Campaign

c-flat-day-reactive-vs-scripted — does the behaviour model leave a sensing fingerprint?

c-flat-day-reactive-vs-scripted · walk-notebook

Archive snapshot, as of 4 h 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.

Sessions

Session	State	Runs	Synthesis	Criticism	Figures	Verdict
2EJS49XP 2026-06-26T00:43	finished	0	✓	✓	2	Residue resolved at 20 seeds/arm: the matched-headcount fingerprint is a tiny, borderline effect (KS D=0.135, p=0.0475) that only crosses significance because n grew 4x — and is partly a residual ~1-person gap. The spatial-behaviour fingerprint is negligible next to headcount.
Q762A3KE 2026-06-26T00:36	finished	10	✓	✓	2	True matched-headcount run-set confirms the stratified result: with co-presence equalised by construction (time-capped FSM exits, peak 16.4 vs 15.4), the per-link CSI fingerprint is not distinguishable (KS D=0.24, p=0.095, down from D=0.40 p=4e-4). The §1 fingerprint was headcount.
MCYXT4SN 2026-06-26T00:24	finished	10	✓	✓	2	Occupancy-matched: the §1 fingerprint was mostly headcount (variance KS D=0.40→0.26, p 4e-4→0.056). A faint spatial residue survives (~2.7 dB at equal N); reactive genuinely disperses the crowd (47% vs 59% in living-0 at matched N).
T6Q002MC 2026-06-25T23:51	finished	0	✓	✓	2	Behaviour leaves a measurable CSI fingerprint (KS D=0.40, p=4.2e-4) — a real train/test mismatch risk for a counting model.

Brief

c-flat-day-reactive-vs-scripted — does the behaviour model leave a sensing fingerprint?

An A/B on the same floor + same 10 receivers, changing only the strategic layer:

Arm A — scripted linear agendas: the v1 c-flat-day-ai cast (each persona a fixed itinerary; mills when a room is full).
Arm B — reactive FSMs: the IP-107 Phase-5 cast (guards divert guests when room_full(living-0), escape the kitchen on count_in > 4, etc.).

Question

Reactive behaviour visibly redistributes the crowd (B spreads into all four bedrooms; living-0 peak 16 vs A's 18). Does that redistribution change the sensed signal — the per-link CSI variance and RSSI distributions — enough to matter for a crowd-counting model trained on one arm and tested on the other? If a counting model is sensitive to the behavioural assumptions baked into the training crowd, that is a real sim-to-real risk worth quantifying.

Plan

Run both arms (5 seeds each) — crowd leg locally (uv run --with jupedsim), the deterministic FSM/agenda configs from experiments/ip107-flat-day-ai/.
Couple each to CSI (exp-csi-crowd, CI) → per-link statistics.
sim_reduction_run over the 10 runs: per-room dwell deltas + per-link CSI-variance distributions, with CIs; test arm separability.

This closes the loop between IP-107 (behaviour realism) and the thesis sensing question: behaviour realism is only worth the effort if it changes what the sensor sees.