diff --git a/configs/phase3.yaml b/configs/phase3.yaml
index 4b335c33..4089ad07 100644
--- a/configs/phase3.yaml
+++ b/configs/phase3.yaml
@@ -22,6 +22,7 @@ selection:
   top_n: 100
   min_novelty: 0.05
   max_safety_risk: 0.40
+  max_disagreement: 0.30
 
 notes:
   - "Phase 3 exploration config. Weights differ from pipeline.yaml to prioritize safety."
diff --git a/configs/pipeline.yaml b/configs/pipeline.yaml
index 82fd99a7..25e80bbb 100644
--- a/configs/pipeline.yaml
+++ b/configs/pipeline.yaml
@@ -15,6 +15,7 @@ selection:
   top_n: 100
   min_novelty: 0.20
   max_safety_risk: 0.70
+  max_disagreement: 0.40
 
 notes:
   - "These weights are transparent baseline heuristics, not validated biological predictors."
diff --git a/src/openamp_foundry/data/loaders.py b/src/openamp_foundry/data/loaders.py
index 2bb6f836..b8ec4273 100644
--- a/src/openamp_foundry/data/loaders.py
+++ b/src/openamp_foundry/data/loaders.py
@@ -23,6 +23,13 @@ def load_candidates_csv(path: str | Path) -> list[PeptideCandidate]:
     rows: list[PeptideCandidate] = []
     with Path(path).open("r", encoding="utf-8") as f:
         reader = csv.DictReader(f)
+        fieldnames = reader.fieldnames or []
+        if "sequence" not in fieldnames:
+            raise ValueError(
+                f"CSV header is missing required 'sequence' column. "
+                f"Found columns: {list(fieldnames)!r}. "
+                "Check that the CSV header spells it exactly as 'sequence'."
+            )
         for i, row in enumerate(reader, start=1):
             cid = row.get("id") or f"candidate-{i:06d}"
             seq = normalize_sequence(row["sequence"])
diff --git a/src/openamp_foundry/pipeline.py b/src/openamp_foundry/pipeline.py
index e231d22f..f85b3739 100644
--- a/src/openamp_foundry/pipeline.py
+++ b/src/openamp_foundry/pipeline.py
@@ -123,6 +123,7 @@ def run_ranking_pipeline(
     selection_cfg = config.get("selection", {})
     min_novelty = float(selection_cfg.get("min_novelty", 0.0))
     max_safety_risk = float(selection_cfg.get("max_safety_risk", 1.0))
+    max_disagreement = float(selection_cfg.get("max_disagreement", 1.0))
 
     ranked = rank_candidates(scored)
     top_n = int(selection_cfg.get("top_n", len(ranked)))
@@ -132,6 +133,7 @@ def run_ranking_pipeline(
         if getattr(item, "valid", True)
         and item.scores["novelty"] >= min_novelty
         and item.scores["safety"] >= (1.0 - max_safety_risk)
+        and item.scores["disagreement"] <= max_disagreement
     ]
     selected = greedy_diverse_select(eligible, top_n=top_n)
 
diff --git a/tests/test_data_loaders.py b/tests/test_data_loaders.py
index f9dbc141..dd808282 100644
--- a/tests/test_data_loaders.py
+++ b/tests/test_data_loaders.py
@@ -5,6 +5,8 @@
 import tempfile
 from pathlib import Path
 
+import pytest
+
 from openamp_foundry.data.loaders import (
     CANONICAL_AA,
     is_valid_sequence,
@@ -141,3 +143,35 @@ def test_uses_demo_file_successfully(self):
         assert len(candidates) > 0
         for c in candidates:
             assert len(c.sequence) > 0
+
+    def test_missing_sequence_column_raises_value_error(self):
+        # A CSV with no 'sequence' column should give a helpful error, not a bare KeyError.
+        # This guards against users passing in a wrongly-formatted file and getting
+        # a cryptic crash deep in the pipeline.
+        with tempfile.TemporaryDirectory() as d:
+            path = Path(d) / "bad.csv"
+            with path.open("w", newline="", encoding="utf-8") as f:
+                writer = csv.DictWriter(f, fieldnames=["id", "seq", "source"])
+                writer.writeheader()
+                writer.writerow({"id": "C1", "seq": "KWKLF", "source": "test"})
+            with pytest.raises(ValueError, match="sequence"):
+                load_candidates_csv(path)
+
+    def test_missing_sequence_column_fires_on_header_only_csv(self):
+        # Header validation happens before iterating rows — zero-row files are caught too.
+        with tempfile.TemporaryDirectory() as d:
+            path = Path(d) / "header_only.csv"
+            path.write_text("id,seq,source\n")  # header only, no data rows
+            with pytest.raises(ValueError, match="sequence"):
+                load_candidates_csv(path)
+
+    def test_missing_sequence_column_error_names_found_columns(self):
+        # The error message should list what columns were found to help the user.
+        with tempfile.TemporaryDirectory() as d:
+            path = Path(d) / "bad.csv"
+            with path.open("w", newline="", encoding="utf-8") as f:
+                writer = csv.DictWriter(f, fieldnames=["peptide"])
+                writer.writeheader()
+                writer.writerow({"peptide": "KWKLF"})
+            with pytest.raises(ValueError, match="peptide"):
+                load_candidates_csv(path)
diff --git a/tests/test_pipeline_filters.py b/tests/test_pipeline_filters.py
index db037814..7ed53284 100644
--- a/tests/test_pipeline_filters.py
+++ b/tests/test_pipeline_filters.py
@@ -164,6 +164,68 @@ def test_eligibility_novelty_at_exact_minimum_passes(tmp_path):
     assert rows[0]["selected"] is True
 
