The orange "AI Rainfall Event Analysis" card is the hardest thing on this site to explain — so here is all of it. The short version: every number is computed by transparent math before any AI sees it. The language model writes the sentences; it cannot change the call.
It arms when a quarter inch of rain (area-weighted) has accumulated over the four headwater drainages above Ponca in the current rain spell, runs a fresh analysis every hour while armed, and stands down after 12 consecutive dry hours — unless a surge it predicted is still traveling toward St. Joe, in which case it holds until the wave's window closes.
The core question the card asks is simple: in 11½ years, what did storms that looked exactly like this one go on to do? The library holds 394 historical storms sliced into 7,319 hourly snapshots, each labeled by how its storm ended: 228 Fizzles (rain fell, Ponca never reached 200 cfs), 82 Moderate (200–900), 18 High (900–1,600), 66 Floods (over 1,600). The current hour is matched against its 18 nearest neighbors on five measurements — rain so far, current flow, the last 6 hours' trend, the pre-storm baseline, and how strongly the river is responding per inch of rain — and the neighbors' outcomes become the odds you see.
Three honesty rules shape the match: no borrowing from bigger storms — candidate storms are only ones whose eventual total rain stayed within 1.75× of what has fallen here so far (most storms with 1" don't grow into 4" storms, and the odds shouldn't pretend otherwise); storm-balanced sampling — one long flood can't flood the neighbor pool with its many hours; and a stage clamp — the card can never call an outcome below what the river has already reached.
Statistical lookalikes have blind spots, so hard physics overrides them in three cases: if Boxley is already surging, a wave is physically 1–3 hours from Ponca and the card floors at what that wave implies, whatever the rain odds say; if the hand-calibrated watershed tripwires on the feeder drainages (Upper Buffalo, Beech, Adkins, Boen Gulf) have fired FLOOD, the call floors at Flood; and past 2.5 inches areal / 3 inches on the wettest drainage, history says Ponca floods ~83% of the time regardless of what the gauge shows yet — the card says so.
The storm's characteristics are captured once, when the event arms: pre-storm ground wetness as of spell start, and rain totals as running storm maximums. A storm is never allowed to change the definition of its own starting conditions mid-event — that lesson (audit 2026-07-01) is now baked in everywhere.
When Ponca peaks, the card projects the surge downstream: Pruitt runs a near-constant ~1.13× Ponca about 5–7 hours later; St. Joe uses the fitted SJ_MODEL equation (see the wave-tracker page for the same idea) — it reads how the storm's rain was split between the headwaters and the middle basin plus ground wetness, because a wave crossing 1,300 dry km² attenuates while one riding wet middle-basin rain amplifies up to ~4×. It only runs in its fitted domain (a real event ≥800 cfs with ≥0.5" of generating rain); below that it claims only the routed wave, no amplification. Travel windows are 9–16 h on wet ground, 11–18 h on dry.
Everything above happens in plain arithmetic first. The result — stage call, odds, floors, downstream numbers — is handed to a local language model (qwen, running on our own hardware) which writes the narrative paragraph you read. It is explicitly instructed never to soften or exaggerate the computed call, and its drafts are logged and graded against what the river actually did.
Honest limits: the analog match is anchored to Ponca's own gauge, which lags the rain — that's exactly why the physics floors exist. And 394 storms is a good library, not an infinite one: genuinely unprecedented storms will match poorly, and the card's confidence language reflects thin matches.
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 2026-07-02T23:52:20+00:00. Experimental — never rely on a single number for safety decisions.