Add files using upload-large-folder tool

tasks/0000_582_582934_qa_4/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0000_582_582934_qa_4/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0000_582_582934_qa_4/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0000_582_582934_qa_4/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_105_1105892_qa_5/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_105_1105892_qa_5/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_105_1105892_qa_5/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_105_1105892_qa_5/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_261_1261978_qa_5/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_261_1261978_qa_5/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_261_1261978_qa_5/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_261_1261978_qa_5/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_439_1439907_qa_3/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_439_1439907_qa_3/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_439_1439907_qa_3/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_439_1439907_qa_3/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_487_1487950_qa_1/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_487_1487950_qa_1/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_487_1487950_qa_1/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_487_1487950_qa_1/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_533_1533644_qa_3/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_533_1533644_qa_3/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_533_1533644_qa_3/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_533_1533644_qa_3/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_683_1683832_qa_5/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_683_1683832_qa_5/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_683_1683832_qa_5/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_683_1683832_qa_5/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_727_1727043_qa_4/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_727_1727043_qa_4/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_727_1727043_qa_4/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_727_1727043_qa_4/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0001_735_1735855_qa_3/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0001_735_1735855_qa_3/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0001_735_1735855_qa_3/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0001_735_1735855_qa_3/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0002_092_2092240_qa_5/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0002_092_2092240_qa_5/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0002_092_2092240_qa_5/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0002_092_2092240_qa_5/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0021_135_21135636_qa_3/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0021_135_21135636_qa_3/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0021_135_21135636_qa_3/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0021_135_21135636_qa_3/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0021_656_21656365_qa_5/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir

tasks/0021_656_21656365_qa_5/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0021_656_21656365_qa_5/tests/grader.py

