One flaw in how we matched titles put wrong figures on seven pages and four articles in a single morning

On 7 August 2026 we changed eleven published figures. Seven were on data pages, four were inside articles. Eight of the eleven came from one flaw, and it had been live for days. The other three were separate mistakes that surfaced only because the flaw made us re-read everything.

The flaw

Netflix does not publish a country of production. To build a Korean panel you have to decide which chart rows are Korean, and the obvious method is to take a list of Korean works from Wikidata and match it against the charts by title text.

That works until two countries make something with the same name, which they do constantly. Teach You a Lesson is Chinese. Hunger is Thai. Forgotten Love is Polish. The Empress is German. Each of them matched a Korean work with the identical English name, and each of them was counted as Korean.

Fifty-two titles were in our catalogue that should not have been. They carried four billion viewing hours with them.

Why counting would never have found it

This is the part worth keeping.

Every total looked healthy. 294 titles is a plausible number. 27.7 billion hours is a plausible number. Nothing was zero, nothing was negative, no check failed, and the figures moved smoothly week to week. A monitor watching for anomalies would have watched this for a year without a flicker.

The error was found because somebody read a page and thought that title does not look Korean. Not a threshold, not an alert — a person recognising a name. We have said before that things which fail loudly are the easy ones; this is the other kind.

The part that surprised us

Fixing the source was maybe a third of the work. The wrong figures had been quoted.

Two of our own articles cited the 294-title catalogue as background. A third used it to compute a different statistic entirely. /screen-split was built on the same base. /industry had an unrelated error — non-disclosing companies left in a denominator — that only surfaced because we were already re-reading everything.

A correction is not finished when the cause is fixed. It is finished when everything that repeated the number has been swept, and something exists to stop it recurring. Eight figures moved because one rule was wrong; only one of them was the rule’s own page. A sweep that stopped there would have left seven standing.

What changed so it cannot happen the same way

The rule lives in one file now. It had been copied into two build scripts, so fixing one left the other wrong. That is now scripts/lib/korean-netflix-titles.mjs, and everything that needs the judgement calls it.

The rule got a real test. Netflix splits its global chart by a title’s primary language, and a Korean work belongs on the Non-English side. Restricting to that removed the 52 automatically rather than by inspection. Nine more came out by hand after we read the largest titles one at a time.

Hand-made data files were converted to scripts. Seven data files on the site had no build script — they had been produced once and edited since. A file nobody generates cannot follow a rule change.

And a check now looks at the output rather than trusting the rule. Putting the rule in one place does not stop the next script from failing to call it, which is exactly what happened later the same day when a new collector we wrote re-introduced the same foreign titles. A check now reads every roster we produce and fails the build if a known foreign title is in it.

What is still wrong

204 of the 405 titles in the Southeast Asia panel have never reached a global Top 10, so the language test cannot be applied to them at all. They are in the panel on a name match — the same weak key that caused all of this — and we cannot tell you what share of them is foreign, because we have not read all 204.

That number is now printed on the pages that use it, and there is a check that fails our build if it stops being printed. It is the honest state of the thing: one flaw found and fixed, and a second, larger version of the same flaw that we can measure the size of but not yet close.