c-csi-roomscale-calibration — session 01KVZ2FNS4DHTQ99R8MSCNFES5 (2026-06-25)
Verdict: room scale is ELIMINATED as the sim→real band-gap confound. By exclusion, the 5 GHz advantage the simulator lacks is the ray-tracer's frequency physics.
This closes the diagnostic chain opened by c-csi-band-calibration: three candidate causes for "the sim doesn't reproduce the WiMANS 5 GHz > 2.4 GHz occupancy-discriminability gap" — scenario, body_loss, room scale — are now all tested and rejected.
Setup
The smallest ResPlan floor carrying a csi-link-* experiment, resplan-12419 (135 m²) vs the 350 m² resplan-12439, to test whether occupants nearer the links recover the band gap. The floor had no seats, so 6 seats were authored in bedroom-0 (near csi-rp12419-rx1; provenance computed) and a smoke confirmed the walker reaches them (walking-CV 0.96). 30 coupled runs N∈{1..5}×{2.4,5.0 GHz}×seeds{0,1,2}, refit scenario duration_units=1200, all gate-passed, explicitly attached.
What we found
| band | sim small (12419) | sim large (12439) | WiMANS real |
|---|---|---|---|
| 2.4 GHz | +0.29 [−0.27, +0.73] | +0.74 | +0.28 |
| 5.0 GHz | +0.29 [−0.28, +0.73] | +0.75 | +0.71 |
| band gap | +0.00 | +0.01 | +0.43 |
- Opposite failure mode, same conclusion. The small floor does not starve the signal — walking-CV is high (0.07–0.12) because occupants near the 3 short links strongly perturb the channel. But it saturates: the perturbation nearly maxes out by N=1, so the CV is only weakly graded (ρ=+0.29, CIs spanning 0). The large floor under-couples (better dynamic range, ρ=+0.74); the small floor over-couples (saturated, ρ=+0.29). Different pathologies — but neither shows any band gap (both +0.00/+0.01).
- Only real data has the 5 GHz advantage (+0.43). 3 genuine seeds (0% inert), honest CIs.
Conclusion — the defect is pinned
Scenario (fixed at dur=1200), body_loss (inert 5→20 dB), and room scale (this run, both floors equal-band) are all eliminated. The simulator treats 2.4 and 5 GHz identically for occupancy variance regardless of floor or scenario, because the ray-tracer uses the same per-band parameters — num_subcarriers/bandwidth_hz, the homogeneous body-scatterer response, and max_depth=3 for both carriers. It physically cannot manufacture the shorter-wavelength sensitivity real 5 GHz hardware shows. The exp-csi-calibration refit target is therefore the frequency-dependent RT model, not scenario/coupling/geometry: per-band subcarrier count + bandwidth, a frequency-dependent body permittivity/scattering response, and possibly higher max_depth at 5 GHz.
In-silico calibration probe; rank + band-gap only. 6 seats added to resplan-12419 (persistent DB change). Figure + metrics under the session prefix and _attachments/wimans-zeroshot/fig_roomscale_3way.png. 30 runs attached.