Rain Gauge Quality Control

Two real cases from the last fortnight. The bars are what the gauge sent every 15 minutes; the dark line is the radar over the gauge. When a gauge reports heavy rain and the radar sees none at all, something is wrong.

The quieter fault is a gauge that stops counting: a funnel blocked with leaves reads zero, and zero looks like a dry day. It only shows against its neighbours.

Every reading is tested against fixed rules: limits on what is possible in 15 minutes, sudden spikes, the gauges within 100 km, the radar overhead, and the patterns a faulty tipping bucket leaves behind. All three gauges above failed at least one.

Rules are fast, give the same answer every time and are easy to inspect. Each is set out in The checks.

The rules are deliberately sensitive, as a missed fault does more harm than a second look. The catch: to a rule, a heavy shower over one gauge looks exactly like a fault. It is a spike, and an outlier against neighbours that stayed dry.

Most flags are like this: real rain. Something has to weigh the evidence before a gauge is called faulty.

Every reading goes through the same three steps, every 15 minutes.

1. 1The eight checks run.
   All pass, and under 20 mm: passed. Not reviewed, not listed here, shown everywhere as normal.
2. 2Three hard rules. 60 mm or more in 15 minutes; stuck at zero while it rains all around (Cynefail, above); or 10 mm or more at 8 times the radar or more, while the gauge's 24-hour total is also far above its neighbours'.
   Likely faulty decided by the rule. The AI is not asked to overrule it, and cannot.
3. 3The AI review. Every other flagged reading, and any reading of 20 mm or more even if every check passed.
   Likely faulty Uncertain Likely genuine
   Still waiting for its turn (at most 40 a run): Not yet reviewed

In short: the AI has the final say on every reading it reviews, except the three hard rules, which always win. A failed check on its own never makes a gauge faulty, and nor does the quality index, even at 0: the 35 mm gross error check, a spike or a dry radar are evidence for the review, not a verdict. Whitburn and Crew Fell were called faulty by the AI; Cynefail by the rule.

Each flagged reading goes to a language model with what a forecaster would look at: the reading and the gauge’s last 24 hours, every check and its numbers, the radar within 5 and 30 km in the runs either side, the nearest gauges, the Met Office UKV forecast and UK rainfall records.

It returns a verdict, likely genuine, uncertain or likely faulty, and two or three sentences in plain language saying why. Recent ones, as written:

A fair question, and it has been checked. Over 928 reviews from 23 September to 2 October 2026, one simple rule (suspect if the radar or 24-hour check failed, or the radar within 5 km saw almost nothing while the gauge had 1 mm or more) agreed with the AI on 908. So rules could make most of the calls. The AI earns its place three ways.

• The hard cases. Where the two disagreed, the AI was right. Whitburn reported exactly 11.5 mm every 15 minutes for 19 hours: only the neighbours check fired, and the drip check ignores rain that heavy, so a rule let it through. Nine real showers a rule would have thrown out were kept, because radar 15 minutes earlier, or a wet neighbour, showed the rain was real.
• A reason you can read. A rule says “spatial outlier, quality index 0.65”. The review says what is going on, for instance a gauge that has recorded nothing all day while its neighbours and the radar had rain, so its funnel is probably blocked.
• Mixes nobody wrote a rule for. Each new fault pattern would otherwise need a new rule, tuned and tested. The review takes an odd mix of evidence as it comes.

When a gauge is first flagged, or a reading reaches 20 mm. While the gauge stays flagged its verdict carries forward, and the AI is asked again about every hour (every fourth run), or sooner if another check starts failing or the reading jumps.

At most 40 readings are reviewed each run, heaviest first; the rest show as not yet reviewed until a later run. That is why the flag history shows each review once, with the runs it covered beneath it.

The rain gauge maps and tables (Rain gauge, Multi view, Rain gauge table, Radar & gauges, the gauge layers on Radar accumulation and Compare, and the homepage) have a Quality control switch. It is on unless you turn it off.

• Only for the time it covers. A gauge is left out only when a faulty reading falls inside the window that page adds up. Crew Fell’s glitch drops it from a 24-hour total that day, but not from a 1-hour total the next morning.
• Uncertain is your call. It stays in, so a real downpour is not hidden on a guess. Open it here to see why before relying on it.
• Nothing is edited or deleted. Turn the switch off and every gauge is back. A gauge’s own page never hides itself; it carries a note when it was last judged likely faulty.

Likely faulty

The AI judged the reading suspect, or a hard rule ruled it out (60 mm or more in 15 minutes, stuck at zero in the rain, or far above the radar with a day's total far above the neighbours'). Left out of the other gauge pages for you.

Uncertain

The evidence points both ways. Still shown elsewhere: check it against radar and nearby gauges before relying on it.

Likely genuine

A check failed, but radar, the gauge’s own history and its neighbours support the reading. Usually a real, local downpour.

Not yet reviewed

Flagged, with no AI verdict yet; it follows on a later run.

The counts at the top of the list filter the list and the map. Move through earlier runs with the slider under the map: red and amber marks show when gauges were judged faulty or uncertain. Flag history lists each episode over the chosen window, grouped by AI review. Runs are kept for 14 days and the history for 90.