Independent research engineer focused on one lane: reproducible AI-claim verification for high-stakes decisions.
I build inspectable evaluations that test whether AI, statistical, benchmark, and performance claims survive reruns, corrected comparators, held-out checks, provenance checks, and adversarial scrutiny. The operating rule is simple: AI proposes; deterministic systems verify; evidence remains traceable.
- RCV-Bench — research-claim verification benchmark for autonomous agents. It separates simple re-execution from harder provenance and robustness judgments, with deterministic baselines and a first real-agent baseline.
- Eval Invariance Engine — reusable tooling for testing whether evaluation scores remain stable under semantics-preserving perturbations, with CLI and Inspect AI integration.
- MMLU Robustness Audit — separately regenerated option-order fragility; the central majority-flip result survived while several historical calibration/stability quantities did not.
- De-Stress Lab — selection-aware scientific stress testing with installable software, frozen manifests, reproducibility tooling, and explicit evidence limits.
- ARC-AGI-2 Occam Baseline — corrected same-holdout calibration and selection analysis, including a published reversal of the earlier optimistic result.
The next threshold is outside use, not more self-authored projects. I welcome:
- independent reruns of the MMLU audit or de-stress package;
- outside agents evaluated on RCV-Bench under a declared tool/network/isolation policy;
- outside Inspect users for Eval Invariance Engine;
- technical review, issue reports, and reproducibility PRs.
The standard is: test the claim that matters, preserve the receipts, publish the correction when the evidence changes, and make the surviving result rerunnable.