SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.08093 · 2026

CausalDS: Benchmarking Causal Reasoning in Data-Science Agents

Yuekai Sun, Andrej Leban

arXiv · PDF

Code that ran

17 of 33 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (2 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 2 have no contract verdict; 14 failures of our instrument; 16 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2607.08093.json · badge: /api/ran/2607.08093.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
andleb/causalds community
Papers with Code archive
17 ran · 16 no recorded run
FunctionStatusWhere it lives
ExamComposition Ran
ran on a synthesized input · constructed an object; no result was computed
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("6fdf6553ceeed2b5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ExamComposition: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from dataclasses import dataclass
from typing import Dict, List, Optional

ARGS = {
    'name': 'weighted_policy_v1',
    'scene_sampling': 'uniform',
    'task_family_weights': {
        'family_a': 0.5,
        'family_b': 0.3,
        'family_c': 0.2,
    },
    'task_family_weights_by_structure_label': {
        'structure_1': {'family_a': 0.6, 'family_b': 0.4},
        'structure_2': {'family_a': 0.2, 'family_c': 0.8},
    },
    'output_variant_weights_by_family': {
        'family_a': {'variant_1': 0.7, 'variant_2': 0.3},
        'family_b': {'variant_1': 0.5, 'variant_2': 0.5},
    },
    'observation_variant_weights': {
        'obs_easy': 0.4,
        'obs_medium': 0.35,
        'obs_hard': 0.25,
    },
    'observation_variant_weights_by_family': {
        'family_a': {'obs_easy': 0.6, 'obs_hard': 0.4},
    },
    'observation_variant_order': [
        'obs_easy',
        'obs_medium',
        'obs_hard',
    ],
    'difficulty': 0.5,
    'rung_tilt_strength': 1.0,
    'output_variant_tilt_strength': 1.0,
    'obs_tilt_strength': 1.0,
    'max_scene_resamples': 4,
    'allow_silent_renorm': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ExamComposition — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ExamComposition output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
TaskType Ran
ran on a synthesized input · constructed an object; no result was computed
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("e0d77746002a838d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TaskType: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from enum import Enum

class TaskType(str, Enum):
    """Question family (semantic intent)."""
    PREDICTION = 'prediction'
    ASSOCIATION = 'association'
    COLLIDER_PHENOMENON = 'collider_phenomenon'
    CAUSAL_SKETCH = 'causal_sketch'
    IDENTIFICATION = 'identification'
    EFFECT_ESTIMATE = 'effect_estimate'
    BIAS_DIAGNOSTIC = 'bias_diagnostic'
    COUNTERFACTUAL_IDENTIFICATION = 'counterfactual_identification'
    COUNTERFACTUAL_EFFECT = 'counterfactual_effect'
    MEDIATION_EFFECT = 'mediation_effect'

    def __str__(self) -> str:
        return self.value

torch.manual_seed(0)

ARGS = {
    'value': 'prediction'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TaskType — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TaskType output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_apply_obs_tilt Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("471a02e4534852ff")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import math
from typing import Any, Dict, List, Optional

# Construct valid arguments for _apply_obs_tilt

# weights: Dict[Any, float] - a dictionary mapping observation keys to their weights
weights = {
    "obs_easy": 1.0,
    "obs_medium": 2.0,
    "obs_hard": 3.0,
    "obs_other": 0.5
}

# gamma: float - the tilt coefficient
gamma = 0.5

# difficulty: float - the difficulty parameter (typically between 0 and 1)
difficulty = 0.7

# obs_order: Optional[List[str]] - ordered list of observation names (easiest first)
obs_order = ["obs_easy", "obs_medium", "obs_hard"]

