CSI room-scale calibration — does a WiMANS-scale room let the coupled sim reproduce the measured 2.4-vs-5 GHz occupancy gap?

Question

c-csi-band-calibration found the coupled sim does NOT reproduce the WiMANS-measured 5 GHz > 2.4 GHz occupancy-discriminability gap, with the low-N walking-CV pinned at the fading floor on the 350 m² resplan-12439 apartment. Its diagnostic isolated the lever: the scenario (duration_units 3600→1200 lifted CV ~750×), not body_loss (inert). But a second confound remains — room scale. WiMANS's rooms are small (5.5×5.5 m meeting, 11×5.5 m classroom); 1–5 occupants there strongly perturb a short link. On a sprawling apartment the same occupants barely touch most Tx–Rx paths. Does a WiMANS-scale room let the sim reproduce the measured band gap?

Framing discipline. Sim-to-real calibration probe, not a hypothesis test. Two priors are now controlled (scenario fixed at duration_units: 1200; body_loss held at the inert literature value 5.0). This campaign isolates the remaining geometry-scale variable. The honest deliverable is the per-band ρ comparison (sim small-room vs WiMANS) with bootstrap CIs.

What we already know

• c-csi-band-calibration (01KVZ0VV94…, refit + 3 seeds): the scenario refit clears the floor; the question is whether the band-gap ordering then matches real data — and whether floor scale matters.
• WiMANS real (2026-06-25 - WiMANS zero-shot — H4 on real CSI): 5 GHz ρ ≈ +0.7 cross-room robust (classroom/meeting/empty, n=1881/cell), 2.4 GHz ρ ≈ +0.3 — measured on small rooms.
• Methodology (audit 2026-06-24): reduction_stats CIs + sanity bounds + explicit attach-run loop.

Prerequisite (gating)

resplan-12419-floor-0 (135 m², the smallest of the six ResPlan floors carrying a csi-link-* experiment) is the candidate small floor. It must have seat occupiables for the seek_and_occupy walker (gis_occupiable --floor resplan-12419-floor-0 --subtype seat); if it has none, either author a handful via gis_create_occupiable or switch the scenario to a sustained through-traffic group (target_subtype: null + goal_bbox_cm) so no seats are needed. Confirm with a one-cell smoke before the fan-out (the floor-finickiness rule — see project_walk_notebook_floor_tuning).

What the supervisor does

1. Smoke one coupled run (N=2, 2.4 GHz) on the small floor; confirm the chain stages and walking-CV clears 1e-3.
2. Fan out N∈{1..5} × {2.4, 5.0 GHz} × seeds {0,1,2} (30 coupled runs, where=local or CI), attach each explicitly.
3. Reduce (csi_temporal_dynamics.py + the band-gap comparison from reduction_stats): per-band ρ with bootstrap CI, multi-seed CI, sim-vs-WiMANS table.
4. Synthesise the scale verdict.

Out of scope

• Hardware impairments / the BLE anchor.
• Absolute CV matching (rank + band-gap only; real CV is noise-inflated).

Expected interpretation

1. Small room reproduces the gap (sim 5 GHz ρ > 2.4 GHz ρ, CIs separated, matching WiMANS) → the large-floor mismatch was a scale artefact; the ray-tracer is band-faithful at WiMANS scale. Strengthens occupancy-csi-variance and validates "prefer 5 GHz".
2. Small room still inverts/flattens the gap → the defect is the ray-tracer band model (e.g. num_subcarriers/bandwidth per band, max_depth, the body scatterer's frequency response), a deeper refit target for exp-csi-calibration than scenario or room scale.
3. Prerequisite unmet (no seats / chain won't stage) → platform finding; author occupiables or switch scenario, then proceed.