feat(manifold): add outcome compatibility guard and conditional market rejection

Reject false-positive matches where Jaccard overlap is high but the outcome is
not equivalent (e.g. Poly nomination vs Manifold "If X is nominee, will he win").

- _is_conditional(): detect conditional Manifold markets (If/Conditional on/
  Assuming/Given that prefixes + mid-sentence " if ...," clauses) -> reject with
  reason "conditional_market".
- _classify_outcome(): classify into nomination|primary_win|general_win|
  conditional|other; reject when poly/mfld types differ or either is conditional
  -> reason "outcome_mismatch: poly=... manifold=...".
- Persist poly_outcome_type/mfld_outcome_type on ManifoldMatchResult, in
  manifold_match_audit (CREATE + idempotent ALTER), save_manifold_audit() and
  the bayesian call site.
- Tests covering classification, conditional detection and the Graham Platner
  regression (now rejected); valid nomination<->nomination still accepted.

Untouched: _MATCH_THRESHOLD (0.40), MANIFOLD_LOGODDS_WEIGHT, edge thresholds,
exposure, trading logic.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

bot/data/db.py

@@ -530,6 +530,8 @@ class Database:
         match_score: Optional[float],
         match_reason: Optional[str],
         match_status: str,
+        poly_outcome_type: Optional[str] = None,
+        mfld_outcome_type: Optional[str] = None,
     ) -> None:
         async with self._pool.acquire() as conn:
             await conn.execute("""
@@ -537,13 +539,15 @@ class Database:
                     id, poly_market_id, poly_question, search_query,
                     mfld_market_id, mfld_market_title, mfld_market_url,
                     prob_raw, prob_final, inverted,
-                    match_score, match_reason, match_status, used_in_trade
-                ) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,FALSE)
+                    match_score, match_reason, match_status, used_in_trade,
+                    poly_outcome_type, mfld_outcome_type
+                ) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,FALSE,$14,$15)
             """,
                 audit_id, poly_market_id, poly_question, search_query,
                 mfld_market_id, mfld_market_title, mfld_market_url,
                 prob_raw, prob_final, inverted,
                 match_score, match_reason, match_status,
+                poly_outcome_type, mfld_outcome_type,
             )
 
     async def mark_manifold_audit_used(self, audit_id: str) -> None:

bot/data/manifold.py

@@ -6,6 +6,15 @@ by keyword overlap and returns a ManifoldMatchResult with full audit metadata.
 
 Match threshold: >= 0.40 Jaccard overlap (raised from 0.25 for stricter semantics).
 
+Outcome compatibility guard (conservative):
+  - Conditional Manifold markets ("If X, will Y?" / "Conditional on..." / "Assuming..."
+    / "Given that..." / mid-sentence "...if X is nominated, will...") are rejected:
+    a premise-gated question is not equivalent to a direct outcome question even when
+    token overlap is high. reason='conditional_market'.
+  - Each side is classified into an outcome_type (nomination | primary_win |
+    general_win | conditional | other). Matches with differing outcome_type — or any
+    conditional side — are rejected. reason='outcome_mismatch: poly=... manifold=...'.
+
 Inversion guard (conservative):
   - If Polymarket question names a party (democrat/republican) AND the matched
     Manifold market names the OPPOSITE party → invert probability (1 - prob).
@@ -59,6 +68,8 @@ class ManifoldMatchResult:
     match_reason: Optional[str] = None   # human-readable explanation
     inverted: bool = False
     search_query: str = ""
+    poly_outcome_type: Optional[str] = None  # nomination|primary_win|general_win|conditional|other
+    mfld_outcome_type: Optional[str] = None
 
 
 def _significant_words(text: str) -> set[str]:
@@ -82,6 +93,45 @@ def _detect_party(text: str) -> Optional[str]:
     return None
 
