Question
c-ble-csi-coregistration found that a co-registered BLE anchor robustly resolves occupancy presence (empty vs occupied, d = 2.3–3.7) but saturates — no graded headcount — on the resplan-12439 multiroom floor, with the anchors dispersed one-per-peripheral-room. Its synthesis named the fix: "anchor placement near occupant zones / per-link calibration." This campaign tests that directly: does topology-informed placement — anchors concentrated on the high-betweenness corridor hub (where occupants transit and cluster) plus the largest connected room — recover the graded count the dispersed placement lost?
Framing discipline (carried from c-ble-csi-coregistration): in-silico motivation for ble-periodic-calibration, not a hypothesis test. The BLE event noise is a self-authored Gaussian; any accuracy headline against it is circular. The honest deliverable is a placement comparison of the graded-count slope + resolvable-ΔN table. The real defeater is hardware (IP-106).
Why this placement, by topology
The resplan-12439 room-adjacency graph (IP-040 NRG via gis_connectivity) is hub-and-spoke:
living-0 is a 135 m² corridor that every room connects to — the highest-betweenness, highest-degree
cell by a wide margin. The dispersed baseline put one anchor in each peripheral bedroom; once any
single body lands on a given Tx→bedroom path that link saturates and adding bodies elsewhere does not
grade it further. The topology-informed placement instead concentrates 5 Rx on the hub (3 in the
living-0 corridor) + the large unanchored bedroom-1 (47 m², 2 Rx) — the rooms occupants actually
occupy and transit, so successive occupants progressively block the short hub-spanning links and the
aggregate ΔRSSI stays graded above the saturation knee. Grounding: zhen2022_bb0b (layout-adaptive
BLE placement +26.4 %), afghantoloee2021_8628 (topology-dependent optimal placement),
depatla2018_97a9 (room geometry dominates count bias).
What the supervisor does
18 runs on the new experiment csi-link-resplan-12439-topo-anchors (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 the Rx placement. Reduction pools ble_links.parquet per run and compares,
head-to-head against the dispersed baseline's 18 noise-2.0 runs: per-placement ρ(N, baseline-normalised
ΔRSSI), the empty→occupied step, and the within-occupied resolvable-ΔN effect-size table. Local
deterministic supervisor (no LLM fan-out); systematic re-run routes to CI.
Out of scope
- Any BLE-vs-CSI estimator MAE (circular against self-authored noise).
- Moving crowds (the coupled variant is a follow-on once this static contrast is read).
- Hardware noise / MAC scheduling — that is IP-106.
Expected interpretation
- Recovered (topology placement: ρ(N,ΔRSSI) steeper + monotone past the saturation knee, finer resolvable ΔN within the occupied range than dispersed) → "topology-informed anchor placement recovers graded count in sim — a placement design input for the hardware campaign." Feeds ble-periodic-calibration (still plausible) + the IP-106 anchor-placement decision.
- Not recovered (both placements saturate equally) → the saturation is intrinsic to single-Tx narrowband blockage at this density, not a placement artefact; the graded count needs multiple Tx or the CSI feature, not anchor relocation. A clean negative that narrows the BLE arm's claim.
- Criterion 1 fails → plumbing/placement finding; fix before interpretation.