diff --git a/hindsight-api-slim/hindsight_api/engine/search/reranking.py b/hindsight-api-slim/hindsight_api/engine/search/reranking.py
index 4888400b69..4698626638 100644
--- a/hindsight-api-slim/hindsight_api/engine/search/reranking.py
+++ b/hindsight-api-slim/hindsight_api/engine/search/reranking.py
@@ -212,16 +212,47 @@ async def rerank(self, query: str, candidates: list[MergedCandidate]) -> list[Sc
         # Get cross-encoder scores
         scores = await self.cross_encoder.predict(pairs)
 
-        # Normalize scores using sigmoid to [0, 1] range
-        # Cross-encoder returns logits which can be negative
-        import math
-
+        # Normalize scores to [0, 1] range.
+        # Local models return logits (any real number) — sigmoid is appropriate.
+        # External API rerankers (SiliconFlow, Cohere, etc.) return pre-normalized
+        # relevance_score in [0, 1] with very small absolute values. Applying
+        # sigmoid to these compresses everything to ~0.5, destroying the ranking
+        # signal and making recency the sole sorting factor. We detect the score range
+        # and choose the appropriate normalization.
         import numpy as np
 
-        def sigmoid(x):
+        def _sigmoid(x: float) -> float:
             return 1 / (1 + np.exp(-x))
 
-        normalized_scores = [sigmoid(score) for score in scores]
+        def _rank_normalize_with_ties(score_list: list[float]) -> list[float]:
+            """Rank-based normalization that assigns equal ranks to equal scores."""
+            n = len(score_list)
+            if n <= 1:
+                return [1.0] * n
+            indexed = sorted(enumerate(score_list), key=lambda x: x[1], reverse=True)
+            result = [0.0] * n
+            i = 0
+            while i < n:
+                j = i
+                while j < n and indexed[j][1] == indexed[i][1]:
+                    j += 1
+                # Average rank for tied scores
+                avg_rank = (i + j - 1) / 2.0
+                norm = 1.0 - (avg_rank / (n - 1))
+                for k in range(i, j):
+                    result[indexed[k][0]] = norm
+                i = j
+            return result
+
+        if scores and min(scores) >= 0.0 and max(scores) <= 1.0:
+            # Scores are already in [0, 1] (e.g. SiliconFlow, Cohere relevance_score).
+            # Use rank-based normalization to preserve relative ordering without
+            # depending on absolute score magnitudes.
+            normalized_scores = _rank_normalize_with_ties(scores)
+        else:
+            # Scores are logits (e.g. local sentence-transformers models).
+            # Sigmoid maps (-inf, +inf) to (0, 1).
+            normalized_scores = [_sigmoid(score) for score in scores]
 
