) vs targets occ_mean ~2.5-2.8/max 6 (pure through-traffic). A linear map fit high applied low inflates from out-of-range extrapolation; within-floor results (fusion null on low-count floors) corroborate that the "drift" is largely the source/target occupancy mismatch, not geometry.
3. Within-floor "fusion helps" is near-noise except on 12439; ~1-2% at sub-person MAE, N=1 seed.
4. Two bands add little (negative control only); single seed => no error bars.

Honest scope: in-silico, 3 floors, self-authored 2 dB Gaussian BLE noise, anchor calibrated on target truth. Establishes a mechanism shape, not a recovery magnitude; says nothing about BLE independence, geometry-vs-extrapolation separation, densities >~10, or hardware fidelity (IP-106).