https://localfirstai.eu/tags/exp_045/Recent content in Exp_045 on Local First AIhttps://localfirstai.eu/images/og-default.pnghttps://localfirstai.eu/images/og-default.pngHugoen-usWed, 21 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-21-the-model-was-the-ceiling/Wed, 21 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-21-the-model-was-the-ceiling/For three experiments we treated every weakness in our lead triage as a problem with the questions or the code rules. Then we put a 27B decision model behind exactly the same rules, and request accuracy went from 0.919 to 0.991 on the same synthetic enquiries. What we shipped instead is two small models splitting the work. Once the reader was strong enough, the remaining errors were in the questions: telling the model today's date nearly doubled its accuracy on timing given only as a month, date or holiday.