         # Create ScoredResult objects with cross-encoder scores
         scored_results = []
diff --git a/hindsight-api-slim/tests/test_reranker_score_normalization.py b/hindsight-api-slim/tests/test_reranker_score_normalization.py
new file mode 100644
index 0000000000..fad4aeac64
--- /dev/null
+++ b/hindsight-api-slim/tests/test_reranker_score_normalization.py
@@ -0,0 +1,163 @@
+"""
+Unit tests for CrossEncoderReranker.rerank() score normalization logic.
+
+Covers:
+1. Rank-based normalization when scores are already in [0, 1].
+2. Tied scores receiving identical normalized values.
+3. Sigmoid normalization when scores are logits outside [0, 1].
+4. Empty candidates returning an empty list without calling predict().
+"""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock
+from uuid import uuid4
+
+import pytest
+
+from hindsight_api.engine.search.reranking import CrossEncoderReranker
+from hindsight_api.engine.search.types import MergedCandidate, RetrievalResult
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _make_candidates(n: int) -> list[MergedCandidate]:
+    """Create *n* minimal MergedCandidate objects."""
+    candidates = []
+    for i in range(n):
+        retrieval = RetrievalResult(
+            id=str(uuid4()),
+            text=f"Document {i}",
+            fact_type="world",
+            occurred_start=None,
+            occurred_end=None,
+        )
+        candidates.append(
+            MergedCandidate(retrieval=retrieval, rrf_score=1.0 / (i + 1))
+        )
+    return candidates
+
+
+def _make_cross_encoder(predict_return: list[float]):
+    """Return a fake cross-encoder whose `predict` is an AsyncMock."""
+    ce = AsyncMock()
+    ce.predict = AsyncMock(return_value=predict_return)
+    ce.provider_name = "local"
+    ce.initialize = AsyncMock()
+    return ce
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_rank_normalization_for_0_1_scores():
+    """Scores already in [0, 1] should use rank-based normalization."""
+    # Scores in [0, 1] but NOT logit-shaped — rank normalization preserves order
+    raw_scores = [0.1, 0.5, 0.9]
+    ce = _make_cross_encoder(raw_scores)
+    reranker = CrossEncoderReranker(cross_encoder=ce)
+    reranker._initialized = True
+
+    candidates = _make_candidates(3)
+    results = await reranker.rerank("test query", candidates)
+
+    assert len(results) == 3
+    # Top-ranked (0.9) gets normalized 1.0
+    assert results[0].cross_encoder_score == pytest.approx(0.9)
+    assert results[0].cross_encoder_score_normalized == pytest.approx(1.0)
+    # Bottom-ranked (0.1) gets normalized 0.0
+    assert results[-1].cross_encoder_score == pytest.approx(0.1)
+    assert results[-1].cross_encoder_score_normalized == pytest.approx(0.0)
+    # Middle gets 0.5
+    assert results[1].cross_encoder_score_normalized == pytest.approx(0.5)
+    # predict was called exactly once
+    ce.predict.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_tied_scores_get_same_normalized_value():
+    """When raw scores contain ties, tied entries must receive the same normalized score."""
+    # Two scores tied at 0.7, one distinct at 0.3
+    raw_scores = [0.7, 0.3, 0.7]
+    ce = _make_cross_encoder(raw_scores)
+    reranker = CrossEncoderReranker(cross_encoder=ce)
+    reranker._initialized = True
+
+    candidates = _make_candidates(3)
+    results = await reranker.rerank("test query", candidates)
+
+    # The two 0.7 entries should share the same normalized value.
+    # With rank-based normalization over 3 items (indices 0,1,2 sorted desc):
+    #   rank 0 & 1 tied → avg_rank = 0.5 → norm = 1 - 0.5/2 = 0.75
+    #   rank 2          → norm = 1 - 2/2 = 0.0
+    # After sorting by normalized score descending, the two 0.7 entries are
+    # at indices 0 and 1; the 0.3 entry is at index 2.
+    scores_by_raw = {r.cross_encoder_score: r.cross_encoder_score_normalized for r in results}
+    tied_norm = scores_by_raw[0.7]
+    assert tied_norm == pytest.approx(0.75)
+    assert scores_by_raw[0.3] == pytest.approx(0.0)
+    # Both 0.7 results have identical normalized scores
+    assert results[0].cross_encoder_score_normalized == pytest.approx(
+        results[1].cross_encoder_score_normalized
+    )
+
+
+@pytest.mark.asyncio
+async def test_sigmoid_normalization_for_logits():
+    """When scores are outside [0, 1] (logits), sigmoid normalization is used."""
+    raw_scores = [2.0, -1.0, 0.0]
+    ce = _make_cross_encoder(raw_scores)
+    reranker = CrossEncoderReranker(cross_encoder=ce)
+    reranker._initialized = True
+
+    candidates = _make_candidates(3)
+    results = await reranker.rerank("test query", candidates)
+
+    assert len(results) == 3
+    import math
+
+    # Results are sorted by weight descending
+    expected_sigmoid = [1 / (1 + math.exp(-s)) for s in raw_scores]
+    expected_sorted = sorted(expected_sigmoid, reverse=True)
+
+    for result, expected in zip(results, expected_sorted):
+        assert result.cross_encoder_score_normalized == pytest.approx(expected, rel=1e-6)
+
+    # Verify the highest logit (2.0) maps to the highest normalized score
+    assert results[0].cross_encoder_score == pytest.approx(2.0)
+    assert results[0].cross_encoder_score_normalized > 0.5
+
+
+@pytest.mark.asyncio
+async def test_empty_candidates_returns_empty_without_predict():
+    """When candidates are empty, rerank must return [] without calling predict()."""
+    ce = _make_cross_encoder([])
+    reranker = CrossEncoderReranker(cross_encoder=ce)
+    reranker._initialized = True
+
+    results = await reranker.rerank("test query", [])
+
+    assert results == []
+    ce.predict.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_mixed_boundary_scores_use_rank():
+    """Boundary values 0.0 and 1.0 (still in [0,1]) should trigger rank normalization."""
+    raw_scores = [0.0, 1.0, 0.5]
+    ce = _make_cross_encoder(raw_scores)
+    reranker = CrossEncoderReranker(cross_encoder=ce)
+    reranker._initialized = True
+
+    candidates = _make_candidates(3)
+    results = await reranker.rerank("test query", candidates)
+
+    # Rank-based: 1.0 → 1.0, 0.5 → 0.5, 0.0 → 0.0
+    by_score = {r.cross_encoder_score: r.cross_encoder_score_normalized for r in results}
+    assert by_score[1.0] == pytest.approx(1.0)
+    assert by_score[0.5] == pytest.approx(0.5)
+    assert by_score[0.0] == pytest.approx(0.0)