On Measuring Scientific Faithfulness in Language Models
A practical note on why answer correctness and citation faithfulness should be evaluated separately for technical domains.
Notes
Short-form writing on methods, experiments, and open problems. Use this section to publish interim insights between full papers.
When Better Loss Curves Hide Worse Calibration
Evidence from synthetic and real biomedical datasets showing optimization gains can mask reliability regressions.
Designing Benchmarks That Survive Distribution Shift
Recommendations for benchmark construction that retain discriminatory power under realistic data drift.
A Minimal Protocol for Reproducible Ablation Studies
A checklist for writing ablation sections that remain reproducible after repository evolution.