diff --git a/Backend.Tests/PolicyRunnerTests.cs b/Backend.Tests/PolicyRunnerTests.cs
new file mode 100644
index 0000000..9c044cd
--- /dev/null
+++ b/Backend.Tests/PolicyRunnerTests.cs
@@ -0,0 +1,55 @@
+using System.IO;
+using Backend;
+using Xunit;
+
+namespace Backend.Tests;
+
+public class PolicyRunnerTests
+{
+    private static string LocateOnnxModel()
+    {
+        var dir = AppContext.BaseDirectory;
+        while (dir is not null)
+        {
+            var candidate = Path.Combine(dir, "models", "ppo_lander.onnx");
+            if (File.Exists(candidate)) return candidate;
+            var parent = Directory.GetParent(dir);
+            if (parent is null) break;
+            dir = parent.FullName;
+        }
+        throw new FileNotFoundException(
+            "models/ppo_lander.onnx not found — run `python Training/export_onnx.py ...`");
+    }
+
+    [Fact]
+    public void LoadsWithoutError()
+    {
+        using var runner = new PolicyRunner(LocateOnnxModel());
+    }
+
+    [Fact]
+    public void SelectAction_ReturnsValidActionForZeroObs()
+    {
+        using var runner = new PolicyRunner(LocateOnnxModel());
+        var obs = new float[7];  // all zeros
+        int action = runner.SelectAction(obs);
+        Assert.InRange(action, 0, 3);
+    }
+
+    [Fact]
+    public void SelectAction_IsDeterministicForSameInput()
+    {
+        using var runner = new PolicyRunner(LocateOnnxModel());
+        var obs = new float[] { 0.1f, -0.2f, 0.05f, 0.0f, 0.0f, 1.0f, 0.0f };
+        int a1 = runner.SelectAction(obs);
+        int a2 = runner.SelectAction(obs);
+        Assert.Equal(a1, a2);
+    }
+
+    [Fact]
+    public void SelectAction_ThrowsIfObsLengthWrong()
+    {
+        using var runner = new PolicyRunner(LocateOnnxModel());
+        Assert.Throws<ArgumentException>(() => runner.SelectAction(new float[3]));
+    }
+}
diff --git a/Backend/PolicyRunner.cs b/Backend/PolicyRunner.cs
new file mode 100644
index 0000000..9ec10b0
--- /dev/null
+++ b/Backend/PolicyRunner.cs
@@ -0,0 +1,52 @@
+using Microsoft.ML.OnnxRuntime;
+using Microsoft.ML.OnnxRuntime.Tensors;
+
+namespace Backend;
+
+/// <summary>
+/// Loads a PPO policy ONNX model once and provides deterministic action selection.
+/// Thread-safe: <see cref="InferenceSession"/> is safe for concurrent Run() calls.
+/// </summary>
+public sealed class PolicyRunner : IDisposable
+{
+    public const int ObsDim = 7;
+    public const int NActions = 4;
+
+    private readonly InferenceSession _session;
+    private readonly string _inputName;
+
+    public PolicyRunner(string onnxPath)
+    {
+        if (!File.Exists(onnxPath))
+            throw new FileNotFoundException($"ONNX model not found at {onnxPath}");
+        _session = new InferenceSession(onnxPath);
+        _inputName = _session.InputMetadata.Keys.First();
+    }
+
+    /// <summary>Run one inference and return the argmax action index.</summary>
+    public int SelectAction(ReadOnlySpan<float> obs)
+    {
+        if (obs.Length != ObsDim)
+            throw new ArgumentException($"expected obs of length {ObsDim}, got {obs.Length}", nameof(obs));
+
+        var tensor = new DenseTensor<float>(new[] { 1, ObsDim });
+        for (int i = 0; i < ObsDim; i++) tensor[0, i] = obs[i];
+
+        using var results = _session.Run(new[]
+        {
+            NamedOnnxValue.CreateFromTensor(_inputName, tensor)
+        });
+
+        var logits = results.First().AsEnumerable<float>().ToArray();
+        // argmax
+        int best = 0;
+        float bestVal = logits[0];
+        for (int i = 1; i < logits.Length; i++)
+        {
+            if (logits[i] > bestVal) { best = i; bestVal = logits[i]; }
+        }
+        return best;
+    }
+
+    public void Dispose() => _session.Dispose();
+}