Notes: David Mashiahhttps://davidmashiah.com/feed.xml2026-07-28T00:00:00ZDavid Mashiahhttps://davidmashiah.comThe training set moves the error 390x. Nothing predicts it.https://davidmashiah.com/notes/training-set-choice-moves-the-error2026-07-28T00:00:00Z2026-07-28T00:00:00ZAcross 231 training-set choices the worst case moved by a factor of 390. I tested nine cheap ways to pick a good one. None of them held up.The metric was measuring my model, not the problemhttps://davidmashiah.com/notes/the-metric-was-measuring-my-model2026-07-28T00:00:00Z2026-07-28T00:00:00ZA transferability score held at 0.54 across two independent budgets, then collapsed to 0.14 when I changed the model class. It was tracking the algorithm.You cannot count the sale you did not makehttps://davidmashiah.com/notes/the-sale-you-did-not-make2026-07-27T00:00:00Z2026-07-27T00:00:00ZSales data records what you sold, not what people wanted. The two differ exactly on the days they matter most, and the gap teaches a forecast to shrink.A scratched label still scans, and that is not luckhttps://davidmashiah.com/notes/a-scratched-label-still-scans2026-07-27T00:00:00Z2026-07-27T00:00:00ZA check digit tells you a barcode scan was wrong. Reed–Solomon parity tells you what it should have been. That difference decides what a damaged label costs.Why a screen cannot measure vernier acuity, and how close it getshttps://davidmashiah.com/notes/why-a-screen-cannot-measure-vernier-acuity2026-07-25T00:00:00Z2026-07-23T00:00:00ZA screen pixel at 60 cm subtends about 90 arcseconds; clinical vernier thresholds are 2–5. Antialiasing still measures below one pixel. Here is the arithmetic.My model scored perfect and that was the problemhttps://davidmashiah.com/notes/my-model-scored-perfect2026-07-25T00:00:00Z2026-07-23T00:00:00ZA perfect held-out score measured one thing: whether the model could repeat what it had seen. Of 30 cross-family tests, 20 fell below a crude baseline.