Coding classes in the East Riding of Yorkshire

We cannot count real birds from a keyboard, so the learner builds a model: a cliff of 400 sections, where 60 are crowded ledges averaging 4,500 birds, 140 are middling at around 1,100, and 200 are sparse faces with about 120. Every number is our own teaching value; the model cliff holds 447,916 birds, the right order of size for Bempton. Now pretend to be a surveyor with time to count only 20 sections, multiply up, and see how close the estimate lands. Then do it ten thousand times with different random choices and look at the spread of answers.

The first row is the trap. Simple random sampling is fair on average, but twenty random sections might catch two crowded ledges or six, and that one accident swings the whole total. Only a quarter of those surveys come within 10 per cent. Stratifying fixes it: divide the cliff into its three types first, sample within each, and add up the parts. Now the lucky or unlucky draw of ledges no longer decides the answer, and four surveys in five land within 10 per cent.

The last row adds a subtler idea. Spend your counts where the numbers vary most, not where the sections are most numerous: the sparse faces barely change from one to the next, so two counts there are enough, while the crowded ledges need nine. This rule, known as Neyman allocation, takes the share of each stratum to be its size multiplied by how much its sections vary, and it pushes the success rate to 93.4 per cent with exactly the same effort. All three methods are right on average; the difference is how often a single survey is right.

So an East Riding teenager should learn to code in 2026 to design the counting, not just to add it up. The longer argument is in why teenagers should still learn to code in 2026.