Live status · Read the latest
monad-knowledge Wi-Fi sensing lab · FIIT STU
designed EXP-F4

EXP-F4 Propagation Surrogate Field Validation

In plain words

Does our radio model of the library predict the signal strength the real nodes measure?

rx1 rx2 rx3 rx4 rx5 rx6 rx7 rx8 rx9 predicted by the modelmeasured by the fleet received power nine receivers, one band at a time
The radio model predicts how strong the signal should be at each receiver. The design measures it with the real fleet in the empty hall and lays the two side by side, after allowing one constant offset per band. Schematic of the design, not a measurement.
Why it matters
A surrogate checked against its own simulator still needs an independent measured spatial comparison.
How it is done
Transmit from monad01 on channel 6 and then channel 36 in the empty hall for at least ten minutes per band while the nine receivers record received power. Compare with the model after one fitted offset per band: the error, the agreement of the spatial ranking, and whether wall counts explain what is left.
Where it stands
The August 3 preregistration remains untested. Its geometry and transmitter settings need matching to a dated as-built survey before acquisition.

FoundNo measured validation data are attributed.

StandingDesigned, not yet run

EXP-F4 Propagation Surrogate Field Validation

Current assessment (2026-09-15)

No measured validation data are attributed. The August 3 predicted RSSI and analysis hash remain the original preregistration. The old transmitter coordinate and 1.8 m height must not be silently combined with later surveys or a changed antenna height.

A fitted global offset can remove a shared calibration error, but nine receivers sharing one source are not nine independent days. A wall-count slope interval that includes zero does not demonstrate a negligible residual effect. Fitting five material losses to nine nodes would need held-out validation before a transfer claim. Publication suitability cannot be inferred from passing this pilot.

Next evidence needed: Version any geometry or analysis amendment before new acquisition and retain the original prediction hash. Report offset fitting and spatial assessment separately, account for shared source/day effects, and reserve independent measurements for any material refit.

Basis: Thesis Narrative, the September 10 synthesis, and the September 15 register review. See the review record for evidence sources and outstanding checks.

Earlier protocol and dated results

The methods and observations below retain their original scope. Read their readiness statements and proposed next steps as historical. The current assessment above records what has changed; frozen thresholds have not been retrospectively revised.

Validate the RT-distilled multi-wall surrogate (and the Sionna radio maps behind it, notebooks/campaigns/python/propagation_atlas.py) against measured received power from the deployed AX210 fleet (DEPLOY-001 AX210 Fleet Deployment FIIT Library). The 2026-08-02 atlas validated frequency scaling only within the ray tracer's own physics; this experiment supplies the missing measured anchor. Without it, publication ceiling is a workshop paper; with it, an AWPL/OJ-AP letter becomes arguable.

Hypothesis (declared before data)

The surrogate predicts the spatial pattern of measured mean RSSI across the 9 fixed receivers to within σ ≈ 5–8 dB after a single global offset fit (unknown antenna gains + cable losses absorbed into one constant per band). Falsifier: rank correlation between predicted and measured per-node RSSI below ρ = 0.6, or per-band residuals structured by wall count (which would indicate the fitted wall losses are wrong, not just offset).

Design

  • TX: monad01 at its mounted position (16.94, 12.42, 1.8 m), 20 dBm, ch 6 (2.437) then ch 36 (5.18). Optionally the monad-app phone as a paced illuminator on the atlas walk path (QR-anchored ground truth per the instrument audit).
  • RX: monad02–10 (csid captures; RSSI per packet from the radiotap/CSI header; sidecar sessions to S3 per IP-120 convention).
  • Protocol: empty hall (early morning, per EXP-F2 CSI Drift and Periodic BLE Calibration windows); ≥10 min per band; report per-node median RSSI + bootstrap CI (unit = 1 s windows to respect burstiness; cf. multicast burstiness finding in the first-captures evaluation).
  • Comparison: per band, fit one offset c: minimize Σ(RSSI_meas − (pred + c))²; report (a) offset-removed RMSE across 9 nodes, (b) Spearman ρ of spatial ranking, (c) residual-vs-wall-count regression slope (should be ≈ 0 if L_concrete/L_drywall are right).
  • Statistics: n = 9 nodes is small — report exact permutation p for ρ, and label CIs per the small-n honesty rules (statistics skill §2). This is a pattern-validation pilot, not a coverage claim.

What each outcome means

  • Pattern matches, residual flat in wall count → surrogate usable as the planning forward model; proceed to write-up.
  • Offset large but pattern matches → antenna/cal constants; fine, expected.
  • Residual slopes with wall count → per-material losses mis-fitted (materials provenance, IP-092 default labels) → refit L on measured data (5 unknowns, 9 equations — feasible), which itself becomes the paper's calibration-transfer result.
  • Pattern fails → the scene geometry/material pipeline is wrong in kind; stop and debug with EXP-008 Active Injection vs Passive Sniffing-style controlled links.

Pre-registration (2026-08-03, before any measurement)

Per-node predicted RSSI (both bands, offset-free form) + conformal q90 intervals + the exact analysis plan are frozen in exp_f4_preregistration.json, SHA-256 d80dd13fd9e9b052e969db0227185a6ea0dd9e76d81c11c069705928cfab3e55, stored at s3://…/benchmarks/propagation-atlas/dt=2026-08-03/exp_f4_preregistration.json. PASS criteria declared: Spearman ρ ≥ 0.6 and offset-removed RMSE ≤ 8 dB; residual-vs-wall-count slope CI must cover 0. Analysis after measurement uses this file, not refitted numbers.

Dependencies

Fleet thermal qualification (monad02 throttling — see the 2026-07-28 diary), csid capture sessions healthy, both bands enabled in the AX210 profile (5 GHz AP-crash caveat from the illuminator bring-up applies to AP mode, not monitor capture).

Provenance

—
not recorded