 
+# ── Conditional-market detection (Task 1) ──────────────────────────────────────
+# A market is "conditional" when its resolution is gated on a premise rather than
+# asking the outcome directly (e.g. "If X is the nominee, will he win?"). Such a
+# market is NOT equivalent to a direct outcome question even with high token overlap.
+_CONDITIONAL_PREFIXES = ("if ", "conditional on", "assuming ", "given that")
+# " if <clause>," — a mid-sentence conditional clause closed by a comma.
+_CONDITIONAL_CLAUSE_RE = re.compile(r"\sif\s[^,]*,")
+
+
+def _is_conditional(text: str) -> bool:
+    """True if the question is phrased conditionally (premise-gated)."""
+    t = (text or "").strip().lower()
+    if t.startswith(_CONDITIONAL_PREFIXES):
+        return True
+    return bool(_CONDITIONAL_CLAUSE_RE.search(t))
+
+
+def _classify_outcome(text: str) -> str:
+    """
+    Coarse classification of what a question is *asking about*, used to reject
+    matches whose outcomes are not equivalent even when tokens overlap.
+
+    Returns one of: nomination | primary_win | general_win | conditional | other.
+    Order matters: conditional is checked first (premise-gated), then nomination
+    (which subsumes "primary nominee"), then primary, then general election.
+    """
+    t = (text or "").strip().lower()
+    if t.startswith(_CONDITIONAL_PREFIXES):
+        return "conditional"
+    if any(k in t for k in ("nominee", "nominated", "nomination")):
+        return "nomination"
+    if any(k in t for k in ("primary", "win the primary", "first round")):
+        return "primary_win"
+    if any(k in t for k in ("win the election", "win the race",
+                            "win the seat", "general election")):
+        return "general_win"
+    return "other"
+
+
 def _find_best_candidate(poly_question: str, results: list[dict]) -> tuple[Optional[dict], float]:
     """Find the highest-scoring open binary Manifold market by Jaccard overlap."""
     poly_words = _significant_words(poly_question)
@@ -137,9 +187,14 @@ class ManifoldClient:
         if cached and (now - cached[0]) < CACHE_TTL_SEC:
             return cached[1]
 
+        poly_outcome = _classify_outcome(question)
+
         query = _build_search_query(question)
         if not query:
-            result = ManifoldMatchResult(status="no_results", search_query="")
+            result = ManifoldMatchResult(
+                status="no_results", search_query="",
+                poly_outcome_type=poly_outcome,
+            )
             self._cache[question] = (now, result)
             return result
 
@@ -152,12 +207,18 @@ class ManifoldClient:
             results = resp.json()
         except Exception as exc:
             log.warning("Manifold API error for %r: %s", question[:40], exc)
-            result = ManifoldMatchResult(status="no_results", search_query=query)
+            result = ManifoldMatchResult(
+                status="no_results", search_query=query,
+                poly_outcome_type=poly_outcome,
+            )
             self._cache[question] = (now, result)
             return result
 
         if not results:
-            result = ManifoldMatchResult(status="no_results", search_query=query)
+            result = ManifoldMatchResult(
+                status="no_results", search_query=query,
+                poly_outcome_type=poly_outcome,
+            )
             self._cache[question] = (now, result)
             return result
 
@@ -176,22 +237,36 @@ class ManifoldClient:
                 match_score=score if best else None,
                 match_reason=reason,
                 search_query=query,
+                poly_outcome_type=poly_outcome,
+                mfld_outcome_type=_classify_outcome(best.get("question", "")) if best else None,
             )
             self._cache[question] = (now, result)
             return result
 
-        # ── Inversion analysis (conservative) ────────────────────────────────
+        # ── Outcome compatibility + inversion analysis (conservative) ─────────
+        mfld_title     = best.get("question", "")
+        mfld_outcome   = _classify_outcome(mfld_title)
         poly_party     = _detect_party(question)
-        manifold_party = _detect_party(best.get("question", ""))
+        manifold_party = _detect_party(mfld_title)
 
         poly_words    = _significant_words(question)
-        mfld_words    = _significant_words(best.get("question", ""))
+        mfld_words    = _significant_words(mfld_title)
         matched_tokens = sorted(poly_words & mfld_words)[:6]
 
