Stateless utility class that transforms LLM conversation traces (ChatCompletionMessageParam) into a formatted string suitable for judge consumption. Handles edge cases like reasoning content, tool calls, structured output, and tool results in a single message, emitting separate blocks to avoid silent data loss. Demonstrates careful design for trace fidelity and can be reused in any LLM evaluation pipeline.

Implements the largest-remainder method to convert weighted entries (e.g., percentages for train/test/val splits) into integer-percent allocations that sum to 100. Clean, well-tested algorithm that can be extracted for any project requiring proportional allocation with integer constraints.

Generic async job runner with configurable concurrency, full-jitter retries, progress reporting via an observer pattern, and clean cancellation. Provides a reusable foundation for any batch processing or background job system, decoupled from the specific domain.

Token-level translation of citation spans through auto-reformatting, with mark_html_range for highlight.js integration. Solves the non-trivial problem of maintaining span accuracy when JSON is reformatted (e.g., pretty-printed), ensuring that highlighted citations remain correct. A reusable utility for any application dealing with annotated JSON spans.

Statistical utilities for paired comparison of evaluation results, including McNemar exact/chi2 tests, Wilcoxon signed-rank test, bootstrap confidence intervals, and Wilson CIs for proportions. Provides a reusable toolkit for A/B testing or model comparison scenarios, independent of the evaluation framework.