Phase 3 →Hypothesis registry + VOACAP correctors — learned correctors now realized as the P4 offline predictor.propagation_hypotheses table. Monthly LLM analysis job: score → synthesize KB writeup → propose new hypotheses. Constantly-updating knowledge loop.
Phase 4Dashboard scorecards — calibration curves, per-band predictability, hypothesis ledger, seasonal views. Full UI fills in as ≥12 months of data accrues.
Scientific Loop
HypothesisFalsifiable statement — e.g., "20m opens to EU when SFI>120 & Kp<3, 12–16 UTC"
PredictionVOACAP REL + learned correctors → per-band open/partial/closed forecast for each UTC hour
Source: sdr-decode.py off co-located KiwiSDR — actual FT8 decoded count per 30-min slot.
decodes>0 = band open for NVIS range (~200km). Ground-truth for VOACAP accuracy scoring.
VOACAP Prediction Accuracy
No scoring data yet — weekly cron runs Sunday 09:00 UTC. Needs ≥7 days of nvis_ground_truth to score.
Accuracy: % slots where round(VOACAP rel/100) matched FT8 open/closed. Sensitive to the 50% threshold — use Brier for calibration.
Brier: mean (predicted − actual)² — lower is better (0=perfect, 0.25=worst-case).
BSS: Brier Skill Score vs climatology (predict open-rate). Positive = adds skill.
Caveats apply at n_obs < 100 (VOACAP is monthly-median; day-to-day solar variance is un-modelled).
GhostNet Receiver × Station Matrix
7.107 MHz USB · JS8 Normal · Thursdays 21:00 ET (01:00 UTC Fri) ·
Rows = KiwiSDR receivers · Cols = heard stations ·
Cell = decode count + best SNR
Loading…
Receiver Summary · all GhostNet-window captures
Loading…
rx_score predicts NVIS reception; this matrix is the empirical ground-truth.
Receivers that decode frames validate the rx_score model.
Source: decodes.jsonl (dial_khz=7107) + js8_messages DB.
Callsigns available only for captures made after 2026-07-12 (js8decode.py update).
Offline-inference layer. Consumes the correction factors learned by the
Science tab's
scoring + fusion pipeline (VOACAP + WSPR ground truth). Science = R&D; P4 = the operational, internet-free forecast.
P4 HF Forecasts
Calibrated WSPR-SNR predictions per circuit × band × horizon.
Cal SNR = VOACAP (100W, physical re-baseline) + WSPR correction factor.
Values in dB; WSPR detects to ~−27 dB. Null = no factor yet (needs ≥30 WSPR samples).
Loading…
Accuracy · calibrated vs VOACAP vs actuals
MAE from propagation_forecast_outcomes. Outcomes accumulate after WSPR
observation windows close (~23:00 UTC). First forecasts committed 2026-07-29; first scored outcomes 2026-07-30 23:00 UTC. P4.9/P4.11 gate: ≥7 days (~2026-08-05).
No scored outcomes yet — first results expected 2026-07-30 23:00 UTC (≥7 days for P4.9/P4.11: ~2026-08-05)
Diagnostics · miss bins + feedback + LLM reasoning