@@ -0,0 +1,257 @@
+"""Shared three-tier grader.
+
+Used by both runtimes:
+  * `env/server.py` — inline inside the `final_answer` ORS @tool method
+  * `tests/test.sh` — Harbor CLI path; this module is mounted at /tests/grader.py
+    and invoked with `python3 /tests/grader.py < /workdir/answer.txt`
+
+Tiers (cheapest → most expensive):
+  1. Exact (case-insensitive, whitespace-collapsed)
+  2. Numeric (parse a float, accept abs/rel diff ≤ 1e-3)
+  3. LLM-judge (gpt-4o-mini, A/B/C verdict, OpenAI simple-evals prompt)
+     — only invoked if OPENAI_API_KEY is set
+
+CLI form (used by tests/test.sh):
+  EXPECTED_ANSWER='5' QUESTION='How many...' python3 grader.py
+    → reads predicted answer from stdin
+    → writes a single float ('1.0' or '0.0') to stdout
+    → exits 0 always; harbor reads stdout from /logs/verifier/reward.txt
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import sys
+from dataclasses import dataclass
+
+
+_NUMERIC_RE = re.compile(r"-?\d+(?:[.,]\d+)?(?:[eE][-+]?\d+)?")
+
+# Locked in PLAN.md §"Locked decisions" — adapted from
+# https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py
+JUDGE_PROMPT = """You are grading short answers from a data-science agent.
+Decide whether the predicted answer is semantically equivalent to the gold answer.
+
+Rules:
+- Case, punctuation, whitespace, articles ("the"), and trailing units don't matter.
+- For numbers: predicted must match to the last significant figure of the gold
+  (gold "0.544341", predicted "0.544" -> CORRECT; "0.5" -> INCORRECT).
+- Extra surrounding prose is fine if the gold value is clearly stated
+  (gold "5", predicted "There are 5 distinct categories" -> CORRECT).
+- Common synonyms/abbreviations count (gold "Not Applicable", predicted "N/A" -> CORRECT).
+- Parenthetical annotations in the gold are equivalence hints, not required
+  (gold "Gandalf (Ainur)", predicted "Gandalf" -> CORRECT;
+   gold "No (correlation coefficient = 0.02)", predicted "No" -> CORRECT).
+- Percent + qualifier: if the gold is "X% in YEAR", a prediction of X or X%
+  is CORRECT as long as the core numeric value matches
+  (gold "21.334% in 2014", predicted "21.334" -> CORRECT;
+   gold "21.334% in 2014", predicted "21.334% in 2013" -> INCORRECT — year mismatch).
+- Multi-part gold like "0.987 (XGBoost and LGBM)" — if the question asks for one
+  thing (e.g. "which model?"), predicting either listed value is CORRECT.
+- If the predicted answer hedges without committing to the gold value -> NOT_ATTEMPTED.
+- If it commits to a different value -> INCORRECT.
+
+Question: {question}
+Gold answer: {gold}
+Predicted answer: {pred}
+
+Reply with exactly one token: A (CORRECT), B (INCORRECT), or C (NOT_ATTEMPTED)."""
+
+
+@dataclass
+class GradeResult:
+    reward: float
+    method: str  # "exact" | "numeric" | "llm" | "miss"
+
+
+def _normalize(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip().lower())
+
+
+def _to_float(s: str) -> float | None:
+    if not s:
+        return None
+    m = _NUMERIC_RE.search(s.replace(",", ""))
+    if not m:
+        return None
+    try:
+        return float(m.group(0))
+    except ValueError:
+        return None
+
+
+def _math_verify_match(gold: str, candidate: str) -> bool:
+    """Deterministic symbolic/numeric equivalence via HuggingFace Math-Verify.
+
+    Catches equivalences the naive float-parse misses: thousands separators,
+    fractions (3/4 == 0.75), percentages (50% == 0.5), and symbolic forms.
+    Fails CLOSED — any parse error / prose gold / timeout returns False (never a
+    false positive). No network, no LLM.
+    """
+    try:
+        from math_verify import parse, verify
+
+        return bool(verify(parse(gold), parse(candidate), timeout_seconds=5))
+    except Exception:  # noqa: BLE001
+        return False
+
+
+def _is_clean_number(s: str) -> bool:
+    """True iff the whole string IS a single number (allowing a leading/trailing
+    %, $, and thousands-commas) — e.g. '95293', '95,293', '0.5', '-3.16', '50%'.
+    A COMPOUND string like 'Y=3 with 95,293 instances' is NOT a clean number, so
+    the numeric tier must not fire on it (its `_to_float` would grab the leading
+    '3' and false-positive-match any answer starting with 3)."""
+    t = (s or "").strip().strip("%$").strip().replace(",", "")
+    return bool(_NUMERIC_RE.fullmatch(t))
+
+
+def grade(
+    gold: str,
+    candidate: str,
+    *,
+    question: str = "",
+    reward_mode: str = "",
+    judge: bool = True,  # noqa: ARG001 — kept for signature compat; LLM tier removed
+    judge_model: str | None = None,  # noqa: ARG001
+    rel_tol: float = 1e-3,
+    abs_tol: float = 1e-3,
+) -> GradeResult:
+    """Deterministic three-tier match — NO LLM judge.
+
+    exact -> numeric (abs_tol/rel_tol, gated to clean-number golds) -> math_verify
+    -> miss. `judge`/`judge_model` accepted but ignored (LLM tier removed).
+    """
+
+    if not gold or candidate is None:
+        return GradeResult(0.0, "miss")
+
+    # ── Tier 1: exact (case-insensitive, whitespace-collapsed) ─��───────
+    if _normalize(gold) == _normalize(candidate):
+        return GradeResult(1.0, "exact")
+
+    # ── Tier 2: numeric — ONLY when the gold is a single clean number ──
+    # (compound golds like "Y=3 with 95,293 instances" would otherwise match on
+    #  the leading number; those must go to exact/math_verify, which fail closed).
+    if reward_mode == "numeric" or _is_clean_number(gold):
+        g, c = _to_float(gold), _to_float(candidate)
+        if g is not None and c is not None:
+            if abs(g - c) <= abs_tol or abs(g - c) / max(abs(g), 1e-9) <= rel_tol:
+                return GradeResult(1.0, "numeric")
+
+    # ── Tier 3: math-verify (symbolic/numeric equivalence, deterministic) ──
+    if _math_verify_match(gold, candidate):
+        return GradeResult(1.0, "math_verify")
+
+    return GradeResult(0.0, "miss")
+
+
+def _tool_calls() -> int | None:
+    """Number of agent tool calls this rollout, for the efficiency reward.
+
+    Written by the harness into the sandbox (`/workdir/.n_tool_calls`); falls
+    back to the `N_TOOL_CALLS` env var. Returns None if neither is present
+    (efficiency reward becomes null — the harness didn't report a count).
+    """
+    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_calls: int | None) -> float | None:
+    """0..1 efficiency signal: fewer tool calls -> higher. None if unknown.
+
+    clip(1 - n_calls / TOOL_BUDGET, 0, 1). TOOL_BUDGET default 15 (solved-4B
+    median was ~9-10 calls). Raw signal — NOT gated on correctness here; the
+    RL reward combiner decides gating/weights.
+    """
+    if n_calls 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_calls / budget))
+
+
+def _tols() -> tuple[float, float]:
+    """Per-task (abs_tol, rel_tol) from ATOL/RTOL in [verifier.env]; 1e-3 default."""
+    def _f(name: str) -> float:
+        try:
+            return float(os.environ.get(name, "") or 1e-3)
+        except ValueError:
+            return 1e-3
+    return _f("ATOL"), _f("RTOL")
+
+
+def main_cli() -> int:
+    """CLI shim for `tests/test.sh` — env-driven, stdout-only.
+
+    Inputs (via env vars set by Harbor's [verifier.env]):
+      EXPECTED_ANSWER — the gold
+      QUESTION        — the original question
+      ATOL / RTOL     — per-task numeric tolerances (default 1e-3)
+
+    Predicted answer is read from stdin.
+    Single float written to stdout. Exit code is always 0.
+    """
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    question = (os.environ.get("QUESTION") or "").strip()
+    candidate = sys.stdin.read().strip()
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    print(f"{result.reward:.1f}")
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} reward={result.reward} method={result.method}",
+          file=sys.stderr)
+    return 0
+
+
+def main_json() -> int:
+    """Multi-reward shim for `tests/test.sh` — writes a `reward.json` dict.
+
+    Three named rewards (Harbor's VerifierResult.rewards carries them all):
+      correctness      — 1.0 iff graded answer matches gold, else 0.0
+      submission       — 1.0 iff a non-empty answer was submitted, else 0.0
+      tool_efficiency  — 0..1 (fewer tool calls higher); null if count unknown
+
+    No submission -> all three 0. Predicted answer on stdin; JSON to stdout.
+    """
+    import json
+
+    gold = (os.environ.get("EXPECTED_ANSWER") or "").strip()
+    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("[grader] no submission -> all rewards 0", file=sys.stderr)
+        return 0
+
+    at, rt = _tols()
+    result = grade(gold, candidate, question=question, reward_mode=os.environ.get("REWARD_MODE","") or "", abs_tol=at, rel_tol=rt)
+    n_calls = _tool_calls()
+    eff = _tool_efficiency(n_calls)
+    print(json.dumps({
+        "correctness": float(result.reward),
+        "submission": 1.0,
+        "tool_efficiency": eff,
+    }))
+    print(f"[grader] gold={gold!r} pred={candidate[:80]!r} "
+          f"correctness={result.reward} method={result.method} "
+          f"tool_calls={n_calls} tool_efficiency={eff}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    if "--json" in sys.argv[1:]:
+        raise SystemExit(main_json())
+    raise SystemExit(main_cli())

