Why this campaign exists
c-csi-ble-fusion supplies the thesis's core "staleness-switch = the fusion architecture" claim, but the audit found both its sessions sealed with zero attached figures, and session 01KW09NZ… issued an unhedged POSITIVE from 3 runs / 1 floor / ~380 samples, R2=0.36 (64% variance unexplained), in a regime where its own MLP arm collapses to R2=-0.43. Our own real-data papers (renet-csi-calibration: "magnitude needs ~40 geometries"; csi-recalibration-plane: sim is self-confirming) would not permit that verdict.
This re-run powers the claim (>=8 seeds, >=3 geometries), forbids CI-less point-estimate headlines, renders + uploads the paired-difference figure, and gates the POSITIVE on a public real-CSI cross-check (meneghello / OPERAnet) — the campaign-rigor audit's top recommendation. The mechanism may well be real; it must be earned, not asserted.
How this executes (two phases + real-data gate)
Phase A — corpus. Fan out coupled exp-csi-crowd runs, ≥8 seeds × ≥3 ResPlan floors, matched through-traffic occupancy. where=local. Runs attach to the session.
Phase B — the fusion reduction (standalone, already critic-hardened). monad_knowledge/notebooks/python/fusion_count_v2.py (BLE as an independent device-counter — no leakage; CSI-only-transferred / BLE-only / Fused arms; paired MAE mean±std over seeds) consumes the Phase-A run dirs. Run it standalone (container can't read Hetzner s3://); it emits per-seed scalars → reduce the paired count-error difference (no-recal minus periodic-BLE-recal) with sim_reduction_run(multi_seed_ci) for the CI. Render + attach the paired-difference figure before seal.
Real-data gate (the POSITIVE lock). The standing anchor is c-sim2real-count-transfer (WiMANS leave-one-environment-out; its csi_sim2real_count_transfer.py reduction is selftest-validated). A POSITIVE here requires sign-agreement with that real anchor — do not seal POSITIVE on sim alone.
Adversarial-review notes (2026-07-15): the estimand is deliberately pivoted from the prior's R² to the paired count-error difference (no-recal minus periodic-BLE-recal) — this tests the actual staleness-switch drift-bounding mechanism, not raw CSI regressability. Do NOT quietly carry the old R²=0.36 forward as if re-earned; it is a different quantity. The pilot_sigma: 0.15 is inherited from the very pilot this brief calls underpowered — treat it as provisional and re-estimate σ from the first real batch before trusting the power sizing; the effect may still land inside noise at 8 seeds if σ was underestimated. "Sign-agreement" on the real anchor is a floor, not proof of transfer (OPERAnet/meneghello carry their own environment/hardware confounds).
Prior: c-csi-ble-fusion (underpowered pilot — direction only, do not cite its POSITIVE).