c-csi-topology-sniffer-count — session 01KVXT9B31TF088GH5TE539QZN (clean re-seal, 2026-06-24)
This session supersedes 01KVXQTTMGCKCGFVB3TN2G8WXD, which sealed with runs_attached: 0: the CLI sim run --label session_id=… does not auto-attach (only the MCP sim_run_launch happy-path does), and the explicit attach-run loop was skipped. The 12 source runs are now properly attached here. The scientific conclusion is unchanged.
Verdict: layout-topology-fingerprint-discriminability stays plausible — mixed, underpowered; the result is robust to a physical noise-floor clip. Surfaced by an adversarial rigor audit against the vault literature.
Two corrections folded in (audit-driven)
- Run attachment (integrity). Prior session's "12 runs attached" was false (manifest showed 0). Fixed: 12 runs attached here; the local-dispatch pattern now requires an explicit
sim_campaign_attach_runloop after CLI fan-out. - Degenerate-scalar robustness. Three runs emit
mean_amp_db = −240 dB(the ray tracer's "no path found" sentinel for fully-blocked single-body placements), which drives a physically-impossible per-runmean_atten_db_per_person = 173 dB/person(free-space across a room is 60–80 dB). That scalar is the runner's own occupancy-slope fit — it is not used by this reduction, which works in linear amplitude where the sentinel → ~0 (bounded, not high-leverage). A defensive physical noise-floor clip (−120 dB) was added to_fingerprint_matrixand is confirmed a no-op on every conclusion (rho_D, rho_kfrac, defeater all identical), proving the topology result was not corrupted by the sentinels.
Result (unchanged, noise-floor-clipped)
- Per-link discriminability D weakly favours topology over area (Spearman ρ: mean_degree +0.37, betweenness +0.26, diameter +0.19, area +0.14); weak at n=6, none significant.
- Required sniffer fraction k/n_links* best predicted by area (ρ = −0.74), beating diameter (−0.63) — the
adeel2019_ada0AP-count baseline holds against the claim's operational form. - Diameter-stratified defeater not triggered (Cohen's d = −1.44) but underpowered (3 floors per stratum, three of six tied at diameter 3); area and diameter confounded (largest floor = highest diameter).
- BLE location-fingerprints ~2× more separable than CSI's on every floor.
Honest scope (sharpened by the audit)
This is simulation-derived experimental design, not externally-validated evidence: no real measurement, single seed effectively (the placement RNG), self-authored body/BLE priors, uniform-material. Against the academic bar (choi2022_17c2 LOSO over real rooms; santos2024_1e39 7 real environments) a single-floor synthetic Spearman is exploratory, not confirmatory. The hypothesis's strength stays plausible; the next step is the area-decorrelated, link-as-unit (wall-count-on-path) design + a small real-layout check. A separate runner bug (degenerate occupancy slope at n_agents=1) is logged for a platform-level sanity-bound validator.