Asking the fleet what it is doing…
monad-knowledge Wi-Fi sensing lab · FIIT STU
Campaign

CSI band calibration — does the coupled sim reproduce the WiMANS-measured 2.4-vs-5 GHz occupancy-discriminability gap?

c-csi-band-calibration · exp-csi-crowd

Archive snapshot, as of 18 h ago — the run corpus is rebuilt once a day, so this page is not a live reading. The fleet panel is the live one; it refreshes every 30 s.

Sessions

state
Session State Runs Synthesis Criticism Figures Verdict
CRQ3CJDE 2026-06-25T09:13 finished 30 ✓ — 1 Refit completed (30 runs, 3 seeds). The lever was SCENARIO (duration_units 3600→1200, ~750× CV lift), NOT body_loss (tested inert 5→20, kept 5.0). With the floor cleared the sim gives clean monotone CV(N) and genuine 3-seed replication (0% inert — fading_seed varied). Per-band ρ: 2.4=+0.74 [0.35,0.90] / 5.0=+0.75 [0.35,0.92] — NO band gap (sim reproduces the original c-csi-crowd-temporal "both bands equal" claim robustly). But this STILL does not match real WiMANS (5 GHz≫2.4 GHz, gap +0.43). Defect localised to the ray-tracer FREQUENCY PHYSICS (not scenario, not body_loss); room-scale is the remaining probe (c-csi-roomscale-calibration). Supersedes the dur=3600 session 01KVYZ674X.
VJDRNKW5 2026-06-25T08:43 finished 10 ✓ — 1 Sim does NOT reproduce the WiMANS-measured band gap. At the WiMANS occupancy range (N=1-5) on the large resplan-12439 floor, the coupled walking-CV signal sits at the fading floor (1e-6–3e-5; cf. real ~0.05–0.24), so band differences are noise: sim ρ(N,CV) 2.4=+0.90 [CI +0.11,+1.0] / 5.0=+0.10 [CI −1.0,+1.0] vs real +0.28 / +0.71 — the sim flattens/inverts the real 5 GHz advantage, and the 5 GHz CI is uninformative. A calibration/scenario refit target for exp-csi-calibration, NOT a refutation of H4 (the real WiMANS data supports it). Confirms the audit's point: small-N single-seed sim ρ is fragile (the prior ρ=+1.0 held only at N=4-14).

Brief

Question

The WiMANS zero-shot (2026-06-25 - WiMANS zero-shot — H4 on real CSI) found, on real Intel-5300 CSI, that 5 GHz is materially more occupancy-discriminative than 2.4 GHz (ρ(moving-count, CV) = +0.71 vs +0.28; presence d = 1.92 vs 1.10). Our coupled JuPedSim→Sionna programme (c-csi-crowd-temporal) reported the two bands tracking identically (ρ ≈ +1.0 both). Either the sim's N-ladder (4–14) hid the gap that only shows in the 0–5 range WiMANS covers, or the ray-tracer's body/material prior is mis-calibrated and over-states 2.4 GHz. Re-run the coupled chain at the WiMANS occupancy range on both bands and compare the simulated band gap to the measured one.

Framing discipline. This is a sim-to-real calibration probe, not a hypothesis test. The honest deliverable is the head-to-head band-gap comparison (sim vs WiMANS) with bootstrap CIs. A match validates the ray-tracer's band behaviour and confirms "prefer 5 GHz" as a design rule; a mismatch localises the defect to the self-authored prior (body_loss_db, wall material) and yields a concrete refit target for exp-csi-calibration. No accuracy headline against the synthetic noise.

What we already know

  • 2026-06-25 - WiMANS zero-shot — H4 on real CSI (real, 5643 samples/band): presence transfers on both bands; graded counting transfers at 5 GHz only — the measured gap this campaign tests.
  • c-csi-crowd-temporal (01KVWYPHMHZVGRYQBW5Z9ZYEQ7): coupled chain validated, ρ=+1.0 both bands — but on N∈{4,6,8,11,14}, a range that barely overlaps WiMANS 0–5, and read without a CI.
  • occupancy-csi-variance (H4) now carries its first real-data anchor; this campaign is the sim-side of closing the calibration loop.
  • Methodology fixes from the 2026-06-24 audit are now available and MUST be used: monad_knowledge/sim/reduction_stats.py (physical sanity bounds, bootstrap_spearman_ci, multi_seed_ci inert flag) and the explicit sim_campaign_attach_run loop after CLI fan-out (the topology session sealed with 0 runs attached because the --label session_id= does NOT auto-attach — see project_local_campaign_run_gotchas).

What the supervisor does

  1. Fan out exp-csi-crowd over the 6 × 2 (n_agents × band) grid on resplan-12439-floor-0 (12 coupled runs, seed 0; trajectory_frame_stride=10 to keep CPU-only RT tractable). Systematic on CI for the full grid; a one-cell smoke (N=2, 2.4 GHz) locally first to confirm the chain stages floor_geometry.json + scene.json.
  2. Attach each run explicitly (sim_campaign_attach_run loop) — do not rely on the launch label.
  3. Run the reduction (csi_temporal_dynamics.py + the band-gap comparison): per-band walking-CV(N), ρ(N,CV) with bootstrap CI, the degenerate-cell sanity pass, and the sim-vs-WiMANS band-gap table.
  4. Synthesise the calibration verdict (match → validate 5 GHz rule; mismatch → prior refit target).

Figure render request

csi_band_calibration — walking CV(N) per band (sim) with bootstrap-CI ribbons, overlaid with the WiMANS per-count CV trend (rank-aligned), and a side bar of ρ(N,CV): sim-2.4 / sim-5.0 / real-2.4 / real-5.0 with CIs.

Out of scope

  • Hardware-impairment modelling (that is c-csi-impairment-sim-to-real / exp-csi-calibration).
  • The BLE anchor (separate thread); this is the CSI-variance feature only.
  • Absolute MAE against WiMANS (scales differ; rank + band-gap only).

Expected interpretation

  1. Sim reproduces the gap (sim ρ_5.0 − ρ_2.4 > 0, CIs separated, matching the WiMANS sign) → the ray-tracer is band-faithful in 0–5; "prefer 5 GHz for the counting feature" is validated end-to-end. Strengthens occupancy-csi-variance toward strong.
  2. Sim shows no gap (both bands ρ≈+1.0 even at N 0–5) → the prior over-states 2.4 GHz; the defect is body_loss_db / wall material, and the refit target (match the measured ρ gap) is the deliverable for exp-csi-calibration. A valuable, honest negative.
  3. Criterion 1 fails (degenerate scalars / chain doesn't stage) → platform finding; fix before interpretation.