c-csi-cross-geometry-scaleout — Phase 1 pilot, session 01KWYBEJQK4307TTP9KMRY0DEY (2026-07-07)
Verdict: the Phase-1 pilot validates the machinery and settles the variance split — geometry dominates seed decisively in the CV(N) sensing signal — but the powered primary criterion is not met, and at 6 floors the inflation-ratio CI is too fragile to reason over. LOFO cross-floor inflation is a pooled ×3.37 (honest leave-one-occupancy-out within-floor baseline 0.026 vs held-out 0.088), floor-cluster 95% CI [1.04, 9.85] — but with only 6 bootstrap clusters that interval has erratic coverage; its bare exclusion of 1.0 is not trustworthy evidence and is not leaned on. Criteria 2/3: C2 (variance split) and C3 (honest baseline) pass; C1 (inflation CI lower bound > 1.5) fails on width.
What we ran
6 ResPlan floors × 3 seeds × 5 occupancy levels {0,2,4,6,8} at 2.4 GHz, 48 Monte-Carlo body placements per cell — 90 exp-csi-static runs, all gate-passed, local Sionna, ble disabled (CSI-only estimand). Scope cap disclosed up front: the brief's Phase-1 names 10 floors, but only 6 are staged in PostGIS (resplan-{12439,1374,147440,7421,16157,12419}; the other four — 10156/10425/6090/8266 — have no experiment). n₀ = z²σ²/d² ≈ 40 floors for the powered primary, so 6 floors is a machinery-and-variance pilot, not the powered test. Reduction csi_cross_geometry_scaleout.py (unit = floor).
What we found
- Geometry dominates seed — decisively (C2 ✓). Nested one-way random-effects decomposition on the per-(floor,seed) mean-CV signature: σ²_floor = 0.0109 vs σ²_seed = 0.000165 — floor variance is 66× the seed variance. Between-floor SD 0.105 vs within-floor between-seed SD 0.0116 (9×). The CV(N) curve is a stable floor signature; placement-seed jitter is small against the geometry spread. This is the pilot's real deliverable: the cross-floor drift the thesis chain leans on is a genuine geometry effect, not seed noise.
- Honest baseline changes the comparator, not the direction (C3 ✓). The original
csi_cross_geometry.pyused an in-sample fit residual as the "within-floor" MAE (the audit's flagged bug). This reduction uses leave-one-occupancy-out within each floor (0.026) as the denominator — a generalization error, never the training residual. LOFO held-out MAE 0.088 → inflation ×3.37. - Primary CI is wide and its width is the finding (C1 ✗, by design). Floor-cluster bootstrap (unit = floor, 6 clusters) gives inflation 95% CI [1.04, 9.85]. The pooled ×3.37 hides extreme per-floor heterogeneity — the per-floor held-out/within ratios span 0.70× (resplan-7421, where held-out is actually better than within — no inflation) to 49.8× (resplan-1374, whose tiny within-MAE 0.005 dominates the pooled numerator); mean-of-ratios is 10.1 vs ratio-of-means 3.37. That spread, not sampling luck, is why the CI is nearly an order of magnitude wide. At 6 clusters the bootstrap coverage is itself unreliable, so the interval is best read as "uninformative but consistent with large inflation," not as evidence the lower bound clears 1.0. Booked false, not rounded up.
What it means for the thesis chain
The CV(N) sensing signal is geometry-dominated (σ²_floor ≫ σ²_seed) — that variance-split result is solid at this n and is the pilot's real contribution. It is distinct from, and does not by itself establish, the MAE-inflation drift claim (C1 is on held-out MAE, C2 on CV(N) variance — related but not the same quantity). The honest inflation point estimate is ×3.37, but it is a heterogeneous mixture (0.70×–49.8× across floors) and the number the thesis leans on cannot carry a > 1.5 lower-bound claim until Phase 2 stages ≥40 floors (the ResPlan converter → gis_rebuild_nrg → gis_validate batch). No hypothesis strength change: this is a design input that quantifies how many distinct geometries the IP-106 field study needs, and the answer is "many — seed replication is cheap, floor replication is what buys the CI." Phase 2 is the powered follow-on.
Honest scope
In-silico, ResPlan geometries, single band (2.4 GHz), CSI-only (ble disabled). 6 of 10 planned floors, 3 seeds each. The variance-split and honest-baseline results are solid at this n; the inflation-ratio CI is deliberately under-powered and labelled so. Reduction + cross_geometry.parquet (per floor×seed×N) + per-floor CSV + 2 figures (csi_cross_geometry small-multiples with the CI panel, seed_trace between-floor vs between-seed spread) under this session's artefacts/ prefix; 90 runs attached.