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

01KT9BJ82A61X7T7HQYY7D9309

finished 2026-06-04 13:01:06.506344+00:00 → 2026-06-04 13:26:14.205733+00:00 · 8 runs · supervisor: react-agent

“Material palette is load-bearing: drywall→concrete swap shifts first-order CSI by KS-D 0.74–1.00 and second-order (delay spread) by 0.27–0.60, all p ≪ 1e-3, on identical geometry/placements — prior default-material results are now explicitly conditional on the (unverified) drywall import tag; per-wall tagging and furniture scatterers escalate. Synthesis inlined by the supervisor (tiny 8-run corpus, IP-084 cost rule); critic skipped with caveats pre-integrated — the conditional framing distinguishes palette sensitivity from mis-tagging error.”

Archive snapshot, as of 1 day 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
Both arms stage and run cleanly on the SAME link layout and placement seeds: arm ∈ {as-stored palette (uniform drywall on this floor), forced `scene_overrides.uniform_material: concrete`} × band {2.4e9, 5.0e9} × occupancy {0, 6} — 8 runs, every one returning the scalar floor + links.parquet (now incl. `delay_spread_ns`); no NaN/Inf. yes
The KS test is computed per (band, occupancy, link-class) on BOTH first-order (per-placement mean_amp_db ECDF) and second-order (delay_spread_ns ECDF) statistics, with the KS-D and p reported in a table — not a prose claim. yes
A decision is recorded either way: KS-D < 0.15 on all second-order cells → 'uniform scenes are licensed for material-insensitive campaigns' (cheaper staging); any KS-D ≥ 0.15 → 'a material-fidelity caveat is attached to every prior default-material CSI result' and the affected sealed sessions are named. yes

Synthesis

What we tested

Eight exp-csi-static runs (sweep 01KT9BPF4V470GJSQQAG1XE67F, all attached, all gate_passed) on the through-wall rig (resplan-12439-floor-0 / csi-link-resplan-12439-multiroom, 24 placements, depth 3, seed 0): arm ∈ {as-stored palette, scene_overrides.uniform_material: concrete} × band {2.4, 5.0 GHz} × occupancy {0, 6}. The runner's 2026-06-04 extensions carried the arm switch (scene_overrides) and the second-order statistic (delay_spread_ns from the CFR power-delay profile). Two-sample KS tests were computed per (band, occupancy) on per-placement mean_amp_db and delay_spread_ns, pooled and per link (table: material_ks.csv in the session scratch dir).

Discovered while authoring the brief: the staged palette is already uniform drywall on every current floor — the import does not tag per-wall materials. The original heterogeneous-vs-uniform framing was therefore counterfactual; this session tests palette sensitivity (drywall → concrete, a large but realistic worst-case perturbation: ε_r 2.94 → 5.31 plus conductivity).

What we found

Criterion 1 — both arms staged and ran cleanly, 8/8 runs with links.parquet incl. delay_spread_ns, no NaN/Inf. Criterion 2 — the KS table is decisive on every cell: pooled first-order KS-D = 0.74–1.00 (mean_amp_db) and pooled second-order KS-D = 0.27–0.60 (delay_spread_ns), all p < 4e-4; per-link maxima reach 1.00. Criterion 3 — the decision rule fires on its strongest branch (KS-D ≥ 0.15 on first and second order): the wall-material palette is load-bearing for both blockage and multipath statistics.

What it means for the thesis chain

Every prior default-material CSI result is conditional on the drywall tag being correct: the sealed sessions c-csi-crowd-temporal/01KT99FX2Z3RG2W2DZ4FVCVTH7 and c-csi-cross-geometry-resplan/01KT98WKH2HYQYP4EP79G09CEG (and the unsealed occlusion/body-EM corpora) all ran on the untagged uniform-drywall default. Because the perturbation tested here is a full palette swap, this session quantifies sensitivity to the palette, not the error of plausible mis-tagging — drywall is the plausible tag for ResPlan interior partitions, so those results are not invalidated, but their fidelity now rests on an unverified import-side assumption. Escalation per the brief: (1) per-wall material tagging at import becomes a real requirement, not nice-to-have; (2) the furniture-scatterer extension (obstacle cells → extruded meshes) rises in priority since material fidelity demonstrably moves second-order statistics; (3) exp-csi-calibration against real hardware remains the only way to fix the absolute palette. Cross-geometry caveat: the ×3.8–×7.7 LOFO drift was measured with the SAME palette on all six floors, so palette uncertainty does not confound the relative drift result — it bounds absolute magnitudes only.

Attached runs

Run Gate Purpose Replay
43NX39VF
J87H22AQ
6K2K9R21
CEZV9AEF
7C754GKA
BNBFRBAV
C5YWZ6R6
RY1AT16E