TR NotesModels, reasoning states & infrastructure.
A reflection on black-box medicine, AI mathematics and dual-use research: when benefits become personal, how should we weigh uncertainty, and who may accept risks on behalf of others?
A complete sanitized account of an authorized client engagement, from initial access, a misleading RAG stack, and deployment remnants to local and domain privileges, SSH identity matching, vector export, candidate recovery, and ten types of AI error.
Starting with a cache hypothesis, I tested cross-session replay, cross-model reuse, and usage differences in Azure encrypted reasoning—and reconsidered the boundaries between state, caching, and trust.