SmolDataEnvs-harbor-eval: remove null tool_efficiency from graders (#1)

- SmolDataEnvs-harbor-eval: remove null tool_efficiency from graders (8f940fef4d877436ef4aafdcb21804902a0cb9b0)


Co-authored-by: Govind Kamtamneni <thegovind@users.noreply.huggingface.co>

tasks/0000_886_886039_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_901_901543_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_901_901543_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_903_903504_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_939_939639_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_939_939639_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_949_949403_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_975_975906_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_986_986420_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_986_986420_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_988_988342_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_988_988657_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_991_991010_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_991_991010_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_991_991010_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_994_994384_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0000_994_994443_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_004_1004716_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_024_1024990_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_027_1027372_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_029_1029150_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_036_1036744_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_042_1042725_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_044_1044175_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_062_1062946_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_065_1065379_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_067_1067596_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_068_1068916_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_068_1068916_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_078_1078204_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_079_1079149_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_082_1082305_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_083_1083006_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_091_1091047_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_091_1091047_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_095_1095486_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_102_1102505_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_104_1104901_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_104_1104901_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_104_1104959_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_105_1105529_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_105_1105892_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_105_1105892_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_116_1116342_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_116_1116342_qa_5/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_123_1123075_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_124_1124736_qa_1/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_124_1124736_qa_2/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_133_1133224_qa_4/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0
 

tasks/0001_133_1133625_qa_3/tests/grader.py

@@ -106,28 +106,20 @@ def grade(gold: str, candidate: str, *, question: str = "", reward_mode: str = "
     return GradeResult(0.0, "miss")
 
 
-def _tool_calls():
-    try:
-        with open("/workdir/.n_tool_calls") as fh:
-            return int(fh.read().strip())
-    except (OSError, ValueError):
-        pass
-    raw = os.environ.get("N_TOOL_CALLS")
-    if raw is not None:
-        try:
-            return int(str(raw).strip())
-        except ValueError:
-            return None
-    return None
-
-
-def _tool_efficiency(n):
-    if n is None:
-        return None
-    budget = float(os.environ.get("TOOL_BUDGET", "15") or "15")
-    if budget <= 0:
-        return None
-    return max(0.0, min(1.0, 1.0 - n / budget))
+# TOOL EFFICIENCY IS NOT COMPUTED HERE, AND CANNOT BE.
+#
+# This grader used to emit a `tool_efficiency` key derived from a tool-call count it read from
+# `/workdir/.n_tool_calls` or `$N_TOOL_CALLS`. Nothing writes that file and no task.toml sets that
+# variable, so the count was always None, the key was always null, and Harbor's reward model
+# (`dict[str, float | int]`) rejected the WHOLE dict -- taking `correctness` down with it. The only
+# path that survived was the empty-submission branch below, so the suite graded failures correctly
+# and crashed on every successful answer. That silently removed 86 of 250 tasks from scoring.
+#
+# The count is not the sandbox's to know: it lives in the capture proxy, which sees every model call.
+# A reward belongs in the verifier only if the sandbox is what makes it computable. Correctness needs
+# the data, the gold answer and the tolerances -- it belongs here. Tool efficiency needs a trace --
+# it belongs to whoever holds the trace, and it is computed there as a shaping term, gated on
+# correctness so that "made no tool calls" can never outscore solving the task.
 
 
 def _tols():
@@ -156,12 +148,11 @@ def main_json() -> int:
     question = (os.environ.get("QUESTION") or "").strip()
     candidate = sys.stdin.read().strip()
     if not candidate:
-        print(json.dumps({"correctness": 0.0, "submission": 0.0, "tool_efficiency": 0.0}))
+        print(json.dumps({"correctness": 0.0, "submission": 0.0}))
         return 0
     at, rt = _tols()
     r = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE", "") or "", abs_tol=at, rel_tol=rt)
-    n = _tool_calls()
-    print(json.dumps({"correctness": float(r.reward), "submission": 1.0, "tool_efficiency": _tool_efficiency(n)}))
+    print(json.dumps({"correctness": float(r.reward), "submission": 1.0}))
     print(f"[grader] gold={gold!r} pred={candidate[:80]!r} correctness={r.reward} method={r.method}", file=sys.stderr)
     return 0