feat(nlp): replace AnthropicEventExtractor with RuleBasedEventExtractor

collectors/nlp/__init__.py

@@ -18,23 +18,28 @@
 
 Today's free-tier implementations:
 
-  loughran_mcdonald.LoughranMcDonaldScorer — finance-domain dictionary
-                                              sentiment, the academic
-                                              gold standard
-  event_extraction.AnthropicEventExtractor  — Haiku-tier structured
-                                              event flag extraction
-                                              (we already pay Anthropic)
+  loughran_mcdonald.LoughranMcDonaldScorer            — finance-domain
+                                                        dictionary sentiment
+                                                        (academic standard)
+  rule_based_event_extraction.RuleBasedEventExtractor — deterministic event
+                                                        classification from
+                                                        vendor tags (Polygon
+                                                        keywords, GDELT codes)
+                                                        + title-keyword regex.
+                                                        Replaced the Haiku-
+                                                        backed
+                                                        AnthropicEventExtractor
+                                                        2026-05-25 per the
+                                                        "LLM calls confined
+                                                        to research module"
+                                                        architectural rule.
 
 Heavier free upgrades that drop in as new adapter classes (Phase 3+):
 
   finbert.FinBERTScorer        — HF yiyanghkust/finbert-tone
   spacy_ner.SpacyEntityExtractor — en_core_web_sm or larger
 """
 
-from collectors.nlp.event_extraction import (
-    AnthropicEventExtractor,
-    DEFAULT_EVENT_CATEGORIES,
-)
 from collectors.nlp.loughran_mcdonald import (
     LoughranMcDonaldScorer,
     load_lm_master_dict,
@@ -48,6 +53,10 @@
     SentimentScore,
     SentimentScorer,
 )
+from collectors.nlp.rule_based_event_extraction import (
+    DEFAULT_EVENT_CATEGORIES,
+    RuleBasedEventExtractor,
+)
 
 __all__ = [
     "EntityMention",
@@ -58,7 +67,7 @@
     "SentimentScorer",
     "LoughranMcDonaldScorer",
     "load_lm_master_dict",
-    "AnthropicEventExtractor",
+    "RuleBasedEventExtractor",
     "DEFAULT_EVENT_CATEGORIES",
     "NewsNLPPipeline",
 ]

collectors/nlp/pipeline.py

@@ -117,11 +117,20 @@ def process(
                         extractor.name, fp, e,
                     )
 
+            # Union vendor tags across all variants — Polygon keywords +
+            # GDELT event codes + Benzinga channels for the same wire
+            # story. Rule-based extractors use this as the primary
+            # classification signal; LLM extractors (if any reactivated)
+            # ignore the kwarg via the EventExtractor Protocol default.
+            article_tags: tuple[str, ...] = tuple({
+                t for v in article.variants for t in v.tags
+            })
             for extractor in self._event_extractors:
                 try:
                     events.extend(extractor.extract(
                         text=text, article_fingerprint=fp,
                         article_tickers=article.tickers,
+                        article_tags=article_tags,
                     ))
                 except Exception as e:
                     logger.warning(

collectors/nlp/protocols.py

@@ -170,4 +170,5 @@ class EventExtractor(Protocol):
     def extract(
         self, *, text: str, article_fingerprint: str,
         article_tickers: tuple[str, ...],
+        article_tags: tuple[str, ...] = (),
     ) -> list[EventFlag]: ...