YouTube extraction: normalize NULL/spaced tickers + asset-specific filter

Two data-quality fixes surfaced by inspecting the first 94 claims:
- _clean_ticker: coerce the literal 'NULL'/'NONE'/'$AAPL' strings the LLM sometimes returns to
  a real None, and collapse ASR space-mangling ('B VC' -> 'BVC').
- require_asset filter: drop claims with neither a ticker nor an asset_name -- generic trading
  commentary ('buy the dip') from tutorial content is not an asset claim. On the backlog this
  cut 94 -> 58 claims (36 generic), leaving clean per-asset signal.
Added the $2,000 Small Account Challenge recap playlist (the daily per-asset trade breakdowns) --
far higher claim density than the evergreen 'Anatomy of a Trade' tutorials. 7 tests green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/unit/test_youtube_extraction.py

@@ -1,7 +1,8 @@
 """Pure-helper tests for the claim extractor — the anti-hallucination quote match + chunking.
 The Ollama call itself is integration-tested live (not mocked here)."""
 
-from youtube_claims.extraction import _chunks, _norm, _normalize, _quote_supported
+from youtube_claims.extraction import (
+    _chunks, _clean_ticker, _norm, _normalize, _quote_supported)
 
 
 def test_norm_collapses_whitespace():
@@ -26,6 +27,23 @@ def test_normalize_drops_claim_without_verbatim_quote():
     assert ok and ok["asset_ticker"] == "AAPL" and ok["timestamp_start"] == 0
 
 
+def test_clean_ticker_coerces_nullish_and_spacing():
+    assert _clean_ticker("NULL") is None and _clean_ticker("none") is None
+    assert _clean_ticker("$AAPL") == "AAPL"
+    assert _clean_ticker("B VC") == "BVC"          # ASR space-mangling
+    assert _clean_ticker("") is None
+
+
+def test_asset_specific_filter_drops_generic_commentary():
+    # a "NULL"-ticker generic tutorial line with no asset is dropped
+    generic = {"asset_ticker": "NULL", "asset_name": None, "verbatim_quote": "buy the dip"}
+    assert _normalize(generic, 0, 6) is None
+    # a real asset claim survives, ticker cleaned
+    real = {"asset_ticker": "null", "asset_name": "Apple", "verbatim_quote": "apple broke out"}
+    keep = _normalize(real, 0, 6)
+    assert keep and keep["asset_ticker"] is None and keep["asset_name"] == "Apple"
+
+
 def test_chunks_group_segments_with_time_span():
     segs = [{"start": float(i), "end": float(i + 1), "text": "x" * 100} for i in range(200)]
     wins = list(_chunks(segs, max_chars=1000))

youtube_claims/extraction.py

@@ -80,14 +80,35 @@ def _chunks(segments: list[dict], max_chars: int = 8000):
         yield buf, start, buf[-1][0].get("end")
 
 
-def _normalize(c: dict, win_start, win_end) -> dict | None:
+_NULLISH = {"", "NULL", "NONE", "N/A", "NA", "-", "--", "UNKNOWN", "TBD", "IT", "THE STOCK"}
+
+
+def _clean_ticker(raw) -> str | None:
+    """The LLM sometimes returns the literal string 'NULL' or a space-mangled ASR ticker
+    ('B VC'); coerce those to a real None / contiguous symbol."""
+    t = (raw or "").strip().upper().lstrip("$")
+    if t in _NULLISH:
+        return None
+    t = t.replace(" ", "")               # 'B VC' -> 'BVC' (tickers are single tokens)
+    return t or None
+
+
+def _clean_name(raw) -> str | None:
+    n = (raw or "").strip()
+    return None if (not n or n.upper() in _NULLISH) else n
+
+
+def _normalize(c: dict, win_start, win_end, require_asset: bool = True) -> dict | None:
     vq = (c.get("verbatim_quote") or "").strip()
     if not vq:
         return None                      # mandatory — drop unsupported claims
-    tkr = (c.get("asset_ticker") or "").strip().upper() or None
+    tkr = _clean_ticker(c.get("asset_ticker"))
+    name = _clean_name(c.get("asset_name"))
+    if require_asset and not tkr and not name:
+        return None                      # not an asset-specific claim (generic commentary)
     return {
         "asset_ticker": tkr,
-        "asset_name": (c.get("asset_name") or None),
+        "asset_name": name,
         "asset_class": (c.get("asset_class") or None),
         "direction": (c.get("direction") or None),
         "claim_text": (c.get("claim_text") or None),