Huseyin Buldurgan · Independent research
AI evaluation, agent safety and mathematical research. Each publication has its own full abstract, open PDF and links to its released resources.
Six preprints. These works have not been peer-reviewed; their pages state the scope and limits of the claims.
A mathematical preprint on minimum-amplitude design under moment and slope constraints. It gives an explicit fourth-order value expansion for finite and countably infinite switch sets under stated hypotheses, exact examples and a regularity obstruction, and a positive theta-kernel application with uniform sixth-order error bounds and a feasible-transport proof of strict ordering.
An unaffiliated audit of all seven RE-Bench task packages at a pinned source revision. It finds three measurement limitations: single-call kernel timings, scaling-law scores that are computed without running the proposed training, and a gap between the documented and implemented parameter-distance metric. It ships eight bounded contract checks. Historical results were not reproduced.
100 synthetic cases in 25 matched four-variant groups, run on two API-delivered systems for three epochs each (600 protocol-complete runs). Neither system acted in any of the 300 invalid-authority runs. Exact authorized execution differed: 150/150 versus 99/150. The results cover this synthetic protocol only; they are not production incident rates or a general model ranking.
Treats a queued agent action as a bounded authority contract that must still hold at execution time. Uses a frozen masked review set of 151 high-priority packets, together with fail-closed measurement checks and synthetic negative controls.
Proposes that evidence authority be treated as an operating control surface between retrieval and downstream use. It separates the retrieval, qualification, generation, action/publication and audit planes.
A controlled 1,000-packet benchmark that separates topical relevance from evidence support, action readiness and same-identity authority. The sealed run scores 2,990 decisions across three commercial LLM APIs, with a relevance reranker as the baseline.