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

01KVZS1KVBW7HBMM9XKYPPCW5A

finished 2026-06-25 16:15:40.907140+00:00 → 2026-06-25 17:21:06.209124+00:00 · 18 runs · supervisor: none

“Topology-informed anchor placement (3 Rx on the high-betweenness living-0 corridor hub + 2 in the large bedroom-1) PARTIALLY RECOVERS the graded count the dispersed one-per-peripheral-room placement lost to saturation. Aggregate ΔRSSI slope flips from flat +0.08 dB/person (dispersed, saturates) to monotone -0.48 dB/person (topology); worst-seed ρ improves -0.09→-0.49 (all seeds now clearly negative); presence preserved (d 2.32→2.56). On the robust effect-size view, topology placement makes coarse graded steps ΔN≥4 reliably resolvable (mean d≥1) where dispersed resolves NONE — but fine ΔN=2 grading stays unresolved and the strongest signal is at N=12. In-silico motivation for ble-periodic-calibration (held plausible); an anchor-placement design input for IP-106, not graded-counting proof.”

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.

Success criteria

CriterionResolved
18 runs on csi-link-resplan-12439-topo-anchors (N{0,2,4,6,8,12}×seeds{0,1,2}, noise 2.0, 2.4 GHz, 5 Rx: 3 living-0 hub + 2 bedroom-1) emitted ble_links.parquet + CSI from one CFR, no NaN/Inf. Co-registered by construction. yes
Head-to-head vs the dispersed baseline (c-ble-csi-coregistration 01KVX0WCD5, 18 noise-2.0 runs): reduction csi_ble_anchor_placement.py compares per-link-baseline-normalised ΔRSSI. Topology graded slope -0.48 vs dispersed +0.08 dB/person; topology ρ∈[-0.89,-0.49] (all seeds monotone) vs dispersed [-0.83,-0.09] (seed2 broken); coarse steps ΔN≥4 resolvable at mean d≥1 under topology, none under dispersed. Partial recovery of graded count. yes
Framing discipline preserved: in-silico motivation (self-authored BLE noise), not a test; the recovery is a hardware anchor-placement design input for IP-106 (concentrate anchors on high-betweenness hubs + large occupant rooms), bounded — fine ΔN=2 grading unresolved, single narrowband Tx, uniform-drywall. ble-periodic-calibration held plausible. yes

Synthesis

c-ble-anchor-placement — session 01KVZS1KVBW7HBMM9XKYPPCW5A (2026-06-25)

Verdict: topology-informed anchor placement PARTIALLY RECOVERS the graded count the dispersed placement lost — a hardware anchor-placement design input, not graded-counting proof. Interpretation #1 (bounded), per the brief.

This answers the open question c-ble-csi-coregistration left: that session found the co-registered BLE anchor resolves occupancy presence (d 2.3–3.7) but saturates — no graded headcount — with anchors dispersed one per peripheral room. Its synthesis named the fix as "anchor placement near occupant zones." We placed 5 Rx topology-informed: 3 on the high-betweenness living-0 corridor hub (the 135 m² cell every room connects to — highest NRG degree/betweenness) + 2 in the large unanchored bedroom-1 (47 m², door-connected to the hub), same Tx, same N-sweep.

Setup

New PostGIS experiment csi-link-resplan-12439-topo-anchors (6 devices placed). 18 exp-csi-static runs N {0,2,4,6,8,12} × seeds {0,1,2}, noise 2.0, 2.4 GHz, n_placements 24 — the noise-2.0 main grid of c-ble-csi-coregistration with the only change being Rx placement. Head-to-head vs that session's 18 dispersed-placement runs. Reduction csi_ble_anchor_placement.py (per-link empty-room-baseline-normalised ΔRSSI).

Result

metric dispersed (baseline) topology (hub+bedroom-1)
graded slope (dB/person) +0.08 (flat — saturates) −0.48 (monotone)
ρ(N,ΔRSSI) per seed −0.83, −0.66, −0.09 −0.89, −0.77, −0.49
presence d (0 vs first occupied) 2.32 2.56
coarse steps ΔN≥4 resolvable (mean d≥1)? none yes (d 1.2→3.2)
  • The aggregate response becomes monotone. Dispersed ΔRSSI drops at N=2 then bounces back toward 0 (the saturation signature); topology stays monotone-decreasing and steepens strongly by N=12. All three topology seeds are clearly negative — the dispersed "broken" seed (ρ=−0.09) recovers to −0.49.
  • Coarse grading is recovered; fine grading is not. On the robust effect-size view (right panel), topology placement makes ΔN≥4 steps resolvable at mean d≥1 where dispersed resolves none — but ΔN=2 stays below threshold (topology d=0.31), and the strongest separation is at high occupancy (N=12). So topology placement buys coarse graded count, not fine headcount, from a single narrowband Tx.

Honest scope

In-silico motivation for ble-periodic-calibration (held plausible) — the BLE event noise is self-authored, so this is a placement design input, not proof. Single Tx, uniform-drywall, static (seated-crowd proxy) regime. The deliverable: concentrate anchors on high-betweenness hubs + large occupant rooms (vs dispersing to peripheral rooms) — a concrete anchor-placement decision for the IP-106 hardware testbed, where the real defeater lives. Reduction monad_knowledge/notebooks/python/csi_ble_anchor_placement.py; fig under the session prefix + _attachments/ble-anchor-placement/. 18 runs attached; dispersed baseline reused from 01KVX0WCD5SXYQDSTJZ77Q1ANT.

Attached runs

Run Gate Purpose Replay
WVB2GK0Q
CTZCC27T
NS19DKDR
WEDC6QVV
DM3SMZNR
044JYHRC
QE2QH8R9
W2J873NM
2K4AWTDE
QW6GRJA2
5P8962RY
63FYHKX4
WK05Q9CQ
JD7Q22BE
BNX95MN2
Q23167D7
49GH09VN
2RVKAJD9