A year of bugs on my own site and services, sorted by how they were found. The expensive ones share a shape: no individual decision was wrong, nothing logged an error, and per-step review could not have caught any of them.
A reranker evaluated with the right answer guaranteed to be in the candidate set is being measured on a problem it never faces in deployment. What is left once you remove that guarantee is a ceiling set by retrieval, not by the reranker.
ICML, ECML PKDD, ICICS and ProbML in a single cycle — and the two habits that mattered more than any individual idea.
Six findings from running a real security audit against this site. The worst one had been live for months and was caused by a single misspelled environment variable.
The VerifAI @ ICLR version and the ICICS main-track version answer the same question. Only one of them proved it at a size anyone cares about.
The paper was accepted. I withdrew it because I could not justify the registration fee — and that is a structural problem worth naming.
Pairing a cheap fast model with an expensive LLM is now a standard pattern, and the trigger is almost always a hand-tuned threshold. Treated as a sequential stopping problem instead, the classical trigger families all turn out to be the same object.