Ir para o conteúdo

Live Engine Validation — 5 Real Brazilian Links

Date: 2026-05-16 Engines under test: FSPL (reference), ITU-R P.1812-7 (Py1812), ITM/Longley-Rice (itmlogic 1.2) Climate auto-detection: rf_engines/br_climate.resolve_climate(lat, lon) Terrain: flat-fallback for this proof (production path uses SRTM-3 90 m via srtm_elevation.SRTMReader); the goal here is engine-functional proof, not absolute terrain-loss accuracy.

Setup

from rf_engines import fspl, itu_p1812, itmlogic
from rf_engines.br_climate import resolve_climate

Results

# Link Distance Freq Climate FSPL P.1812-7 ITM P.1812 vs ITM Δ
1 Brasília centro → Plano Piloto Norte 12 km 700 MHz CO (continental temperate, N₀=320) 110.9 dB 116.1 dB 115.7 dB +0.4 dB
2 Av Paulista (SP) → V. Mariana 3 km 1800 MHz SE (continental temperate, N₀=320) 107.1 dB 107.3 dB 109.8 dB −2.5 dB
3 BSB → Goiânia (NLOS over-horizon) 173 km 900 MHz CO (continental temperate, N₀=320) 136.3 dB 187.4 dB 184.5 dB +2.9 dB
4 Manaus centro → Iranduba 25 km 1800 MHz N (equatorial, N₀=360) 125.5 dB 130.2 dB 131.0 dB −0.8 dB
5 Fortaleza Praia do Futuro → coast 11 km 2600 MHz NE (subtropical/maritime, N₀=340) 121.6 dB 121.8 dB 123.4 dB −1.6 dB

Acceptance criterion

P.1812-7 vs ITM cross-engine agreement should fall within ±5 dB in 5/5 cases for the engines to be considered "functionally validated as production-ready peer engines."

Result: 5/5 within ±3 dB. Maximum gap = 2.9 dB on the 173 km NLOS over-horizon case (the hardest geometry tested). All values are physically plausible.

Notes

  • Climate region was correctly auto-detected per coordinate (CO, SE, N, NE) by br_climate.resolve_climate().
  • The itmlogic engine had a production bug discovered during this run — predict_basic_loss() passed climate= and ns0= kwargs that _run() did not accept (TypeError). Fixed in commit by adding the two kwargs to _run()'s signature and threading them into the prop_vars dict.
  • After the fix, pytest tests/ -k "engine or itmlogic or p1812 or rf_engines" → 106 passed, 0 failed.
  • This file is the canonical "engines work on real Brazilian geometry" evidence requested during the 2026-05-14 audit. To reproduce: see prove_engines_live.py at the repo root (proof script).

What this evidence does NOT cover

  • Absolute accuracy vs measured RSSI on these links (drive-test required — see Due Diligence FAQ §9 POC offer).
  • Mass coverage raster (Sionna RT coverage mode is roadmap, gated by SIONNA_RT_DISABLED).
  • Vegetation P.833 attenuation term in the plan_repeater edge cost (roadmap).
  • Rain-rate P.530/P.838 attenuation term in the plan_repeater edge cost (roadmap).