tasks/0021_656_21656365_qa_5/tests/test.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -u
+mkdir -p /logs/verifier
+
+answer_path="/workdir/answer.txt"
+if [ ! -s "$answer_path" ]; then
+  echo "0.0" > /logs/verifier/reward.txt
+  echo "[grader] no answer at $answer_path" >&2
+  exit 0
+fi
+
+pip install --quiet math-verify >/dev/null 2>&1 || true
+python3 /tests/grader.py < "$answer_path" > /logs/verifier/reward.txt

tasks/0023_757_23757075_qa_3/environment/pull_bucket.py

@@ -0,0 +1,41 @@
+"""Pull this task's bucket prefix into /home/user/input/.
+
+Invoked by Harbor's [environment.healthcheck] command (declared in task.toml)
+— runs after container start, before the agent. Idempotent: skips work if
+files are already present from a prior pull.
+"""
+
+import os
+import sys
+from pathlib import Path
+
+from huggingface_hub import download_bucket_files, list_bucket_tree
+
+
+def main() -> int:
+    bucket = os.environ.get("HF_BUCKET") or "AdithyaSK/jupyter-agent-kaggle-all"
+    prefix = os.environ["BUCKET_PREFIX"].rstrip("/") + "/"
+    dest = Path("/home/user/input")
+    dest.mkdir(parents=True, exist_ok=True)
+
+    existing = [p for p in dest.iterdir() if p.is_file()]
+    if existing:
+        print(f"[pull_bucket] {dest}/ already has {len(existing)} file(s); skipping", flush=True)
+        return 0
+
+    targets = [
+        (it.path, str(dest / Path(it.path).name))
+        for it in list_bucket_tree(bucket, prefix=prefix, recursive=True)
+        if getattr(it, "type", None) == "file"
+    ]
+    if not targets:
+        print(f"[pull_bucket] FATAL: no files at hf://buckets/{bucket}/{prefix}", flush=True)
+        return 2
+
+    download_bucket_files(bucket, files=targets)
+    print(f"[pull_bucket] staged {len(targets)} file(s) from {bucket}/{prefix}", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tasks/0033_308_33308132_qa_3/environment/Dockerfile

@@ -0,0 +1,2 @@
+FROM docker.io/savatar101/env-data-agent-train:base
+WORKDIR /workdir