         inverted = False
         rejection_reason: Optional[str] = None
 
-        if poly_party is not None:
+        # Task 1 — conditional Manifold market is never equivalent to a direct
+        # outcome question, regardless of token overlap.
+        if _is_conditional(mfld_title):
+            rejection_reason = "conditional_market: manifold question is conditional"
+        # Task 2 — outcome types must match; any conditional side is rejected.
+        elif (poly_outcome == "conditional" or mfld_outcome == "conditional"
+              or poly_outcome != mfld_outcome):
+            rejection_reason = (
+                f"outcome_mismatch: poly={poly_outcome} manifold={mfld_outcome}"
+            )
+        elif poly_party is not None:
             if manifold_party is None:
                 # Poly specifies a party; Manifold does not → can't verify inversion safety
                 rejection_reason = (
@@ -219,6 +294,8 @@ class ManifoldClient:
                     f"jaccard={score:.2f}, tokens={matched_tokens}, {rejection_reason}"
                 ),
                 search_query=query,
+                poly_outcome_type=poly_outcome,
+                mfld_outcome_type=mfld_outcome,
             )
             self._cache[question] = (now, result)
             return result
@@ -257,6 +334,8 @@ class ManifoldClient:
             match_reason=match_reason,
             inverted=inverted,
             search_query=query,
+            poly_outcome_type=poly_outcome,
+            mfld_outcome_type=mfld_outcome,
         )
         self._cache[question] = (now, result)
         return result

bot/data/schema.sql

@@ -207,13 +207,19 @@ CREATE TABLE IF NOT EXISTS manifold_match_audit (
     match_score       DOUBLE PRECISION,
     match_reason      TEXT,
     match_status      TEXT NOT NULL,
-    used_in_trade     BOOLEAN DEFAULT FALSE
+    used_in_trade     BOOLEAN DEFAULT FALSE,
+    poly_outcome_type TEXT,
+    mfld_outcome_type TEXT
 );
 
 CREATE INDEX IF NOT EXISTS idx_mfld_audit_timestamp ON manifold_match_audit(timestamp DESC);
 CREATE INDEX IF NOT EXISTS idx_mfld_audit_status    ON manifold_match_audit(match_status);
 CREATE INDEX IF NOT EXISTS idx_mfld_audit_poly_mkt  ON manifold_match_audit(poly_market_id);
 
+-- Backfill outcome-type columns on pre-existing tables (idempotent).
+ALTER TABLE manifold_match_audit ADD COLUMN IF NOT EXISTS poly_outcome_type TEXT;
+ALTER TABLE manifold_match_audit ADD COLUMN IF NOT EXISTS mfld_outcome_type TEXT;
+
 -- ─────────────────────────────────────────────────────────────────────────────
 -- Metric exclusion — administrative closure flag
 --

bot/strategy/bayesian.py

@@ -470,6 +470,8 @@ class BayesianStrategy:
                             match_score=manifold_result.match_score,
                             match_reason=manifold_result.match_reason,
                             match_status=manifold_result.status,
+                            poly_outcome_type=manifold_result.poly_outcome_type,
+                            mfld_outcome_type=manifold_result.mfld_outcome_type,
                         )
                     except Exception as exc:
                         log.warning("Failed to save manifold audit: %s", exc)

