Refactor Deep-Research Evaluation to Use YESciEval (Remove Legacy LLM-as-a-Judge Code)

[jd-coderepos]

Description

The scripts/ directory currently contains legacy LLM-as-a-Judge evaluation code that has already been consolidated and extended in YESciEval. To avoid duplication and ensure consistency across projects, this repository should be refactored to depend on YESciEval for all evaluation logic.

Tasks

• Review the contents of
  scripts/
  https://github.com/sciknoworg/deep-research/tree/main/scripts
• Delete all legacy LLM-as-a-Judge evaluation implementations that are now part of YESciEval
• Identify evaluation scripts that should remain (e.g., orchestration, report-level execution)
• Refactor remaining evaluation scripts to:
  • Import evaluation rubrics, scoring logic, and metrics from the YESciEval library
  • Avoid re-implementing judge prompts, criteria, or aggregation logic locally
• Ensure that:
  • Evaluation is still supported per individual report
  • Evaluation can still compute aggregated scores over a collection of reports

Acceptance Criteria

• No duplicated LLM-as-a-Judge logic remains in this repository
• Evaluation scripts cleanly depend on YESciEval as an external library
• Report-level and collection-level evaluation workflows remain functional
• Codebase is easier to maintain and aligned with the shared evaluation framework

Rationale

YESciEval is the canonical location for evaluation rubrics and LLM-as-a-Judge logic. This cleanup improves maintainability, avoids divergence, and strengthens reuse across projects.