ARGS = {
    "weights": weights,
    "gamma": gamma,
    "difficulty": difficulty,
    "obs_order": obs_order
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_apply_output_variant_tilt Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("e5dd3c76df618a46")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import math
from typing import Dict, List, Optional
from enum import Enum

class OutputVariant(str, Enum):
    """Output contract for a question family."""
    UNKNOWN = 'unknown'
    POINT_PREDICTOR = 'point_predictor'
    PREDICTION_INTERVAL = 'prediction_interval'
    SIGN_ONLY = 'sign_only'
    EFFECT_SIZE_POINT = 'effect_size_point'
    SIGN_BEFORE_AFTER = 'sign_before_after'
    DELTA_POINT = 'delta_point'
    DELTA_SIGN_ONLY = 'delta_sign_only'
    ARGMAX_CHANGE = 'argmax_change'
    INDUCED_ASSOCIATION_BOOLEAN = 'induced_association_boolean'
    INDUCED_ASSOCIATION_SIGN_ONLY = 'induced_association_sign_only'
    INDUCED_ASSOCIATION_STRENGTH_POINT = 'induced_association_strength_point'
    EDGES_ONLY = 'edges_only'
    SKELETON_EDGES = 'skeleton_edges'
    ONE_VALID_ADJUSTMENT_SET = 'one_valid_adjustment_set'
    IDENTIFIABLE_BOOLEAN = 'identifiable_boolean'
    METHOD_LABEL = 'method_label'
    MINIMAL_ADJUSTMENT_SET_SIZE = 'minimal_adjustment_set_size'
    N_VALID_ADJUSTMENT_SETS = 'n_valid_adjustment_sets'
    ALL_MINIMAL_ADJUSTMENT_SETS = 'all_minimal_adjustment_sets'
    ATE_POINT = 'ate_point'
    ATE_UQ_95 = 'ate_uq_95'
    ATE_SIGN_ONLY = 'ate_sign_only'
    ATE_VS_ASSOC_SIGN_MATCH = 'ate_vs_assoc_sign_match'
    COLLIDER_BIAS_BOOLEAN = 'collider_bias_boolean'
    FORBIDDEN_CONTROLS_LIST = 'forbidden_controls_list'
    EFFECT_POINT = 'effect_point'
    EFFECT_UQ_95 = 'effect_uq_95'
    DIRECT_VS_INDIRECT_DOMINANCE = 'direct_vs_indirect_dominance'

    def __str__(self) -> str:
        return self.value

weights = {
    OutputVariant.POINT_PREDICTOR: 0.5,
    OutputVariant.SIGN_ONLY: 0.3,
    OutputVariant.PREDICTION_INTERVAL: 0.2,
}

delta = 0.5

difficulty = 0.7

variant_order = [
    OutputVariant.POINT_PREDICTOR,
    OutputVariant.SIGN_ONLY,
    OutputVariant.PREDICTION_INTERVAL,
]

ARGS = {
    "weights": weights,
    "delta": delta,
    "difficulty": difficulty,
    "variant_order": variant_order,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_portable_repo_path Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("7eb5370f70e56a33")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from pathlib import Path
from typing import Any, Dict, List, Optional

# The function _portable_repo_path takes a single parameter 'path' of type Any
# It converts the path to a Path object, then tries to make it relative to the repo root
# We need to provide a valid path argument

# Create a test path - we'll use a relative path that could exist in a repo
test_path = Path("src/module/file.py")

ARGS = {
    "path": test_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_read_scene_structure_label Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("25b539302a57ef8b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
from pathlib import Path

# Create a temporary directory structure to simulate the benchmark_dir
temp_dir = tempfile.mkdtemp()
benchmark_dir = Path(temp_dir)

# Create the scenes_private subdirectory structure
scene_id = "test_scene_001"
private_scene_dir = benchmark_dir / 'scenes_private' / scene_id
private_scene_dir.mkdir(parents=True, exist_ok=True)

# Create a ground_truth.json file with test data
ground_truth_data = {
    "metadata": {
        "structural_label": "test_structure",
        "motif": "test_motif"
    },
    "graph": {
        "augmentation": {
            "applied_grafts": 2
        },
        "auxiliary_graph_grafts": [],
        "main_graph": {
            "motif": "main_motif"
        }
    }
}

gt_path = private_scene_dir / 'ground_truth.json'
with open(gt_path, 'w') as f:
    json.dump(ground_truth_data, f)

