Coding classes in Cumbria

The learner takes the first volume of each poet's collected works from Project Gutenberg, splits the text into chunks of 300 words, and labels each chunk with its author: 342 Wordsworth, 572 Coleridge. Four-fifths train the model, the rest test it. Naive Bayes is simple: for each author, count how often every word appears, then score a new chunk by how likely its words would be under each author's counts, with one added to every count so an unseen word does not wreck the sum. The chunk goes to whichever author scores higher.

The first row is the lesson people forget. A model that always says "Coleridge" would already be right 60.7 per cent of the time, because there are more Coleridge chunks. Any accuracy figure is meaningless until you compare it with that lazy baseline. Ours beats it comfortably. The last row shows the other honest limit: with only ten words to go on, the model is right 79.1 per cent of the time, because short passages simply carry less evidence.

Then the twist. Asked which words most strongly pointed to Coleridge, the model listed "linenotes", "MSS" and "sibylline", editorial markers from his edition. Its strongest Wordsworth clues included "Marmaduke" and "Oswald", characters in his play The Borderers, and the words "variant" and "footnote". The model had partly learned to recognise the editors, not the poets. Removing the giveaway words cost only two points, and a separately published Ancient Mariner was still classified correctly every time, so some genuine style signal is there. But a learner who never asks "what did it learn?" would never know how much.

So a Cumbrian teenager should learn to code in 2026 to be able to open up a model and see what it has really learned, a skill worth more than any single score. The longer argument is in why teenagers should still learn to code in 2026.