mlchrzan/pairadigm: [1.0.0] Summer Body

pairadigm has reached its first breaking version! This package started as me just trying to port some code across projects easily but now is becoming a serious open-source project. As such, it required a massive refactoring undertaking that unfortunately did lead to breaking prior pieces, such as saving and loading pairadigms (which is now much more stable!). Excited to share it with you all and to continue to build out the functionality. Best, Michael Here's the formals [1.0.0] - 2026-04-16 - 'Summer Body' Added Safer Saving Logic: Instead of using pickles, pairadigm now saves and loads data using individual parquet files, which are more robust and efficient. This also means that pairadigm objects are now much smaller and faster to load. It also saves the instance construction parameters in a metadata.json file, which is used to reconstruct the object when loading. LLM API Cost Estimation: Added estimate_costs() method in Pairadigm class to estimate token usage and cost using tiktoken or a heuristic fallback prior to large API operations. Client Addition Workflows: Enhanced add_client() logic to allow users to optionally generate breakdowns and pairwise comparisons exclusively for a newly added model. Dawid-Skene Enhancements: dawid_skene_annotator_ranking() now returns the confusion matrix in addition to ranking metrics. Added warnings for 3-class (tied) classifications. Reasoning Models: Introduced support for passing a reasoning level parameter to LLM clients (simplifying the OpenAI wrapper and removing redundant temperature tracking). Base URLs Persistence: Base URLs defined for LLMClient instances are now persisted within persistence.py metadata and seamlessly rehydrated during load_pairadigm(). Updated Unified Breakdowns: Consolidated generate_breakdowns() and generate_breakdowns_from_paired() into a single generate_breakdowns() method. Module-Level Ordinal Logic: Decoupled pair_from_ordinal() from the Pairadigm class into module-level logic. It now processes multi-annotator decision columns robustly. Robust Extraction Regex: Enhanced pairwise_compare() internal parsing logic to be far more forgiving for weaker or local models that deviate from strict formatting. Documentation Overhaul: Complete rewrite of docstrings across all public functions in core.py, providing rich parameters, examples, and researcher-friendly contexts. Column Prefixing: Standardized human_ prefixes for manual annotations to eliminate collisions with the LLM decision_ columns. Fixed AltTest Filtering Constraints: Fixed logic in alt_test() where valid data entries were being erroneously dropped. Dawid-Skene Arguments: Resolved a duplicate kwargs collision error within dawid_skene_annotator_ranking(). Sparse Ordinal Data Bugs: Addressed sparse annotation matrix routing issues when handling multi-annotator datasets without globally shared coverage.