Oleksii Simon

Independent researcher (solo, Ukraine) working on evaluation reliability for AI systems that track human state over long conversations. Built DriftBench — an open-source (Apache-2.0), fully deterministic benchmark for belief/state drift with no LLM judge. Measured a three-layer noise decomposition on a peer-reviewed memory benchmark (ES-MemEval, ACM WebConf 2026): judge ≈0.10, answerer ≈0.05, ingest ≈±0.40 — the dominant noise source is the one nobody reports. Pre-registered, adversarial-to-self methodology; negative results published as artifacts. All work to date on free-tier inference, $0 external funding.

Personal Website

https://github.com/simon9679/driftbench

Projects

No grants recorded.