I am an independent researcher and research software engineer focused on the reliability of AI safety evaluations. I build reproducible probes and verification tools to check whether benchmark inputs, tests, and scoring rules support the conclusions drawn from them.

Selected evaluation work (status checked 2026-11-14)

• ControlArena input audit (report) — Checked 24 stored backdoor inputs across 10 sampled problems and identified an unresolved conflict between a stated input constraint and recorded data. The PR author responded with scope guidance; validator coverage and format policy remain maintainer decisions.

• ControlArena side-test probe (report) — Tested reconstructed Django functions locally for tasks 15103 and 16661. In 15103, an or-combined check accepted both escaped and unescaped & output. On 2026-11-05, another contributor reproduced that behavior on SWE-bench Docker images. Maintainer decisions on the fix location and 16661's intended test remain pending.

• FACE-Eval Inspect scorer report — Used a local AST-extracted scorer harness with stubbed judge responses. For a valid cue-aligned answer with missing or unparseable reasoning, observed VCR=0 and UAR=1. The no-trace VCR result conflicts with the documented NaN behavior; UAR and parse-failure eligibility remain questions for upstream. No effect on published paper results was demonstrated.

• Debate control schemes sensitivity note — Separated the Phase 2 erratum's world-standardized question-side weights from the strict unit-weighted realized schedule and derived conditional sensitivity bounds on archived Phase 2 outputs.

• Manifund funding-bench — Proposed an offline diagnostic and unit tests showing how proposal attribution and scoring windows affect a synthetic prediction. Production scoring and data are unchanged.

My studies above are offline or archived-output investigations. I did not run full evaluation pipelines, SWE-bench Docker environments, or live model calls.

Engineering background

My wider verification work includes an exact rational checker for a HiGHS optimality bug confirmed by a maintainer. Upstream contributor Opt-Mucca authored fix #3408, which merged on 2026-10-16 and shipped in HiGHS v1.16.0; my role was reporting and verification. In Ethereum infrastructure, I authored op-reth PR #23334 for a historical-proof unwind bug originally reported by omerfirmak. An Optimism maintainer reproduced the bug, approved the PR, and merged it into develop on 2026-11-12 at commit 082e041. Taiko's alethia-reth then pinned a commit containing the fix to close #265. As of 2026-11-14, tagged op-reth and alethia-reth releases do not yet include this commit, and the funding-bench PR remains open.

Open to research grants, collaborations, and scoped evaluation audits. Corporate contracting is available through MIKOTO Co., Ltd. Further code, reports, and validation limits are linked from my GitHub profile.

AI tools assisted implementation, experiments, analysis, and writing. Linked evidence distinguishes local findings from upstream confirmation.