Case Study 4: Clinical Safety Benchmark
Five fictional patient scenarios with hidden safety traps, 5 models, 25 responses. Found that a visible rubric inflated scores, then reran clean.
Read the case study →Portfolio
AI evaluation work
Benchmarks I designed, ran and scored, with written reasoning for every call.
What I do
Rate & rank responses
Side-by-side comparisons on accuracy, helpfulness, tone, instruction following and safety, each with a written justification.
Catch what sounds right but isn't
Hallucinations, made-up citations and logic slips, checked against primary sources rather than trusted on sight.
Clinical safety review
15+ years of nursing judgment in triage, medication safety, dialysis, cardiac, psychiatric and substance use care, applied to how AI handles health questions.
Case studies
Benchmarks
Case Study 1: Live Research & Citation
How each model finds, cites and verifies sources, and where it invents them. From chapter 4 of the book.
In the book →Case Study 2: Voice & Analogy
Whether each model can hold an author's voice and build a fresh analogy without drifting into filler.
In the book →Case Study 3: Editorial & Logic
Developmental editing and structural reasoning: which models catch logic gaps and which smooth them over.
In the book →How I score
Every response is scored against a written rubric on safety, accuracy, helpfulness and instruction following, and every score gets a short written reason. I test each prompt in a fresh chat with the prompt text only, record the model and version shown at test time, and rerun when test conditions could have skewed the result.