01KW0N0JSVNXR6078XMCYXT4SN

Occupancy-matched A/B — spatial-pattern fingerprint vs headcount confound

Why. The first session found reactive≠scripted CSI (per-link variance KS D=0.40, p=4.2×10⁻⁴), but the dominant separator was co-present headcount (reactive FSMs loop → mean N 14.3 vs scripted 10.1). A counting model keyed on N would absorb that. This session holds headcount fixed and asks whether the spatial arrangement alone leaves a trace. Analysis-only over the same 10 runs (10,380 frame×link observations); no new compute — stratification matches the full occupancy distribution, not just the peak.

Method. Overlap window N∈[3,14]. (1) Conditional attenuation slope: fit amp ~ N per arm within the window. (2) Residual KS: pooled amp = f(N) quadratic fit, KS the residuals arm-vs-arm. (3) Matched-subsample KS: resample both arms to a common integer-N histogram (matched mean N = 9.5 both arms), recompute per-link amplitude variance, redo KS. (4) Mechanism: living-0 occupancy share at matched N. CIs/KS via the IP-108 reduction_stats primitives.

Result — the fingerprint is mostly headcount, with a faint spatial residue.

• Matched-subsample variance KS: D=0.26, p=0.056 — the headline §1 fingerprint (D=0.40, p=4×10⁻⁴) falls below significance once occupancy is matched. Most of the original separation was simply "more bodies present".
• A small genuine spatial residue survives. At equal N the reactive links sit ~2.7 dB lower on average (residual means −1.96 vs +0.76 dB; residual KS D=0.08, p=4×10⁻⁸ — significant but trivial effect over the large pooled, non-independent sample). Conditional attenuation slope: reactive −0.39 vs scripted +0.06 dB/person — adding a guest attenuates more under the reactive arrangement.
• Mechanism confirmed. At the same headcount, the reactive crowd places 47% of present guests in living-0 vs scripted's 59% — the room_full(living-0) guards genuinely disperse it into the bedrooms. The spatial difference is real; it registers only faintly on these sensors because the dispersed bodies still mostly miss the bedroom links (which, per the sibling c-flat-day-csi paradox, are sparse-but-legible while living-0 is saturated).

Refined claim (supersedes §1's framing). Behaviour realism does change the sensed signal, but at matched headcount the effect on the per-link variance distribution is small and borderline (D≈0.26, p≈0.06). The actionable sim-to-real risk is therefore predominantly that reactive and scripted crowds differ in how many people are co-present, not in a strong spatial-arrangement signature. A counting model robust to total-occupancy distribution shift would absorb most of the fingerprint; a residual ~2–3 dB arrangement effect remains.

Caveats. Residual KS p is optimistic (links×seeds not independent; effective n ≪ pooled n) — the matched-variance KS on per-(link,seed) variances (D=0.26, p=0.056) is the honest test and it is null/borderline. Overlap window trims the tails where the arms differ most, by construction. Quasi-static RT, single evening, derived (not independent) scripted arm.</synthesis_md> ["Hold headcount fixed and test whether spatial arrangement alone leaves a CSI fingerprint", "Quantify how much of the session-1 fingerprint was headcount vs spatial", "Confirm the spatial mechanism (reactive disperses the crowd at matched N)"]

Self-critique

• Stratification trims the discriminative tails. Restricting to the overlap window N∈[3,14] removes exactly the high-N reactive frames where the arms differ most — so the matched test is conservative by construction. The honest reading is "at occupancies both crowds reach, the spatial signature is weak", not "there is no signature anywhere".
• Residual KS is over-powered, matched-variance KS is the trustworthy one. The residual test treats 10k correlated rows as independent → p=4e-8 is meaningless as a significance statement; report it only as a direction + effect size (−2.7 dB). The per-(link,seed) variance KS (D=0.26, p=0.056) respects the true unit and is the claim to cite.
• A true matched-headcount run-set (cap reactive co-presence by construction, e.g. inject time-bounded exit states) would beat analysis-side stratification by also matching the temporal occupancy profile, not just the marginal. Deferred — the analysis-side control already shows the effect is small.
• Derived scripted arm + single evening + quasi-static RT, as before.