tests/test_manifold_outcome.py

@@ -0,0 +1,152 @@
+"""
+Tests for the Manifold outcome-compatibility guard.
+
+Regression: a Polymarket *nomination* question must not match a Manifold
+*conditional* question ("If X is the nominee, will he win?") even at Jaccard=1.0.
+"""
+import asyncio
+
+import pytest
+
+from bot.data.manifold import (
+    ManifoldClient,
+    _classify_outcome,
+    _is_conditional,
+)
+
+
+# ── _is_conditional ────────────────────────────────────────────────────────────
+def test_is_conditional_prefixes():
+    assert _is_conditional("If Graham Platner is the nominee, will he win?")
+    assert _is_conditional("Conditional on a recession, will rates fall?")
+    assert _is_conditional("Assuming Trump runs, will he win?")
+    assert _is_conditional("Given that X happens, will Y?")
+
+
+def test_is_conditional_midsentence_clause():
+    assert _is_conditional("Will Biden, if he is nominated, win the election?")
+
+
+def test_is_not_conditional():
+    assert not _is_conditional("Will Graham Platner be the Democratic nominee?")
+    assert not _is_conditional("Will the GOP win the Senate?")
+    # "if" without a closing comma clause is not flagged
+    assert not _is_conditional("What happens if everything goes right")
+
+
+# ── _classify_outcome ───────────────────────────────────────────────────────────
+def test_classify_nomination():
+    assert _classify_outcome("Will X be the Democratic nominee for Senate?") == "nomination"
+    assert _classify_outcome("Will X be nominated?") == "nomination"
+    # "primary nominee" → nomination (checked before primary)
+    assert _classify_outcome("Will X be the primary nominee?") == "nomination"
+
+
+def test_classify_primary_win():
+    assert _classify_outcome("Will X win the primary?") == "primary_win"
+    assert _classify_outcome("Will X advance in the first round?") == "primary_win"
+
+
+def test_classify_general_win():
+    assert _classify_outcome("Will X win the election?") == "general_win"
+    assert _classify_outcome("Will X win the seat?") == "general_win"
+    assert _classify_outcome("Will X win the general election?") == "general_win"
+
+
+def test_classify_conditional():
+    assert _classify_outcome("If X is the nominee, will he win?") == "conditional"
+    assert _classify_outcome("Assuming a runoff, who wins?") == "conditional"
+
+
+def test_classify_other():
+    assert _classify_outcome("Will it rain tomorrow?") == "other"
+
+
+# ── End-to-end get_match with a stubbed Manifold API ────────────────────────────
+class _StubResponse:
+    def __init__(self, payload):
+        self._payload = payload
+
+    def raise_for_status(self):
+        pass
+
+    def json(self):
+        return self._payload
+
+
+class _StubHTTP:
+    def __init__(self, payload):
+        self._payload = payload
+
+    async def get(self, *args, **kwargs):
+        return _StubResponse(self._payload)
+
+    async def aclose(self):
+        pass
+
+
+async def _match(poly, mfld_market):
+    client = ManifoldClient()
+    client._client = _StubHTTP([mfld_market])
+    try:
+        return await client.get_match(poly)
+    finally:
+        await client.close()
+
+
+def test_graham_platner_conditional_rejected():
+    """Poly nomination vs Manifold conditional → rejected (Task 4.1)."""
+    poly = ("Will Graham Platner be the Democratic nominee for Senate "
+            "in Maine in 2026?")
+    mfld_market = {
+        "outcomeType": "BINARY",
+        "probability": 0.55,
+        "question": ("If Graham Platner is the Democratic nominee for Senate "
+                     "in Maine, will he win the general election?"),
+        "id": "abc123",
+        "slug": "graham-platner-win",
+        "creatorUsername": "someone",
+    }
+    result = asyncio.run(_match(poly, mfld_market))
+
+    assert result.status == "rejected"
+    assert result.match_reason is not None
+    assert ("conditional" in result.match_reason
+            or "outcome_mismatch" in result.match_reason)
+    # outcome types are classified and available for persistence
+    assert result.poly_outcome_type == "nomination"
+    assert result.mfld_outcome_type == "conditional"
+
+
+def test_outcome_mismatch_nomination_vs_general_rejected():
+    """Poly nomination vs Manifold general_win (non-conditional) → rejected."""
+    poly = "Will Jane Doe be the Republican nominee for Governor?"
+    mfld_market = {
+        "outcomeType": "BINARY",
+        "probability": 0.4,
+        "question": "Will Jane Doe win the election for Governor?",
+        "id": "x", "slug": "jane-doe", "creatorUsername": "u",
+    }
+    result = asyncio.run(_match(poly, mfld_market))
+
+    assert result.status == "rejected"
+    assert "outcome_mismatch" in result.match_reason
+    assert result.poly_outcome_type == "nomination"
+    assert result.mfld_outcome_type == "general_win"
+
+
+def test_matching_nomination_accepted():
+    """Poly nomination vs Manifold nomination (same outcome) → accepted."""
+    poly = "Will Graham Platner be the Democratic nominee for Senate in Maine?"
+    mfld_market = {
+        "outcomeType": "BINARY",
+        "probability": 0.62,
+        "question": "Will Graham Platner be the Democratic Senate nominee in Maine?",
+        "id": "ok", "slug": "platner-nominee", "creatorUsername": "u",
+    }
+    result = asyncio.run(_match(poly, mfld_market))
+
+    assert result.status == "accepted"
+    assert result.poly_outcome_type == "nomination"
+    assert result.mfld_outcome_type == "nomination"
+    assert result.prob_final == pytest.approx(0.62)