← RFC

Propagation Calibration Against Real Malla Data — Result

Status: Stage 2 result. Date: 2026-06-09. Tag: sim-result-v8. Reproduce: python3 tools/fetch_malla.py then cd sim && python3 -m experiments.propagation_calibration.

We found that malla.mtnme.sh (a Malla instance on the same North Georgia network, distinct from the Meshview view.mtnme.sh) exposes a public, no-auth API with per-gateway reception rows — snr, rssi, gateway_id, from_node_id, hop fields, payload_length (real wire bytes). This is the per-gateway RF data we had been waiting on the operator's DB export for. It lets us calibrate the sim's propagation model against real measurements rather than a literature assumption — directly addressing the v2 validation gap (north-georgia-validation.md).

Method

• Pulled ~20,000 reception rows from two independent Malla instances — malla.mtnme.sh and malla.gamesh.net — each joined to node positions from its own Meshview instance. The two networks cover the same metro (centroids 33.89,−84.25 vs 33.95,−84.23) but use largely different gateways, so each is an independent vantage onto the same RF environment.
• Critical filter — direct receptions only (obs_hops == 0). For a multi-hop packet, the gateway heard a relay, not the originator, so from_node_id → gateway distance is meaningless; only hop-0 rows give a valid originator→gateway RF link. This reduced ~20k → 1,125 valid points.
• Fit rssi = A + B·log10(d) → path-loss exponent n = -B/10; residual std → shadowing σ.

Result — cross-validated across two independent networks

Network (independent gateways)	direct points	exponent n	R²
malla.mtnme.sh	413	2.52	0.37
malla.gamesh.net	712	3.36	0.11
combined	1,125	2.50	0.26

The exponent is the headline, and it's good news: two independent networks (different gateways) both land in the physical outdoor-LoRa range and bracket the assumed 2.7 (2.52 and 3.36, combined 2.50). The PHY model's single most important propagation parameter is now data-confirmed against two real networks, not assumed. The spread (2.5–3.4) is itself realistic — the effective exponent depends on gateway siting (gamesh's lower/different gateways give steeper, noisier path loss).

What the data can't give (honestly): the intercept (PL@1 m → non-physical by naive extrapolation) and shadowing σ (a fit value ~24 dB) conflate path loss with per-node tx-power / antenna heterogeneity and survivorship bias (only decoded links are logged, so the low-RSSI tail is truncated). A passive third-party reception log cannot separate these. So we apply only the data-derived exponent and keep PL@1 m and σ at their literature values.

Effect on the v2 gate

Re-running the North Georgia validation gate with the data-derived exponent (2.50):

channel_utilization (%)	median	mean	p90	max
SIM, prior (n=2.7)	3.7	~3.5	5.0	5.6
SIM, data exponent (n=2.50)	7.8	7.4	8.0	8.0
REAL	5.7	7.0	14.8	37.3

The data-derived exponent brings the mean into agreement (7.4 vs 7.0) and the median to the same order (7.8 vs 5.7) — up from the prior ~1.5× under-prediction. The distribution is still narrow (no long tail), exactly as before.

What this changes (and what it doesn't)

• Upgraded: v2's central tendency is now calibrated and in agreement with the real network (mean 7.4 vs 7.0), on a propagation exponent confirmed against real measurements — not a literature guess. The "order-of-magnitude" caveat tightens to "central tendency matches."
• Unchanged (and now better-supported): the real distribution's upper tail (p90 14.8, max 37.3) is still not reproduced. With propagation now ruled in as approximately correct, this confirms — with evidence — that the tail is structural: co-channel traffic from overlapping metro deployments, gateway/high-site concentration, and the 2.3× diurnal swing — none of which an isolated, steady-state single-cluster sim contains. Reproducing it needs the device LocalStats loss counters (num_packets_rx_bad, num_rx_dupe), which Malla does not decode — the one remaining item worth asking the operator for.
• Caveat: survivorship bias means n=2.5 is a mild lower bound (failed links are unlogged); controlled measurements or the DB export would refine it. But the agreement with the independent literature value is itself reassuring.

Bottom line: the public Malla data let us confirm the propagation model against real measurements and bring the gate's central tendency into agreement with the real network — without operator permission. The residual is now a well-characterised, evidence-backed structural gap, not a PHY uncertainty.