add a metrics.json

Author: Mengtao Zhang

README.md

@@ -121,6 +121,9 @@ representative Harbor run on 2026-07-29 took 20 minutes 53 seconds end to end:
   `verifier/reward.json`. Submitted artifacts are collected from the
   `mcp-server` service, not the agent container.
 
-The verifier writes numeric `reward.json` fields for `primary_score`, `reward`,
-`sharpe`, `cagr`, `max_drawdown`, and `pearson_ic`. `primary_score`, `reward`,
-and `sharpe` are the same hidden after-cost Sharpe value.
+The verifier writes Harbor-compatible numeric reward fields to
+`reward.json`: `primary_score`, `reward`, `sharpe`, `cagr`, `max_drawdown`, and
+`pearson_ic`. `primary_score`, `reward`, and `sharpe` are the same hidden
+after-cost Sharpe value. `metrics.json` is a separate diagnostic artifact
+containing the full official scoring payload, including the complete backtest
+metrics, causal-audit result, and refit diagnostics.

environment/mcp-server/feature_engineering/evaluation/verifier.py

@@ -28,6 +28,7 @@
 
 def _write_outputs(
     rewards: dict[str, float],
+    metrics: dict[str, object],
     *,
     reward_path: str | Path,
     metrics_path: str | Path,
@@ -41,7 +42,7 @@ def _write_outputs(
         encoding="utf-8",
     )
     metrics_file.write_text(
-        json.dumps(rewards, sort_keys=True, allow_nan=False, indent=2) + "\n",
+        json.dumps(metrics, sort_keys=True, allow_nan=False, indent=2) + "\n",
         encoding="utf-8",
     )
 
@@ -77,6 +78,7 @@ async def verify_submission(
         }
         _write_outputs(
             rewards,
+            {"ok": False, "metrics": rewards},
             reward_path=reward_path,
             metrics_path=metrics_path,
         )
@@ -152,6 +154,7 @@ async def verify_submission(
 
     _write_outputs(
         rewards,
+        scored,
         reward_path=reward_path,
         metrics_path=metrics_path,
     )

environment/mcp-server/tests/test_end_to_end_behavior.py

@@ -157,9 +157,15 @@ async def test_real_training_submission_and_verification_workflow(self) -> None:
             self.assertEqual(
                 json.loads(reward_path.read_text(encoding="utf-8")), rewards
             )
+            metrics = json.loads(metrics_path.read_text(encoding="utf-8"))
+            self.assertTrue(metrics["ok"])
+            self.assertEqual(metrics["primary_score"], rewards["primary_score"])
             self.assertEqual(
-                json.loads(metrics_path.read_text(encoding="utf-8")), rewards
+                metrics["metrics"]["annualized_sharpe"],
+                rewards["sharpe"],
             )
+            self.assertIn("fit_diagnostics", metrics)
+            self.assertNotEqual(metrics, rewards)
 
 
 if __name__ == "__main__":

environment/mcp-server/tests/test_verifier_behavior.py

@@ -188,7 +188,7 @@ async def test_missing_submission_writes_negative_reward_outputs(self) -> None:
         )
         self.assertEqual(
             json.loads((self.root / "metrics.json").read_text(encoding="utf-8")),
-            expected,
+            {"ok": False, "metrics": expected},
         )
         score.assert_not_awaited()
 
@@ -336,6 +336,13 @@ async def score_hidden(**kwargs: object) -> BacktestResult:
         self.assertEqual(calls["hidden_artifact"], replacement_bytes)
         self.assertIs(calls["hidden_public_data"], calls["refit_public_data"])
         self.assertEqual(rewards["primary_score"], 1.25)
+        metrics = json.loads(
+            (self.root / "metrics.json").read_text(encoding="utf-8")
+        )
+        self.assertTrue(metrics["ok"])
+        self.assertEqual(metrics["primary_score"], 1.25)
+        self.assertEqual(metrics["metrics"]["annualized_sharpe"], 1.25)
+        self.assertEqual(metrics["causal_audit"], {"visibility": "hidden_fixed"})
 
     async def test_hidden_prediction_attachment_contains_no_outcome_columns(
         self,