Coding and AI classes in Dartford

The learner downloads the second volume of Francis Trevithick's biography of his father, about 16,000 lines. The first check is a plain exact search, and it fails even for a genuine sentence, because the book breaks that line in the middle, after the word "these". The fix is normalising: turning every run of spaces and line breaks into one space, straightening curly quotation marks, and ignoring capital letters. Now both genuine quotations are found exactly. For everything else the program slides a window the length of the quote across the whole book and scores how similar each window is, keeping the closest match.

The invented quotation scores 0.512 and is easy to reject. The altered ones are the real test. Changing "pioneer of locomotion" to "pioneer of the railway" drops the score to 0.848. But changing one year, 1832 to 1834, in "the Waterwitch Company made experiments with those plans", leaves the score at 0.984, a figure most people would read as a match. A rule such as "accept anything above 0.9" would approve a false date. And the stitched quote, real fragments joined into a sentence the book never contains, scores 0.795, uncomfortably close to a lightly misquoted genuine line.

So the learner builds a stricter checker. A quotation passes only with an exact match after normalising. A near match is reported with the book's actual wording beside it, so a person can see the difference. Any number in the quotation must appear identically in the matched text. This is exactly how a careful researcher, or a well-designed AI agent, should treat quotations: similarity is a clue for finding the passage, not proof that the quote is right.

Being able to check whether AI output is real is one of the most useful skills a Dartford teenager can build by coding in 2026. The longer argument is in why teenagers should still learn to code in 2026.