How the waterfall rule works, and how well it works

The rule reads published elevation contours along mapped watercourses and looks for a drop of at least 12.0 m over a run of no more than 40.0 m. Where the terrain falls that fast over that short a distance, the rule proposes a waterfall candidate.

That is all it does. It is a statutory-free, purely topographic inference over an official layer, so a candidate it finds is published at Statutory inference — evidence class — never higher. It cannot see water, flow, season, or whether anything is there at all in a dry year.

A rule identifier is versioned on purpose. If the rule changes, published candidates change with it, and the identifier on each place page says which version found it.

The rule is scored against waterfalls the state already maps: a candidate that matches a mapped waterfall counts as a hit, one that does not counts as a miss, and a mapped waterfall the rule fails to propose counts against recall. At the chosen parameters it scores under 1% precision and 46% recall, for an F1 of 0.00.

Read that plainly: roughly over 99% of the candidates this rule proposes do not correspond to a waterfall the state maps, and it misses roughly 54% of the ones it does. Both numbers are the reason a rule-found place is published as statutory inference and labelled as a rule output on its own page.

Precision is scored against the state's own map, which is itself incomplete — a candidate counted as a false positive may be a real, unmapped fall. The score is honest about the comparison, not a claim that the state layer is the truth.

It is a score for the rule, and not for this site. It counts every candidate the rule proposes, and a candidate on its own is never published: one becomes a place here only where something independent of the rule agrees a place is there, either the register naming a waterfall at it or a community name doing so. Everything else is discarded as steep ground. So the rate above is what the sweep found, and it is not the rate at which the waterfalls on this site are wrong.

The parameters are the strictest pair in the grid below — the most drop required over the shortest run — and not the best-scoring pair. Scoring a rule for unmapped falls against the map would reward whichever setting best reproduces the map, so the sweep does not choose on score at all. It takes the most conservative definition and publishes what that costs.