Campaign

Graded BLE-anchor occupancy, at power — confirm or refute the hub-placement slope with a CI that can actually exclude the null

c-ble-graded-count-powered · exp-csi-static

Archive snapshot, as of 5 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
7J8TJQP6 2026-07-06T23:33	finished	96	✓	✓	—	REFUTED at n=8: the pilot's graded-slope claim does not survive powering. Paired (topology−dispersed) occupancy-response slope diff −0.126 dB/person (95% seed bootstrap CI [−0.412,+0.113], exact sign-flip p=0.461, d=−0.32) — CI includes 0, magnitude LB 0.0 < 0.30; the pilot's −0.498 (n=3) is below the powered CI lower bound. Graded grid 0/20 BH-pass, ΔN=2 topology 0/3. Both placements carry a negative response at power (dispersed does NOT saturate as the pilot claimed); topology's fitted slope is more consistent (8/8 negative, but p=0.0078 is the attainable floor, uncorrected post-hoc) — the between-arm gap is within seed noise. 96/96 runs gate-passed; no sentinel artefacts. ble-periodic-calibration held plausible, no strength change. Deliverable for IP-106: concentrate anchors for response consistency, not fine graded count.

Brief

Why this campaign exists

c-ble-anchor-placement's session 01KVZS1K… reported "coarse resolvability recovered" from a paired slope difference of −0.498 dB/person — but on n=3 seeds, where the exact sign-test floor is p=0.25 and the effect sizes were computed over ~360 pooled events whose honest replication count was 3. The 2026-07-06 statistical overhaul's audit derived the powered design from the observed variance: σ_d=0.379, δ=0.498 → n≈5 for 80% power, n≈7 for 90%. This campaign re-runs the same design at n=8 seeds so the graded-resolvability claim is either confirmed with a CI that excludes the null, or cleanly refuted.

Design

Two arms — dispersed-4 vs hub-4 anchor layout — paired by seed: each seed's crowd trajectory runs through both layouts. Occupancy grid {0, 2, 4, 6, 8, 12} and noise_db=2.0 held identical to the original session (so pilot_sigma is comparable); same floor and placements. ~8 seeds × 2 layouts × 6 N-levels = 96 runs.

Artefact contract (hard gate, supervisor Step 6.7)

ble_graded.parquet (per seed × layout × N cell), the per-cell BH-FDR table, metrics.json with the paired slope difference + seed-block-bootstrap CI. Reduction must use the cluster bootstrap at the seed level, never pooled events.

Figure render request (mandatory)

seed_trace — RSSI vs N, every seed a light trace, mean bold, one panel per layout.
paired_diff — per-seed (dispersed − hub) slope-difference dots + CI on the difference.
perm_null — paired-permutation null histogram for the primary endpoint, observed statistic marked.

Honest scope

One floor, one occlusion geometry, synthetic BLE noise model. This confirms the mechanism at power; the field claim still routes to the IP-106 hardware protocol. Moves ble-periodic-calibration-adjacent hypotheses at most candidate→chosen, never beyond plausible.