Calibration-as-control — does a BLE anchor rescue a drifting CSI occupancy filter (sim)?

Question

Static per-window CSI counting saturates at a hard information ceiling (PUB-1), and CSI models drift over time (brunello2025). We showed (a) a birth-death temporal filter beats the per-window ceiling by 60–78% on real CSI, and (b) calibration-as-control — a periodic BLE anchor that not only resets the count but re-identifies the drifted CSI→count emission — recovers 90% of drift-induced error vs 20% for a naive reset. But that 90% needed large drift; real WiMANS cross-environment drift is too small to show it (a plain BLE reset already suffices there).

This campaign supplies the missing regime: a controlled large-drift CSI-counting stream with a co-located BLE anchor, so we can demonstrate that emission re-identification (Bu_t) beats a plain reset in the regime where it matters — the load-bearing experiment for PUB-9 (publications/csi-calibration-control/PROPOSAL.md).

Design (the sim provides ground-truth drift + BLE)

A parameter sweep over a real floor (resplan-12439-floor-0, 5 GHz — matching WiMANS) provides the per-cell CSI + BLE; a reduction then streams them through the calibration-as-control filter.

• Occupancy axis scenario.n_agents ∈ {0,…,5} — the count classes (match WiMANS).
• Drift axis scene_overrides.device_jitter_m ∈ {0.0, 0.5, 1.0, 1.5, 2.0} — with a fixed seed this displaces the Tx/Rx geometry monotonically, so increasing jitter = a progressive environmental drift of the CSI→count emission (the refit lever hierarchy: device-jitter is the fine drift knob).
• occupancy_diffuse ON (IP-118) so simulated CSI carries the realistic count→variance response.
• BLE ON (ble.enabled) → ble_links.parquet co-located with the CSI (same solve) = the anchor.
• Seeds {0,1} for a small replication. Grid = 6 × 5 × 2 = 60 cells; each emits csi.hdf5 + ble_links.parquet.

Configs: sweeps/cac-drift.base.yaml + sweeps/cac-drift.grid.yaml. Reduction: monad_knowledge/notebooks/python/cac_drift_reduce.py.

Method (reduction)

1. Per cell: read csi.hdf5, compute the per-window CSI amplitude-variance feature Φ (the same feature reduced for WiMANS/OPERAnet); the run's n_agents is the ground-truth count; the BLE anchor count at calibration times = n_agents (a reliable presence/count anchor).
2. Fit the emission p(Φ|count) at drift level 0 (the calibration set at deployment time).
3. Assemble a stream: a birth-death count path + a monotone drift schedule (drift index rising over the stream); at each step sample a window from the cell matching (count_t, drift_level_t).
4. Run the birth-death HMM filter under four arms — no-cal (stale emission), reset-only (BLE state reset), calibration-as-control (reset + emission re-ID), oracle (current-drift emission) — and report count MAE + the accuracy-vs-recalibration-cadence curve.

Success criteria

• The sim reproduces meaningful drift: no-cal MAE rises with drift level (a defined degradation).
• Calibration-as-control beats reset-only once drift is large (the PUB-9 core claim, on ground-truth sim drift + a co-located BLE anchor) — with the crossover pinned to a drift magnitude.
• The cadence curve gives a deployable BLE-anchoring interval for a target error.
• Honest scope: sim drift (device-jitter) is a controllable proxy; the hardware campaign supplies real weeks-scale drift.

Dispatch

Corpus generation is deterministic (no LLM supervisor). Build the image once, then:

uv run monad-knowledge sim build --sim-id exp-csi-crowd
export MONAD_SIM_SIMULATORS="$(uv run monad-knowledge sim build --sim-id exp-csi-crowd \
  --emit-env 2>/dev/null | sed -n 's/^export MONAD_SIM_SIMULATORS=//p')"
uv run monad-knowledge sim sweep exp-csi-crowd \
  --grid monad_knowledge/sim/exp_csi_crowd/sweeps/cac-drift.grid.yaml --where local
# then: uv run python monad_knowledge/notebooks/python/cac_drift_reduce.py <figdata_dir>

Material/temperature drift — was INERT under IP-118, RESOLVED by IP-119

The first material-drift check (concrete vs metal walls, scene_overrides.uniform_material, IP-118 overlay ON) came back byte-identical (log-var 1.1177, |H| 8.18) — material completely inert. Cause: occupancy_diffuse's absolute power (p_d0=35, gamma=4000) produced a diffuse floor amplitude ≈5.9 while the ray-traced channel is |H_det|≈1e-4 (≈80 dB below) — the overlay swamped the physics, so the "sim CSI" was essentially the scene-independent synthetic overlay and structural drift was negligible. Only post-overlay drift (SNR AWGN) moved the emission — why the SNR-drift test worked and the material test did not.

IP-119 fixed this by reparameterising the overlay to channel-relative Rician K: diffuse_power_ℓ(N) = |H_det,ℓ|² / K(N) with K_dB(N) = k0_db − k_slope·N (dimensionless K, no pdim dilution; a bounded fraction of each link's ray-traced power, never swamping it). The diffuse now inherits the scene physics, so wall material propagates into the count→variance response. Verified on Metal Sionna runs: concrete vs metal are no longer byte-identical (rel L2 0.49, mean|H| ratio 0.84); runner log augmented … det_power=2.76e-7 diffuse_power=1.74e-7 mean_K=1.58 (diffuse/det = 1/K exactly).

Phase 2 result — re-ID ≫ reset on GROUND-TRUTH structural (wall-material) drift

50-cell material-drift sweep (sweeps/cac-drift-material.grid.yaml: n_agents 1-5 × uniform_material {concrete, glass, brick, drywall, wood} × seed {0,1}, IP-119 overlay + BLE, Metal). Drift coordinate = |eps_r(material) − eps_r(concrete)| (concrete = deployment-calibration baseline). Reduction monad_knowledge/notebooks/python/cac_drift_material_reduce.py. Unlike the analytic SNR axis (cac-drift2), this drift is re-raytraced — a ground-truth structural emission shift no reduction can synthesise, only possible because IP-119 makes the diffuse track |H_det|².

The material moves the emission consistently across counts (mean log amp-var shift vs concrete: glass +0.22, brick +0.15, drywall +0.08, wood +0.12 — stable across all 5 counts, i.e. a whole-curve recalibration target). Count MAE (calib every 40 windows): no-cal 0.559, BLE reset-only 0.524 (7% recovery — stale emission mis-reads the drifted channel between resets), calibration-as-control 0.215 (74% recovery), oracle 0.096. Re-ID beats reset by +0.309 MAE, and CaC dominates reset-only at EVERY recalibration cadence (0.14 vs 0.46 at interval 15 → 0.36 vs 0.55 at 250). figdata/cac-drift-material/ in the publication. Same qualitative result as the SNR run (re-ID ≫ reset) but now on ray-traced physics, not overlay-dominated CSI — the material/temperature-drift case the IP-118 note said needed a channel-relative overlay. Honest scope: sim structural drift (one material palette, static per run) is a controllable proxy; the hardware campaign still supplies real weeks-scale drift.

Changelog