01KW0KGZV29HF9YGWME2S3XCPB

c-flat-day-csi — the ray-traced CSI behind the explorer

Setup. The IP-107 reactive-FSM crowd (21 personas / 29 agents / 18 behaviour FSMs) was coupled to ray-traced Sionna CSI over the dense-10 receiver layout (csi-link-resplan-12439-dense10: 10 csi-rp-rx-* + csi-rp-tx, all on resplan-12439-floor-0). 5 seeds, run locally via the exp-csi-crowd chain (walk-notebook(from_floor) → sionna-csi-runner, --where local, ~44 s each). Runs 01KW0K4S2RJ636WPMFAXDC4J1S, …6ZQVC3B, …8F11CSQ, …A1G1BG, …BJ1DG3.

1 — Chain runs end-to-end (✓). All 5 seeds gate_passed, emitting links.parquet (per-frame mean_amp_db/rician_k_db/delay_spread_ns, 103 RT frames × 10 links) + csi.hdf5, uploaded to S3. Occupancy arc 1→21 over the 35-min evening.

2 — Ray-tracer reproduces the occlusion signature (✓). Amplitude collapses from −60.5 dB (clear link) to −90.1 dB when a body crosses the Tx→Rx line; mean attenuation 29.0 dB/person, Rician-K slope −9.8 dB/person, variance-monotonicity −1.0. The qualitative behaviour the first-order proxy assumed is reproduced by full multipath EM.

3 — The living-0 observability gap is CONFIRMED, and sharper than the proxy claimed (✓). living-0 is the busiest room (peak ≈ 13–16) and the best-instrumented (4 of 10 receivers + the transmitter). Yet the correlation between its live occupancy and its in-room link attenuation is −0.06 — essentially zero — and its amplitude variance is the lowest of any occupied room (51.9). A near-empty bedroom (bedroom-2, peak ≈ 1.6, single receiver) reads its occupancy at +0.53. Across rooms, Spearman(peak occupancy, observability) = −0.70: the busier a room, the less its CSI tracks it. This is an observability gap, not merely a coverage gap — the dense hub saturates (constant occlusion + rich multipath → amplitude is always crushed and stops responding to marginal occupancy).

4 — Per-link variance vs room occupancy (✓, with nuance). In legible rooms the link's attenuation tracks in-room occupancy (bedroom-2 0.53, bedroom-3 0.24, kitchen 0.19); in the saturated hub it does not. So the sensing target is recoverable where the crowd is sparse, exactly inverse to where counting matters most.

5 — Artefacts feed the explorer (✓). A self-contained interactive dashboard ("The Flat-Day Observatory") was built over the ray-traced bundle (identical schema to the proxy links.parquet): trajectory replay with occlusion-reddened Tx→Rx rays, per-link signal wall, the 18-FSM atlas with live current-state pulse, and the observability scatter. Published as a claude.ai artifact; bundle + dashboard on S3.

Caveats. Single authored evening; quasi-static ray-tracing (Doppler off — temporal CV is therefore band-independent here); 5 seeds; a synthetic cast, not a measured crowd. The paradox is an in-silico mechanism; magnitude awaits real CSI (IP-106).</synthesis_md> ["Coupled chain runs end-to-end (FSM crowd -> Sionna -> links.parquet + csi.hdf5)", "Ray-traced CSI reproduces the occlusion signature the proxy shows", "Quantify the living-0 gap: is the busiest room still the least observable?", "Per-link CSI variance correlates with that link's room occupancy", "Artefacts feed the CSI explorer (same schema) for ray-traced replay"]

Self-critique (deterministic supervisor)

• Observability metric is a linear correlation. corr(in-room occupancy, −in-room link amplitude) will miss non-monotone or saturating relationships — which is precisely the living-0 regime. A near-zero correlation there is consistent with both "blind" and "saturated-but-informative-nonlinearly". The honest claim is the linear signal vanishes; a non-linear feature (or the variance itself) might still carry count information, untested here.
• Quasi-static RT. Doppler is off, so the temporal micro-fading that earlier lifted CSI→count R² from 0.09→0.36 (project_csi_ble_fusion_campaign) is absent. The observability gap may narrow with a Doppler-rich temporal feature; this campaign measures the amplitude channel only.
• Single seed for the replay/observability detail; 5 seeds only for the scalar slopes. Per-room correlations are pooled over 10 runs but not bootstrapped per-room — the −0.70 Spearman is over 6 room points (small n).
• Synthetic cast. The crowd is LLM-authored, not measured; the geometry is one flat. Generalisation to real layouts/crowds is asserted by mechanism, not shown.