RSS/parameter configuration UI, feedparser migration, effective-params logging

Set FAKE_THRESHOLD=0.75, benchmark results P=0.909 R=0.227 F1=0.364 Acc=0.650 FP=1 FN=34

threshold sweep + TF-IDF/RF/SVM baselines, event-date KG +64 canonical events (1169->1233 facts), V8 action-before-office check, _check_event_date() in external.py

DATE_TOLERANCE 400->200, cross-entity overlap check with same_role_category filter, Reference KG +22 manual facts (Nixon/Ford/LBJ/JFK/Truman/Blair/Thatcher/Starmer/Macron/Tusk + legislation) — F1 0.345->0.393, Recall 0.227->0.273, Sep +0.129->+0.151

RSS verification panel, benchmark split (100/150), remove Ollama/REBEL from frontend, sync Qwen3 labels

Some fix

Replace Ollama/llama3 with spacy-llm + Qwen3-1.7B (Pipeline B + Explainer), archive REBEL (F1=0.038), fix TCS n_claims=0 score, add ISOT/RAGuard eval scripts, update start.sh and requirements

RSS feeds updated + Reference KG 517->1161 facts + entity_label fix + Neo4j reconnect

Refactor

Testes + refactor

Testes + refactor

Wikidata inverse, Neo4j cache decoupled, auto-title, cross-article button, batch checkboxes, threshold 0.55

Add new articles and script for liar dataset

Entity subject matching threshold + V5 HOLDS_POSITION only + KG tolerance 400d

Extended benchmark 25 articles + run_evaluation.py with LIAR/FakeNewsNet/boxplot

Manual benchmark 5 articles + run_benchmark.py with Precision/Recall/F1

Add compare_pipelines.py - batch evaluation A vs B

Sprint 1: project structure