Statistics

How Big Does an A/B Test Need to Be?

Sample size is usually decided by how long someone will wait. Simulated across a grid, a genuine one-point lift on a ten percent baseline needs 20,000 per arm to be found reliably — and is detected 17.8% of the time at 2,000.

23 Aug 2026 · 5 min read

Statistics

What Happens When You Check an A/B Test Early

Twenty thousand simulated experiments with no real difference in them. Read once at the end, 4.9% came back significant, as they should. Checked twenty times and stopped when green, 24.8% did.

23 Aug 2026 · 5 min read

Statistics

It Was Not a Fivefold Rise in Terabit DDoS Attacks. It Was Sixfold.

Cloudflare mitigated 935 attacks above 1 Tbps in the first half and reported a 519 per cent quarterly rise. A 519 per cent increase is a multiple of 6.19, and the published totals prove which reading is right.

11 Aug 2026 · 6 min read

Statistics

Cloudflare Says Humans Will Be a Rounding Error. We Checked Our Own Logs and Could Not Tell.

Cloudflare's CFO told analysts non-human traffic could reach a thousand times human traffic within five years — and that its own measurement already put the crossover in May 2026, a year earlier than it had forecast. We went to check our own server logs and found we cannot answer the question at all: the user-agent column is empty on all 321,536 rows.

9 Aug 2026 · 7 min read

Statistics

16 Working Viruses Out of 285 Built — and the Survivors Were Not Copies

Genome language models designed 16 viable bacteriophages out of 285 physically built. The reflexive debunk says the survivors were near-copies of the template; the researchers report up to 392 novel mutations, an altered capsid protein and phages fitter than the original. The number nobody has published is how many candidates there were.

8 Aug 2026 · 9 min read

Statistics

Grades, GPAs and What They Convert To

NUS publishes separate grade profiles for A-level and polytechnic applicants, and none at all for IB — because, it says, the sample is too small to be accurate. Malaysia's UPU puts STPM, Matrikulasi, Asasi and STAM into one category with one calculator. Two systems, opposite answers, same unanswerable question. Plus the reason our own converter told a first-class Malaysian student they had 33%.

1 Aug 2026 · 9 min read

Statistics

Kati, Tahil, Cups and Grams

Ask for a catty of pork and you get 500 g in Guangzhou, 600 g in Taipei and 604.8 g in Singapore — all three legally correct. A recipe's "1 cup" spans 180 ml to 284 ml depending on who wrote it. Here is what the units in an ASEAN kitchen actually equal, why cup measures betray you, and why the knuckle trick for rice works when a ratio does not.

1 Aug 2026 · 12 min read

Statistics

Three Breaches Caused Eighty Per Cent of the Half-Year's Victim Notices. The Other 1,800 Shared the Rest

The Identity Theft Resource Center counted 471 million victim notices in six months, more than all of last year. Take out the single largest incident and the total falls below 2025 again.

1 Aug 2026 · 6 min read

Statistics

Mean, Median, Mode: The Average Is a Choice, Not a Fact

Three different numbers are all correctly called "the average", and which one gets published changes what a reader concludes. Singapore's DOS reports a mean of S$16,159 against a median of S$12,446 for 2025; Malaysia's DOSM reports RM9,155 against RM7,017 for 2024 — the same 1.30× gap from two separate surveys, and both agencies lead with the median. Plus the US Census Bureau's written rule, what one outlier does to each average, and the currency-column bug that made our own calculator report a mean of zero until today.

26 Jul 2026 · 15 min read

Statistics

A High R² Does Not Mean the Model Is Right

Anscombe printed one regression summary for four data sets in 1973 and said the features that matter are "invisible if we see only the usual quadratic summaries". Reproducing his table at full precision shows the collision is at print precision, not to the last digit — 0.66654, 0.66624, 0.66632, 0.66671 — and he never printed the correlation everyone attributes to him. Plus the mirror image (thirteen published data sets at R² ≈ 0.004, one of them a dinosaur), a noiseless parabola whose R² walks from 0.9985 to 0.0000 as the sampling window widens, and the morning our own calculator called a perfectly vertical line "essentially random".

26 Jul 2026 · 20 min read