diff --git a/src/pipeline/gates.py b/src/pipeline/gates.py
index a7bd8062..161f3ba8 100644
--- a/src/pipeline/gates.py
+++ b/src/pipeline/gates.py
@@ -257,12 +257,56 @@ def check(self, recipe: RecipeSpec, ctx: GateContext) -> GateResult:
                           f"{float(size_mb):.1f} MB budget")
 
 
+class GraphIoGate(BaseGate):
+    """WO83: no artifact ships fp16 at the graph boundary. Weights may
+    stay fp16 (WO79 owner decision); every graph input and output must
+    be fp32 (or non-float). Enforced here so the exporter default can
+    never silently regress."""
+
+    name = "graph_io"
+
+    def check(self, recipe: RecipeSpec, ctx: GateContext) -> GateResult:
+        art = (ctx.extra or {}).get("artifact_path")
+        if not art:
+            return GateResult(self.name, True,
+                              "artifact not provided — IO guard not "
+                              "evaluated (enforced at export/publish)")
+        path = Path(art)
+        if not path.is_file():
+            return GateResult(self.name, False,
+                              f"artifact missing: {path}")
+        import zipfile
+
+        import onnx
+        from onnx import TensorProto
+
+        offenders = []
+        with zipfile.ZipFile(path) as z:
+            for member in z.namelist():
+                if not member.endswith(".onnx"):
+                    continue
+                m = onnx.load_from_string(z.read(member))
+                for kind, infos in (("input", m.graph.input),
+                                    ("output", m.graph.output)):
+                    for t in infos:
+                        tt = t.type.tensor_type
+                        if tt.elem_type == TensorProto.FLOAT16:
+                            offenders.append(f"{member} {kind} {t.name}")
+        if offenders:
+            return GateResult(
+                self.name, False,
+                "fp16 at the graph boundary (WO79): "
+                + ", ".join(offenders))
+        return GateResult(self.name, True, "graph boundary fp32-clean")
+
+
 register_gate(QualityGate())
 register_gate(ParityGate())
 register_gate(LicenseGate())
 register_gate(ConformanceGate())
 register_gate(RegressionGate())
 register_gate(BudgetGate())
+register_gate(GraphIoGate())
 
 
 def evaluate(
diff --git a/tests/test_pipeline_graph_io_gate.py b/tests/test_pipeline_graph_io_gate.py
new file mode 100644
index 00000000..e7ee5dca
--- /dev/null
+++ b/tests/test_pipeline_graph_io_gate.py
@@ -0,0 +1,85 @@
+"""WO83 — the graph-IO dtype gate: no artifact ships fp16 at the graph
+boundary (weights may stay fp16; ruby has no Float16 and node ORT
+corrupts the boundary — WO79).
+"""
+
+from __future__ import annotations
+
+import zipfile
+from pathlib import Path
+
+import onnx
+from onnx import TensorProto, helper
+
+from pipeline.gates import GateContext, GraphIoGate, registered_gates
+
+
+def _save_graph(path: Path, in_type: int, out_type: int) -> None:
+    g = helper.make_graph(
+        [helper.make_node("Cast", ["x"], ["c1"], to=TensorProto.FLOAT),
+         helper.make_node("Cast", ["c1"], ["y"], to=out_type)],
+        "g",
+        [helper.make_tensor_value_info("x", in_type, [2, 2])],
+        [helper.make_tensor_value_info("y", out_type, [2, 2])],
+    )
+    m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 14)])
+    m.ir_version = 10
+    onnx.save(m, str(path))
+
+
+def _artifact(tmp_path: Path, in_type: int, out_type: int) -> Path:
+    graph = tmp_path / "encoder.onnx"
+    _save_graph(graph, in_type, out_type)
+    art = tmp_path / "artifact.zip"
+    with zipfile.ZipFile(art, "w") as z:
+        z.writestr("metadata.yaml", "format: imf-v1\n")
+        z.write(graph, "encoder.onnx")
+    return art
+
+
+def test_graph_io_gate_is_registered() -> None:
+    assert "graph_io" in registered_gates()
+
+
+def test_fp16_boundary_fails(tmp_path: Path) -> None:
+    art = _artifact(tmp_path, TensorProto.FLOAT16, TensorProto.FLOAT16)
+    result = GraphIoGate().check(
+        recipe=None, ctx=GateContext(extra={"artifact_path": str(art)}))
+    assert not result.passed
+    assert "fp16" in result.detail
+
+
+def test_fp32_boundary_passes(tmp_path: Path) -> None:
+    art = _artifact(tmp_path, TensorProto.FLOAT, TensorProto.FLOAT)
+    result = GraphIoGate().check(
+        recipe=None, ctx=GateContext(extra={"artifact_path": str(art)}))
+    assert result.passed
+
+
+def test_fp16_weights_with_fp32_boundary_pass(tmp_path: Path) -> None:
+    """The WO79 shape: fp16 graph math inside, fp32 IO — legal."""
+    art = tmp_path / "artifact.zip"
+    graph = tmp_path / "encoder.onnx"
+    g = helper.make_graph(
+        [helper.make_node("Cast", ["x"], ["h"], to=TensorProto.FLOAT16),
+         helper.make_node("Cast", ["h"], ["y_raw"], to=TensorProto.FLOAT16),
+         helper.make_node("Cast", ["y_raw"], ["y"], to=TensorProto.FLOAT)],
+        "g",
+        [helper.make_tensor_value_info("x", TensorProto.FLOAT, [2, 2])],
+        [helper.make_tensor_value_info("y", TensorProto.FLOAT, [2, 2])],
+    )
+    m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 14)])
+    m.ir_version = 10
+    onnx.save(m, str(graph))
+    with zipfile.ZipFile(art, "w") as z:
+        z.writestr("metadata.yaml", "format: imf-v1\n")
+        z.write(graph, "encoder.onnx")
+    result = GraphIoGate().check(
+        recipe=None, ctx=GateContext(extra={"artifact_path": str(art)}))
+    assert result.passed
+
+
+def test_missing_artifact_notes_without_failing() -> None:
+    result = GraphIoGate().check(recipe=None, ctx=GateContext())
+    assert result.passed
+    assert "not evaluated" in result.detail
diff --git a/tests/test_pipeline_plan_gate.py b/tests/test_pipeline_plan_gate.py
index 13163f5a..2703a8c7 100644
--- a/tests/test_pipeline_plan_gate.py
+++ b/tests/test_pipeline_plan_gate.py
@@ -158,6 +158,7 @@ def test_gate_results_files_end_to_end(tmp_path: Path) -> None:
         "conformance": False,
         "regression": True,
         "budgets": True,
+        "graph_io": True,
     }