CSI sensability across public-building geometry at scale — the ArchCAD ensemble the residential set couldn't cover

Why this campaign exists

Everything we know about which rooms are hard to sense in comes from apartments — the only geometry our floorplan corpus could load. The residential cross-geometry study (csi-cross-geometry-generalization) established the mechanism (occupancy-sensitivity is geometry-bound, σ²_floor ≈ 66× σ²_seed on N=6) but its powered ≥40-geometry confirmation was never executed — its latest_session was a phantom (NoSuchKey in S3), and the honest pilot's floor-clustered CI [1.04, 9.85] failed its own gate. Meanwhile the buildings that actually matter for crowd sensing — offices, halls, concourses, stair cores, parking decks — were entirely absent.

archcad-400k changes that: 41,097 real drafted public-building floor slices, primitive-level walls/glass/columns already in the archcad PostGIS schema, and a validated extractor (09_archcad_sionna_scene.py) that renders each into a Sionna-RT scene producing physically-real CSI (verified end-to-end: path loss −40…−80 dB, Rician-K 6…42 dB, delay spread 26…113 ns on a stair-core tile). This campaign uses that to do two things the residential set could not:

1. Replicate the geometry≫seed variance split at scale (≥40 geometries) on an independent, non-residential substrate — the powered run that never happened, now on richer geometry.
2. Rank public-building archetypes by sensability (open-hall vs partitioned vs column-forest vs stair/parking) — a first "which building types are intrinsically good/bad for CSI crowd sensing" atlas.

This is a replication-at-scale + a new archetype ranking, not a discovery: geometry-dominance is already known. The value is independence (a wholly different geometry source), scale (≥40, powered), and the public-building question the thesis actually needs.

The one enabling step (Step 0)

exp-csi-static normally builds scene.json from a GIS floor; ArchCAD tiles are not GIS floors. The campaign's Step 0 is to stage the archcad-derived scene.json into each run's inputs/ (the runner already reads --scene/inputs/scene.json), driven by 09_archcad_sionna_scene.py. Two design constraints, both first-class success criteria:

• Room-aware Tx/Rx placement. ArchCAD tiles are multi-room; a whole-tile Tx-west/Rx-east split gets 0 paths when a full-height party/shaft wall severs the link (observed in the prototype smoke). Confine Tx + the Rx cluster to one wall-bounded open region (flood-fill over the wall raster). Log the per-tile signal-fraction and exclude trivially-severed runs from H-A rather than averaging them in.
• Archetype stratification. Draw the ≥40 tiles across structural archetypes using archcad.tiles flags (has_stairs/elevator/parking) + wall-density quartiles + an openness proxy, so H-B has balanced groups and H-A's between-tile term spans real diversity, not 40 near-identical office slices.

Honest scope (stamped on the synthesis)

Sim-only, drafted CAD geometry: thin dielectric-sheet walls, no furniture, uncalibrated 5 dB body-loss prior, unvalidated ITU material coefficients, ~0.014 m/unit empirical scale. Direction and ranking only — no magnitude travels to the thesis without the standing real-data anchor (archcad-400k is drafted geometry; the real cross-check is WiMANS / meneghello, per csi-cross-geometry-generalization's gate). Design mirrors c-csi-cross-geometry-scaleout-v2 and inherits its honesty fixes: tile-bootstrap CI (heavy-tailed ratio), every named figure present in S3, no headline statistic living only in prose/parquet.