m monad-knowledge FIIT STU · open research vault
Campaign

Wall-material fidelity — does the ITU wall palette shift second-order CSI statistics (drywall vs concrete sensitivity probe)?

c-csi-fidelity-material · exp-csi-static

Sessions

No sessions yet — this campaign has a brief but no execution.

Brief

Question

Does the wall material palette measurably shift CSI statistics — drywall (as stored) vs concrete (forced override) on identical geometry and placements — or is the material a fidelity knob the campaigns can safely ignore?

Discovered while authoring this brief: the 12439 staged scene palette is already uniform drywall — the import does not tag per-wall materials, so the original heterogeneous-vs-uniform framing is counterfactual on every current floor. That makes palette sensitivity the testable question, and per-wall material tagging an import-side follow-on if this probe shows sensitivity.

A null result licenses cheap uniform scenes for every material-insensitive campaign; a hit forces a fidelity caveat onto every prior CSI result that ran on default materials. Either answer changes how the platform is operated — that is what makes a methodological arm worth one CPU-hour.

What we already know

  • The scene bridge resolves each cell_space_boundaries.material to an ITU-R P.2040 RadioMaterial (vault names: drywall/concrete/brick/glass/...; MATERIAL_TABLE in monad_knowledge/sim/scenegeom.py); the 12439 scene stages walls as drywall only.
  • Mean amplitude alone cannot answer this — material redistribution mostly moves multipath (delay structure), so the discriminating statistics are second-order: the runner emits delay_spread_ns per (frame, link) since 2026-06-04 exactly for this arm.
  • The override is scene_overrides.uniform_material (added 2026-06-04): the runner forces every wall to the named material and fails loudly if it is not in the staged palette.
  • c-csi-crowd-occlusion / c-csi-through-wall-blockage / c-csi-cross-geometry-resplan all ran on the as-stored palette — they are the results a positive KS outcome would caveat.

Prerequisites

  1. The 2026-06-04 sionna-csi-runner image (uniform-material override + delay_spread_ns).
  2. Nothing else — same floor, same link layout, same seeds as the through-wall rig.

What the supervisor does

Run the 8-cell grid (2 arms × 2 bands × 2 occupancy, seed 0, 64 placements). The reduction is a small KS-test pass over the paired runs' links.parquet (paired by band × occupancy; ECDFs over per-placement mean_amp_db and delay_spread_ns, split by link class where the layout allows); the analysis-writer reports the KS-D table and applies criterion 3's decision rule verbatim.

Figure render request

csi_material_ks — ECDF small-multiples (heterogeneous vs uniform, per band × occupancy) for both statistics + a KS-D summary table panel.

Out of scope

  • Furniture scatterers — obstacle cells are not yet extruded into the scene (the named follow-on extension); this arm tests the wall palette only.
  • Material calibration against real data (exp-csi-calibration).
  • The geometry-continuum claim — gated on real data regardless of this outcome.

Expected interpretation

  1. Null (KS-D small everywhere) → uniform scenes licensed; staging simplified; note added to the simulator's curated note.
  2. Hit on second-order only → prior campaigns' mean-amplitude conclusions stand, but variance/delay-sensitive claims (Rician-K slopes, CV bundles) acquire a material caveat; name the affected sessions explicitly.
  3. Hit on first-order too → palette fidelity is load-bearing for everything; escalate to a furniture-scatterer + material-audit follow-on before further CSI campaigns.

Notes

  • scene_overrides.uniform_material: null in the grid means "as stored" (no override) — the resolver passes the null through and the runner skips the override branch.
  • The override material name is the VAULT-level name (concrete, not itu_concrete) and must be in the staged palette or the runner exits loudly. NOTE: the staged materials list only carries palette entries actually referenced by walls — the runner-side override therefore needs the override material injected into the palette at staging, or the supervisor passes a material the bridge already stages. Verify on the first run; a loud failure here is a scene-bridge finding.