https://selinkurt.com/journal/ Notes from the middle of building AI agents: what the tests showed, what broke, and what changed. en kselin26@m.fudan.edu.cn (Selin Kurt) Mon, 02 Nov 2026 09:00:00 +0000 https://selinkurt.com/journal/signal https://selinkurt.com/journal/signal Mon, 02 Nov 2026 09:00:00 +0000 The agent was on the desk and I was not. What I built to close that gap, the three failures it hid, and the two things it still cannot do. https://selinkurt.com/journal/v3-distillation https://selinkurt.com/journal/v3-distillation Sat, 31 Oct 2026 09:00:00 +0000 Weights alone answered 100 MPa for a material that does not exist. Runtime on, the same question got no number. The evidence that honesty here is an intervention, not a disposition. https://selinkurt.com/journal/demo-launch https://selinkurt.com/journal/demo-launch Sat, 31 Oct 2026 08:00:00 +0000 Try NEXUS yourself: every answer's record, honest refusals, citable transcripts. Plus what V3 adds and where its precision gate stands. https://selinkurt.com/journal/v2-build https://selinkurt.com/journal/v2-build Fri, 23 Oct 2026 09:00:00 +0000 The check-first runtime, three training rounds and what each one taught, the tests behind the claims, and what still fails. https://selinkurt.com/journal/first-runs https://selinkurt.com/journal/first-runs Wed, 21 Oct 2026 09:00:00 +0000 Voice and register held better than instruction-following in the first NEXUS runs, and what that points at for V2.