L-moments
How hard can it rain here?
Every dam in the country is designed around that question, and nobody has a rain record long enough to answer it directly. The standard answer comes from regional frequency analysis with L-moments — pooling many rain gauges to estimate a storm rarer than anything in the record. I’m running that method uniformly, from one codebase, at every dam in the national inventory — roughly 92,000 of them.
The code, in the openAsk about the work
A screening and research tool, full stop. It shows where an expert should look; it does not produce engineering numbers, and no result speaks to any individual dam’s safety.
What’s being built
One method, every dam
24-hour and 72-hour design storms out to the 1-in-10,000-year event, computed the same way everywhere — with the estimate, the uncertainty around it, and a map of every gauge used or rejected, with the reason.
Run more than one defensible way
The interesting result isn’t a single number. Reasonable experts make different, equally defensible choices — and at the rare storms dam safety actually cares about, those choices move the answer by double digits. Running the whole inventory several ways makes that spread measurable instead of anecdotal.
Checked against the sources you already trust
Validation is pre-registered: the acceptance thresholds were committed to the repository before any comparison ran, against the national rainfall standard and against a peer-reviewed regional study that a state adopted as its dam-safety rule. The git history is the timestamp.
Reproducible on purpose
The statistical core is the method co-inventor’s own reference implementation; the pipeline reproduces the textbook’s worked examples and an independent hand calculation to the digit, and two independently built runs months apart came back essentially byte-identical — which is how a cache defect that made results depend on processing order got caught at all.
Why it’s being done this way
Anyone can publish a national table of big numbers. The harder and more useful thing is a national table whose every number can be taken apart by a reviewer who doesn’t trust you.
- Audit trail over assertionEach result carries the decisions behind it — which gauges entered the region, which were screened out and why, which distribution was chosen — so a reviewer can follow the analysis step by step rather than take a depth on faith.
- The misses get published firstA validation that reports 100% agreement reads like a brochure. Every unexplained mismatch against the reference studies is published and listed ahead of the successes.
- Inputs are treated as unverifiedThe precipitation record and the dam inventory are public data with known defects — including coordinate errors this work has already catalogued. Nothing is assumed clean.
- Findings before conclusionsWhere estimates disagree with the standard, the disagreement turns out to be predictable from measurable properties of the surrounding gauges — so the honest claim is not “we match” but “we match, and we can tell you where to trust us less.”
What’s coming
The national sweep and the pre-registered validation tiers are running now, with the findings — a failure atlas, the geography of the rare-storm tail, and the measured cost of expert judgment — written up as they clear. There will be more to say publicly before long.
Follow along on GitHubGet in touch
Work by True Ascent Labs on public data from NOAA/NCEI and the USACE National Inventory of Dams. Code is MIT-licensed; results are for screening and research, and are not valid for engineering decisions without expert review.
