c-csi-band-calibration — session 01KVYZ674XX011J30PVJDRNKW5 (2026-06-25)
Verdict: the coupled sim does NOT reproduce the WiMANS-measured 2.4-vs-5 GHz occupancy-discriminability gap. This is a sim-to-real calibration finding (interpretation #2 of the brief), executed entirely locally (12 coupled walk-notebook→sionna runs, 10 gate-passed). It directly answers the question the WiMANS zero-shot (2026-06-25 - WiMANS zero-shot — H4 on real CSI) raised.
What we ran
exp-csi-crowd on resplan-12439-floor-0 / csi-link-resplan-12439-multiroom, validated seek_and_occupy/seat scenario (copied from c-csi-crowd-temporal), grid N∈{0,1,2,3,4,5} × {2.4, 5.0 GHz}, seed 0. N=0 failed (the seat-seeker scenario needs ≥1 agent — 6 s gate-fail); the 10 N≥1 runs passed (~117 s each). Runs explicitly attached via the attach-run loop (the launch label does not auto-attach — the bug that left the first topology session at 0 runs).
What we found
| band | sim ρ(N, walking-CV) [CI95] | WiMANS real ρ |
|---|---|---|
| 2.4 GHz | +0.90 [+0.11, +1.00] | +0.28 |
| 5.0 GHz | +0.10 [−1.00, +1.00] | +0.71 |
- Root cause — signal starvation at low occupancy. The simulated walking-window CV is 1e-6 – 3e-5 across N=1–5 — at the fading floor (cf. WiMANS real CV ~0.05–0.24, and the sim's own ~5e-3 at N≥6 in c-csi-crowd-temporal). On this large multiroom floor 1–5 agents park too fast / are too sparse to perturb the channel, so the walking window is noise-dominated. The 5 GHz curve is non-monotone (rises to N=2–4, crashes at N=5) → ρ=+0.10 with a CI spanning the entire [−1, +1] (
bootstrap_spearman_ci, the audit's honesty tool — a bare point estimate would have hidden this). - The sim flattens/inverts the real band gap. Real WiMANS: 5 GHz ≫ 2.4 GHz. Sim: a weak 2.4 GHz trend (driven only by the N=1→2 jump, then saturated) and an uninformative 5 GHz. The sim's band behaviour at the deployment-relevant low-occupancy range is not trustworthy.
- Methodological confirmation of the audit. The earlier "ρ=+1.0 at both bands" (c-csi-crowd-temporal) held only at N=4–14 with 3 seeds; at the WiMANS range with 1 seed it collapses. Small-N single-seed sim ρ is fragile — exactly why CIs are now mandatory.
Honest scope and the refit target
This does not refute occupancy-csi-variance (H4) — the real WiMANS data supports it (5 GHz ρ=+0.71). It refutes the simulator's fidelity at low occupancy on a large floor. The concrete deliverable for exp-csi-calibration:
- Refit the body-loss / wall-material prior so a single body perturbs the channel enough to clear the fading floor at low N (the self-authored
body_loss_db=5.0+ uniform-drywall under-couples here). - Scenario fix: a smaller/denser room or sustained-motion scenario (not park-and-stop on a 350 m² apartment) so 1–5 occupants produce measurable walking variance.
- Re-run with ≥3 seeds and a wider N before trusting any band claim.
In-silico calibration probe; single seed (effective_n=1, CFSM deterministic); the comparison anchor is the real WiMANS rank gap, not absolute CV. Figure + csi_band_calibration.metrics.json under the session prefix and _attachments/wimans-zeroshot/fig_band_calibration_sim_vs_wimans.png. 10 runs attached.