LatentAtlas is the independent research line of Huseyin Buldurgan. I run controlled API experiments, build Python evaluation tooling and audit benchmark source code. I publish technical reports whose results anyone can recompute from public artifacts.
Six preprints, with full abstracts, directly hosted PDFs and versioned resources. These papers have not been peer-reviewed. Each one states the limits of its claims.
Browse all publications and PDFs
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.
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.
Public artifacts that let anyone check the reported numbers without trusting the author.
A standard-library Python package: an Inspect evaluation, deterministic verifiers, a transcript-free 600-row results ledger, unit tests and GitHub Actions. It also hosts the RE-Bench contract checks. MIT licence.
The 100 synthetic cases and the 600-row results ledger, byte-identical to the GitHub release (same SHA-256 hashes). The dataset card carries a training-data canary string. MIT licence.
How the work is built so that its claims stay checkable.
I’m Huseyin Buldurgan, an independent researcher based in Adana, Türkiye. I started LatentAtlas in January 2026 to study how AI systems should treat evidence before they act on it. The work grew out of practical data problems, where similarity-based matching kept producing confident but unsupported decisions.
Before research I spent a decade in business roles. I was a sales specialist and then sales manager at Vestel (2016–2021), onboarding investors and business partners and bringing new stores into the regional network. I also worked in investor sourcing at mbco Strategy Consulting and in a family agricultural business. I hold a BA in Management from Sabancı University.
I’m looking for a research or research-engineering role where I can work full-time on evaluation validity and agent safety, and I’m willing to relocate to the UK or EU.
Research collaboration, role enquiries or questions about the work.