The local-rain cards cover the storms the wave tracker can't see yet — rain that hasn't reached any gauge. Here is exactly when they fire, how the river's own baseflow tunes them, and the rules that make a prediction stand down.
Each gauge fires on short-burst intensity (rain in 6 hours over its local drainage). The prediction category scales with the same ladder: 1× the trigger = slight, 2× = moderate, 4× = large. A separate 24-hour soaker trigger exists only where 11½ years of data said it carries real signal — the big lower-river drainages.
| Gauge | Intensity trigger (normal ground) | Soaker trigger | Rain-to-peak window | No-show grace |
|---|---|---|---|---|
| Boxley | 0.45" / 6h · 'large' also needs 3"/24h | — | 5–10 h (median 7) | 3 h |
| Ponca | 0.45" / 6h | — | 5.5–12 h (median 8) | 3 h |
| Pruitt | 0.45" / 6h | — | 8–15 h (median 11) | 9 h |
| St. Joe | 0.4" / 6h | 5.4" / 24h | 9.5–22.5 h (median 16) | 12 h |
| Harriet | 0.45" / 6h | 5.6" / 24h | 9–20.5 h (median 14) | 12 h |
Triggers scale by how primed the ground is, and the best sensor for that is the river itself: each gauge's pre-storm baseflow (6–12 hours back, so a rise can't wet its own tier) is ranked against 11½ years of history for that month. Below the 25th percentile the bar rises by a third (dry); above the 60th it drops by a third (wet). The proof it matters — fizzle rate of rain episodes (rain fell, river shrugged):
| River says | Storms that fizzled |
|---|---|
| dry (<25th pct) | 30% |
| normal | 29% |
| wet (>60th pct) | 17% |
Recent-rainfall history is the fallback when gauge data is thin.
Four honest exits, all watching the river, not a timer: delivered — the predicted rise happened and the gauge is falling again; no-show — rain has been over for the basin's response time (see table) and the gauge hasn't moved; overdue — past the window's median with no qualifying rise underway and no rain; handoff — a tracked wave is already inbound for this gauge, and the wave forecast (which is better) takes the card's place.
Every number on this page is generated from the deployed model files and an 11½-year hourly replay (2014-12 → 2026-06, baseline v6). Generated 2026-07-02T22:57:24+00:00. Experimental — never rely on a single number for safety decisions.