Implements `parse_judge_output()` and `extract_choice_answer()` — robust multi-strategy parsers for extracting structured scores and choices from LLM judge outputs. Uses cascading extraction methods (markdown code blocks, brace matching, JSON repair) to handle imperfect model outputs. Includes `_remove_thinking_tags()` (re-exported from output_parser) to strip chain-of-thought content. The choice extraction handles \boxed{} patterns and fallback text matching. This module is self-contained, has no internal dependencies beyond `task_configs`, and could be extracted as a standalone library for any project needing to parse structured data from LLM text completions.

Provides `_remove_thinking_tags()` and `parse_model_output()` — utilities for cleaning and extracting structured content from model-generated text with XML-like tags. The `_remove_thinking_tags` function is particularly sophisticated: it handles nested, orphaned, and multiple think/response tag blocks with three-pass processing (paired matching, orphan closing, orphan opening). This is a general-purpose text processing utility that solves a common problem in LLM workflows (separating reasoning from final answer) with edge-case handling that goes far beyond simple regex.