 
+def test_disagreement_gate_blocks_high_uncertainty_candidate(tmp_path):
+    """Candidates where activity and Boman index strongly disagree should be blocked.
+
+    max_disagreement caps the |activity - boman_activity| gap. The test proves the
+    gate is the cause of blocking by showing the SAME candidate is SELECTED when the
+    gate is relaxed to 1.0 (no restriction).
+    """
+    _COMMON_CONFIG = (
+        "pipeline_version: '0.1.0'\n"
+        "filters:\n  min_length: 8\n  max_length: 35\n"
+        "  allowed_amino_acids: 'ACDEFGHIKLMNPQRSTVWY'\n"
+        "weights:\n  activity: 0.40\n  safety: 0.25\n  synthesis: 0.15\n  novelty: 0.20\n"
+        "selection:\n  top_n: 100\n  min_novelty: 0.0\n  max_safety_risk: 1.0\n"
+    )
+    strict_config = tmp_path / "strict_disagree.yaml"
+    strict_config.write_text(_COMMON_CONFIG + "  max_disagreement: 0.0\n")
+    permissive_config = tmp_path / "permissive_disagree.yaml"
+    permissive_config.write_text(_COMMON_CONFIG + "  max_disagreement: 1.0\n")
+
+    seq = "KWKLFKKIGAVLKVL"
+    csv_path = tmp_path / "candidates.csv"
+    refs_path = tmp_path / "refs.csv"
+    csv_path.write_text(f"id,sequence,source\nTEST-001,{seq},test\n")
+    refs_path.write_text("id,sequence,source\n")  # empty refs → novelty=1.0
+
+    # Strict gate: disagreement > 0 for any real sequence → blocked
+    out_strict = tmp_path / "strict.jsonl"
+    run_ranking_pipeline(
+        candidate_path=csv_path, reference_path=refs_path,
+        out_path=out_strict, config_path=strict_config,
+    )
+    rows_strict = [json.loads(ln) for ln in out_strict.read_text().splitlines() if ln.strip()]
+    assert len(rows_strict) == 1
+    assert rows_strict[0]["scores"]["disagreement"] > 0.0
+    assert rows_strict[0]["selected"] is False  # disagreement gate blocks it
+
+    # Positive control: same candidate passes when gate is off (max_disagreement=1.0)
+    out_permissive = tmp_path / "permissive.jsonl"
+    run_ranking_pipeline(
+        candidate_path=csv_path, reference_path=refs_path,
+        out_path=out_permissive, config_path=permissive_config,
+    )
+    rows_permissive = [json.loads(ln) for ln in out_permissive.read_text().splitlines() if ln.strip()]
+    assert len(rows_permissive) == 1
+    assert rows_permissive[0]["selected"] is True  # relaxed gate allows it through
+
+
+def test_disagreement_gate_config_values_documented(tmp_path):
+    """pipeline.yaml max_disagreement=0.40, phase3.yaml max_disagreement=0.30 (stricter)."""
+    repo_root = Path(__file__).parents[1]
+    pipeline_cfg = load_config(repo_root / "configs" / "pipeline.yaml")
+    phase3_cfg = load_config(repo_root / "configs" / "phase3.yaml")
+
+    pipeline_max = float(pipeline_cfg["selection"]["max_disagreement"])
+    phase3_max = float(phase3_cfg["selection"]["max_disagreement"])
+
+    # phase3 is stricter (lower tolerance for disagreement)
+    assert phase3_max < pipeline_max
+    assert abs(pipeline_max - 0.40) < 0.01
+    assert abs(phase3_max - 0.30) < 0.01
+
+
 def test_phase3_config_has_stricter_safety_filter_than_pipeline():
     """phase3.yaml max_safety_risk=0.40 (safety>=0.60) vs pipeline.yaml 0.70 (safety>=0.30).