cross-lingual value drift
an eval on inspect that asks the same value-laden question in matched translations and measures whether the answer moves along a named axis.
37 findings across an api, a frontend, dns and the infrastructure behind it: one critical, twelve high, all disclosed.
agent infrastructure in production
more than twenty mcp servers on azure container apps, each encoding a whole workflow from style-learning to compliance validation.
chatgpt agent-mode connectors
mcp servers as chatgpt connectors: oauth 2.1 and a manifest generated at runtime, later re-pointed at claude desktop.
an mcp server that gives an agent beat matching, eq and transitions in ableton live, and lets it mix a set unattended.
a moral judgement classifier: roberta and bert against svm baselines, with shap to see which tokens moved a prediction.