The deployed Buffalo prediction math (verified against the production code by 5,100 golden tests) evaluated on YOUR hypothetical storm — not a forecast. Hypothetical rain never touches the live predictions.
You paint on the 1 km radar grid the live system reads (3,540 cells). Exactly like the live radar reader, the models then average that rain over each gauge's catchment — and over each sub-watershed for the history floor and the locally-intense watch — so detail you paint inside a catchment is averaged away. What the models see shows those averages.
Rain falls at a uniform rate for the chosen duration; the rain→crest step at each gauge is a historical fit on 11 years of storms including the ones that fizzled, with a leaf-season term (median error ≈ 35–50% — wide on purpose; the live system avoids this step by watching real crests, which is why its downstream bands are tighter); downstream reach bands are the live router's own LOYO bands; recession starts from the predicted crest with no live-event anchor.
A history floor can lift a crest above the models: the lowest response — as a multiple of the starting flow — that three quarters of the past storms in the same season, at a similar flow for the month, with at least this much rain on the wettest sub-watershed AND at least this much averaged over the catchment produced — never a storm wetter on either count, and silent where fewer than 6 such storms exist in the 12-year radar record.
The big number on each card is the models' crest; In past storms like this is where the real crest landed, 8 times in 10, when this simulator was replayed on 303 real storms 2014–2026 (each storm scored by models fit without its year) and called the crest the same way — in summer the river usually came in lower than the models say, and the ranges are wide because summer rivers are: the same storm total on the same drought river has crested anywhere from under 100 to nearly 6,000 cfs at St. Joe.
The real storms are 36-hour radar totals; their timing (two nights of storms, a slow drought rise) is lost in a single-shot scenario. The live system's rain-event state machine (consumed rises, no-show cuts) doesn't apply either.
BIG FLOOD marks a models' crest of at least twice the gauge's flood level; the headline is then rounded, because at that size the exact number matters less than the call. Beyond the models' history marks a big flood where the models' crest passes half the biggest flood ever measured at the gauge, or the rain over its catchment passes what 99 in 100 storms in the radar record dropped. Both are labels only — they never change a number, a colour or a range. Could be a big flood marks a card whose models' crest is under twice the flood level while its "storms like this" range reaches it: a watch, not a warning, and a label only too.
Each access-to-access stretch has its own floatable and optimal levels, set on one gauge on the assumption that the river's flow grows with its drainage area. A storm that raises only part of the river breaks that assumption, so a stretch between two gauges also reads a crest blended from both by drainage area and shows the higher of the two levels: the blend can only raise a colour. Hover a stretch to see which one set it. Two local rules sit on top: when Pruitt → Hasty is floatable, every stretch from Hasty down to Grinders Ferry is at least Low but Floatable (the water reaching Woolum is always at least that, and the river around St. Joe holds water year-round); and when Kyles Landing → Erbie is floatable, Erbie → Ozark and Ozark → Pruitt are at least Low but Floatable (Erbie → Ozark never needs more water than Kyles Landing → Erbie). Below Harriet there is no second gauge, so those stretches read Harriet. Buffalo TV colours the river with the same rules, on the gauges' current readings.
Records are the USGS annual peaks; one marked provisional is the gauge's own measured reading from a year USGS has not published yet (hover a card's warning for its source).