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

01KT9BJ5172M9HHXXBE67DYS11

finished 2026-06-04 13:01:03.399550+00:00 → 2026-06-04 13:40:53.159937+00:00 · 48 runs · supervisor: react-agent

“Hypothesis inverted, not confirmed: blockage-dominance lives on the strong-LOS in-room link (N* ≈ 1.3–6.8), while 1-wall links are variance-dominant from N=0 with a seed-robust −0.05 dB/person slope; all N=0 ladders monotone at depths 3/4/6 (the earlier non-monotone ladder did not reproduce), and the 2.4 GHz drywall step is only ~0.5 dB. Criterion 1 failed as a layout finding — the staged layout has no ≥2-wall link (geometric ground truth); re-stage before any re-test. Two criteria honestly failed; the null + inversion is chapter-relevant either way.”

Archive snapshot, as of 20 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
The multi-room link layout stages with three link classes (in-room, 1-wall, ≥2-wall) and every run returns the scalar floor + links.parquet; no NaN/Inf. no
Baseline (N=0) path loss increases monotonically with wall count between Tx and Rx — the ITU through-wall material chain is physically ordered — OR the non-monotonicity is shown to be a ray-depth truncation artefact (the N=0 ladder monotonizes as max_depth rises from 4 to 6); a persistent non-monotone ladder at depth 6 is reported as a small-scene physics caveat, not silently passed. yes
Through-wall links show a blockage-dominant low-occupancy regime: crossover N* > 0 and a steeper mean-attenuation slope than the in-room control link (whose N*≈0 in csi-static-occlusion), with the per-link-class slope reported as mean ± CI across seeds {0,1,2} — single-seed point estimates are not acceptable (the 2026-06-03 diary's <1 dB single-seed result is the weakness this session repairs). no

Synthesis

What we tested

52 exp-csi-static runs on resplan-12439-floor-0 / csi-link-resplan-12439-multiroom, all attached and gate_passed: the 48-cell main grid (n_agents {0,1,2,3,4,6,8,12} × bands {2.4, 5.0 GHz} × seeds {0,1,2}, 24 placements, depth 3) plus the 4-cell depth-diagnostic arm (N=0, depth {4,6}, both bands). The corpus was assembled resiliently (two sweep-level S3/tunnel aborts; completed cells inventoried and the 30 missing cells gap-filled as independent runs — cell coverage verified 52/52). Reduction: csi_through_wall.py — per-link wall counts derived geometrically from the staged scene (segment-crossing count, ground truth), N=0 path-loss ladders per (band, depth), per-class attenuation/CV slopes with across-seed spread, and the normalized blockage↔variance crossover.

What we found

Criterion 1 — FAILED as stated (layout finding). The layout does not deliver three link classes: geometric wall counts are {link 4: 0 walls; links 0–3: 1 wall each}. No ≥2-wall link exists — the bedroom-3 path crosses a single wall segment (likely through a doorway-gap alignment). All class-resolved conclusions below therefore compare in-room (1 link) vs 1-wall (4 links) only.

Criterion 2 — MET, with the depth question resolved. The N=0 wall-count path-loss ladder is monotone at all of depth {3, 4, 6} on both bands — the previously-reported non-monotone 2.4 GHz ladder did not reproduce on this seed-replicated corpus. Magnitudes: the 1-wall step is 11.0 dB at 5 GHz but only ~0.5 dB at 2.4 GHz (ITU drywall is nearly transparent there) — which by itself explains the 2026-06-03 diary's "<1 dB" through-wall result as seed-robust physics, not noise.

Criterion 3 — FAILED as stated, and inverted. Through-wall links show crossover N* ≈ 0 on every seed and band (variance-dominant from the first occupant) with a small attenuation slope (−0.050 ± 0.019 dB/person at 2.4 GHz, −0.064 ± 0.024 at 5 GHz; mean ± std across 3 seeds). The in-room control is where blockage dominates at low occupancy: crossover N* ≈ 2.0–6.8 at 2.4 GHz (seed spread; noisy) and ≈ 1.3–3.6 at 5 GHz. A physically coherent reading: blockage-dominance requires a strong LOS to block — present in-room, already absent behind a wall where the field is diffuse. This inverts the brief's expectation rather than merely failing it.

What it means for the thesis chain

The geometry-continuum claim needs reformulating: the L4.5↔L4.6 axis is not "in-room=variance, through-wall=blockage" but "strong-LOS=blockage-first, weak-LOS=variance-only" in this residential geometry at these ranges. Depatla-style through-wall counting (LOS-crossing dips) presupposes a strong through-wall LOS at short range — a regime these drywall apartment links at 2.4 GHz do not produce (0.5 dB wall step). Caveats bounding all of this: one floor, one layout with only two link classes, ray-traced with the uniform-drywall palette that c-csi-fidelity-material/01KT9BJ82A… has just shown to be load-bearing, depth ≤ 6, 24 placements. Follow-ons: (1) re-stage the multiroom layout with a true ≥2-wall link (move the bedroom-3 Rx off the doorway axis) before any re-test; (2) a concrete-palette variant of this grid — at 11 dB/wall-class sensitivity the inversion may not survive heavier walls; (3) the Depatla inter-event-time recreation needs a short-range strong-LOS through-wall link, not this layout.

Attached runs

Run Gate Purpose Replay
43NX39VF
6K2K9R21
7C754GKA
C5YWZ6R6
9PBX0AW3
W3DJWXYR
K82YM78E
0BWW747M
VT823XYA
2EAFT795
H20NCTZB
Y6DN5DJ7
Z5PBDAXW
D461ZN2G
NE0K8WS7
JW1GA0J7
K30VT1K2
A92GVZ0H
09TX7F6G
3CYQC0QN
BER8R9P8
1VAA8YAC
JMP12X0N armt-gapfill
A66HBY1F armt-gapfill
BAD2H0J3 armt-gapfill
7KZ47YAN armt-gapfill
176N21H9 armt-gapfill
KNG7YWRC armt-gapfill
0A4BS9MZ armt-gapfill
4RF6F8VG armt-gapfill
HQ7WCZB2 armt-gapfill
7V6CSHV0 armt-gapfill
9KV5F7KW armt-gapfill
SZXP9YKW armt-gapfill
S9125FRT armt-gapfill
97KR943R armt-gapfill
1R0FQKKJ armt-gapfill
WYTA4JTX armt-gapfill
6NHG63VR armt-gapfill
1XP5TRSE armt-gapfill
2QE2W4B9 armt-gapfill
KYPK1YVR armt-gapfill
H3BRQFYS armt-gapfill
JA00A0Z8 armt-gapfill
B86079EQ armt-gapfill
FJJE0GSN armt-gapfill
5N9Q3Y71 armt-gapfill
063VR9YA armt-gapfill