Hydrometric / Evidence atlas
The evidence, before the conclusion.
We test UK flood-hydrology methods at national scale, publish what survives, show what fails, and state exactly where the evidence stops.
UK hydrology in numbers
- adjacent same-river gauge pairs
- 85
- stations in the QMED revalidation
- 918
- basins in the finite-sample test
- 27
- catchments in the ReFH parameter refit
- 640
Featured publication
What National-Scale Analysis Reveals About UK Hydrology
We asked what you can see about UK flood estimation from every gauged river at once that you cannot see from one catchment at a time. Four things: neighbouring gauges disagree about the same flood, an easy test flatters a new method, forty years of records only pin down so much, and which years you hold back changes the answer.
Read the evidence- Status
- Current publication
- Version
- 1.0.0
- Evidence cut-off
Evidence library
Current publications
Newest first. Each card carries the headline number, the two-line summary and the decision the work drove.
National hydrology
What National-Scale Analysis Reveals About UK Hydrology
85Neighbouring gauges compared
Look at every gauged river at once and four things show up that one catchment never does: neighbours disagree, easy tests flatter new methods, forty years only buys so much, and the years you pick change the answer.
What we decided: Report network disagreement, method performance, record-length sampling and time-period sensitivity as four separate lines of evidence — never combined into one number.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
QMED
Can Machine Learning Improve QMED?
918River gauges in the test
We gave a machine-learning model every advantage and it still lost to the FEH 2025 statistical method on 918 river gauges. Here is the losing scoreline, and why the losing test was the right one.
What we decided: Keep the FEH 2025 statistical method as the default for ungauged sites; the machine-learning model stays a screening and review tool, not an estimator.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
Statistical methods
The 2025 Statistical Method: What Changed and Why It Matters
5Linked stages that changed at once
Five linked stages of the UK flood-estimation standard changed at once in 2025. What each one does, which three numbers our software is held to, and which decisions still belong to the hydrologist.
What we decided: Treat the 2025 descriptor set, donor rule, pooling rule and urban adjustment as one chain that cannot be mixed with older versions, and gate three quantities against licensed reference values on every change.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
Growth curves
Skew, Scale and the Shape of UK Flood Growth
874Gauged rivers examined
A flood growth curve has four moving parts, and two promising shortcuts to it did not work. One needed the very gauge it was predicting; the other was not saved by the shortness of real records.
What we decided: Report flood size, spread, lopsidedness and tail growth as four separate quantities; drop rating-derived inputs as an ungauged route and stop tuning the c6_t3zero simulation configuration.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
Uncertainty
How Certain Is a Design Flood?
23.4%Method against the measured flood
We have three measurements that all sound like the uncertainty in a design flood. They answer three different questions on three different sets of rivers, so we publish them separately and never add them up.
What we decided: Report design-flood uncertainty as separate labelled lines of evidence; publish no single combined uncertainty band without a fitted joint model behind it.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
Rainfall-runoff modelling
From ReFH Design Events to Continuous Hydrology
640Catchments in the regional refit
A stocktake of four rainfall-to-flood modelling threads: one regional refit held up, one gap measurement landed near 50%, one daily model failed its own test, and one result was withheld for missing evidence.
What we decided: Stop tuning the daily lumped two-store branch, keep the baseflow-lag refit, and publish no regional continuous score until its immutable evidence artefacts are present.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
Flood estimation practice
The Modern Flood Estimation Report
1 of 11Required capabilities produced in full
We marked our own report generator against the eleven things a Flood Estimation Report must record. One is complete, three are partial and seven are missing — so we call the output a calculation report.
What we decided: Describe the current output as a structured calculation report, never as a Flood Estimation Report, and build the seven missing decision-record sections in the order this audit lists them.
- Status
- Current publication
- Version
- Version 1.0.0
- Evidence cut-off
- Evidence through
Publication standard
How we publish
We test a method the way it would be used on a real job, report the spread as well as the average, and say plainly where each finding stops. When a result turns out to be wrong we withdraw it in public.
A fair contest
Both methods get the same catchments and the same information. We test them the way they would actually be used.
The spread, not just the average
Every headline number arrives with how many catchments it came from, and how bad the worst cases were.
We say where it stops
Each finding states what it shows and what it does not. A result for 918 gauges is not a result for your site.
We publish the corrections
When better evidence changes a finding we publish a new version. The old one stays online, marked.
Living record
Recent revision ledger
- Version 1.0.0
What National-Scale Analysis Reveals About UK Hydrology
Initial publication of the national hydrology evidence synthesis.
- Version 1.0.0
Can Machine Learning Improve QMED?
Initial publication of the ML QMED adversarial revalidation.
- Version 1.0.0
The 2025 Statistical Method: What Changed and Why It Matters
Initial practitioner report on the FEH 2025 statistical method and bounded local implementation scope.
- Version 1.0.0
Skew, Scale and the Shape of UK Flood Growth
Initial publication separating scale, L-moment shape, growth, gauged rating evidence and finite-sample model failure.
- Version 1.0.0
How Certain Is a Design Flood?
Initial publication separating population validation, network consistency, finite-record sampling and proposed model-family sensitivity.
- Version 1.0.0
From ReFH Design Events to Continuous Hydrology
Initial publication of bounded ReFH parameter, baseline-gap, failed daily-branch and gated regional-continuous evidence.
Audit boundary
Source and disclosure principles
We disclose
Cohorts, validation design, baselines, distributions, important failure modes, limitations and traceable evidence identifiers.
We withhold
Deployable recipes, learned parameters, private data mechanics, customer information and licensed commercial outputs.
Stable current-report addresses begin with /research/; material revisions retain immutable versioned addresses.