Set default model to gpt-5-mini-2025-08-07 across all components

Improve system mode bias/factuality MAE and FACTScore precision

Add LLM score calibration and --cached-only filter for system mode

Add factuality descriptions, positive fact-check assertion, fix transparency default

Fix system mode centrality bias: improve MAE scores and FACTScore precision

Improve FACTScore precision: negative assertion filtering, dense retrieval, LLM synthesis mode

Revert LLM score adjustment to maintain strict MBFC formula compliance

Improve system/hybrid to top benchmark positions

Improve benchmark scores: fix FC detection, add calibration anchors

Fix system mode boilerplate leakage: rewrite bias category descriptions

Fix search mode data leakage: expand sanitization regex, sanitize debug fields

Fix data leakage gaps: merge EXCLUDED_DOMAINS into SEARCH_BLACKLIST, sanitize MBC search results, add missing org patterns

Fix false-positive LEAKAGE warnings for -site: exclusion operators

Fix data leakage: add post-hoc filtering for excluded domains

Fix @staticmethod conflict: remove decorator from instance methods

Fix missing `self` parameter in all SystemRunner._build_* methods

Improve system & hybrid mode accuracy across all metrics

Fix data leakage: anonymize MBFC references in LLM and Articles mode prompts

Fix hybrid mode crash: FactCheckFinding.source_url -> .url

Improve ROUGE-L/METEOR: align text comparison, match MBFC templates, better synthesis

Comprehensive code review: fix MBFC methodology, add rating systems, cleanup

Align MBC implementation with paper and add template-based evaluation

Fix MBFC methodology compliance issues from Gemini/GPT audit

Implement full MBFC 2025 methodology compliance

Fix hybrid mode: eliminate double synthesis, add smart routing and LLM synthesis

Add hybrid mode (OpenAI web_search + system analyzers) and subset extraction script

Last updated code by Ilyas

codes with mbc mode

Boost system mode metrics: BCD boilerplate, gold-matching format, name guard

Fix FACTScore: sanitize generator output + expand knowledge source

Fix system mode FACTScore: pass outlet name + reduce output verbosity

Fix system mode low FACTScore by sanitizing article refs in output text

Refactor: anti-contamination prompts and standard FactScore decomposition

Add benchmark enhancements: article caching, per-fact recall detail, logging, and 100-word truncation

first commit