feat(rewrite): formal judge rubric selection and score distribution analysis

[lipikaramaswamy]

Background

The rewrite pipeline's final judge currently uses Option A: privacy, quality, naturalness on a 1–10 scale, adapted from an earlier research repo. PR #86 improved the judge prompt wording and iterated on scoring behavior as an interim fix, but the formal rubric decision was never made. This issue closes that out.

Closes #37.

Options under consideration

Option	Dimensions	Scale	Notes
A (current)	privacy, quality, naturalness	1–10	Baseline; quality and naturalness overlap in practice
B	privacy, utility, naturalness	1–10 with improved anchors	Renames quality → utility for clarity; adds explicit anchor descriptions per score band
C	privacy, utility, faithfulness, fluency	1–5	More granular separation of reservation (faithfulness) vs. readability (fluency); narrower scale may reduce score clustering

Scope of work

• Run score distribution analysis on TAB and NVIDIA Synthetic Biographies using current rubric (Option A) — check for score clustering, ceiling effects, and correlation between dimensions
• Evaluate Options B and C against the same datasets and compare differentiation
• Document chosen rubric with justification and sample score distributions
• Update anchor point descriptions for chosen option
• Update FinalJudgeWorkflow implementation
• Update test coverage

Related

• Supersedes feat(rewrite): judge rubric refinement #37
• See feat: LLM-as-a-judge for REPLACE evaluation #98 for LLM-as-a-judge coverage for REPLACE mode