This work is done with AI assistance — primarily Claude (Anthropic), via Claude Code — used as a pair-programmer and research assistant across code, experiments, analysis, and writing.
What that means here, concretely:
The research is human-directed. I choose the questions, the hypotheses, and what counts as a result. Every experiment is pre-registered by me before I see the outcome.
Every number is real and independently run. No figure in this portfolio is AI-estimated or asserted without a script I executed and can reproduce. Negative and boundary results are reported the same way as the positive ones — that discipline is the point of the work, and it is not delegated.
I review and own every line. AI-generated code is read, understood, and tested by me before it ships, and I can explain any of it. The same goes for prose, and for any upstream bug reports or pull requests.
Upstream contributions to third-party projects disclose AI assistance per that
project’s policy (an Assisted-by: trailer, a PR-template checkbox, or a sentence in the
PR body), and I only file a fix I have reproduced and can defend.
If you find something that looks machine-generated and wrong, that is on me — open an issue.
— Gavin Branaa