ARGS = {
    'benchmark_dir': benchmark_dir,
    'scene_id': scene_id
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_resolve_family_weights Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("cd7aca1869f92036")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import logging
from typing import Any, Dict, List, Optional
from enum import Enum

class TaskType(str, Enum):
    """Question family (semantic intent)."""
    PREDICTION = 'prediction'
    ASSOCIATION = 'association'
    COLLIDER_PHENOMENON = 'collider_phenomenon'
    CAUSAL_SKETCH = 'causal_sketch'
    IDENTIFICATION = 'identification'
    EFFECT_ESTIMATE = 'effect_estimate'
    BIAS_DIAGNOSTIC = 'bias_diagnostic'
    COUNTERFACTUAL_IDENTIFICATION = 'counterfactual_identification'
    COUNTERFACTUAL_EFFECT = 'counterfactual_effect'
    MEDIATION_EFFECT = 'mediation_effect'

    def __str__(self) -> str:
        return self.value

logger = logging.getLogger(__name__)

def parse_task_type(value: Any) -> TaskType:
    """Parse a task type into a TaskType enum."""
    if isinstance(value, TaskType):
        return value
    if isinstance(value, str):
        v = value.strip().lower()
        return TaskType(v)
    raise ValueError(f'Unrecognized task_type: {value!r}')

def _resolve_family_weights(comp: 'ExamComposition', scene_structure_label: Optional[str]) -> Optional[Dict[TaskType, float]]:
    """Pick P(Q | structural_label) weights for a scene.

    Conditional table wins if it has a row for the scene's structural label;
    else the marginal; else None (→ uniform over eligible families).
    """
    raw_weights: Optional[Dict[str, float]] = None
    if comp.task_family_weights_by_structure_label and scene_structure_label is not None:
        raw_weights = comp.task_family_weights_by_structure_label.get(scene_structure_label)
    if not raw_weights and comp.task_family_weights:
        raw_weights = comp.task_family_weights
    if not raw_weights:
        return None
    enum_weights: Dict[TaskType, float] = {}
    for k, v in raw_weights.items():
        try:
            enum_weights[parse_task_type(k)] = float(v)
        except ValueError:
            logger.warning('Unknown task family %r in composition; ignoring', k)
    return enum_weights or None


class ExamComposition:
    """Stand-in class for ExamComposition with required attributes."""
    def __init__(
        self,
        task_family_weights: Optional[Dict[str, float]] = None,
        task_family_weights_by_structure_label: Optional[Dict[str, Dict[str, float]]] = None,
    ):
        self.task_family_weights = task_family_weights
        self.task_family_weights_by_structure_label = task_family_weights_by_structure_label


# Test case 1: Both marginal and conditional weights, conditional takes precedence
comp1 = ExamComposition(
    task_family_weights={
        'prediction': 0.3,
        'association': 0.4,
        'effect_estimate': 0.3,
    },
    task_family_weights_by_structure_label={
        'linear': {
            'prediction': 0.5,
            'causal_sketch': 0.5,
        },
        'nonlinear': {
            'mediation_effect': 0.6,
            'counterfactual_effect': 0.4,
        },
    },
)

ARGS = {
    'comp': comp1,
    'scene_structure_label': 'linear',
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_resolve_obs_weights Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("cedc4f415bde2b77")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from typing import Any, Dict, List, Optional
from enum import Enum

class TaskType(str, Enum):
    """Question family (semantic intent)."""
    PREDICTION = 'prediction'
    ASSOCIATION = 'association'
    COLLIDER_PHENOMENON = 'collider_phenomenon'
    CAUSAL_SKETCH = 'causal_sketch'
    IDENTIFICATION = 'identification'
    EFFECT_ESTIMATE = 'effect_estimate'
    BIAS_DIAGNOSTIC = 'bias_diagnostic'
    COUNTERFACTUAL_IDENTIFICATION = 'counterfactual_identification'
    COUNTERFACTUAL_EFFECT = 'counterfactual_effect'
    MEDIATION_EFFECT = 'mediation_effect'

    def __str__(self) -> str:
        return self.value


class ExamComposition:
    """Stand-in class for ExamComposition."""
    def __init__(self, observation_variant_weights_by_family=None, observation_variant_weights=None):
        self.observation_variant_weights_by_family = observation_variant_weights_by_family
        self.observation_variant_weights = observation_variant_weights


# Create test instances
comp_with_family_weights = ExamComposition(
    observation_variant_weights_by_family={
        'prediction': {'variant_a': 0.6, 'variant_b': 0.4},
        'association': {'variant_x': 0.3, 'variant_y': 0.7},
    },
    observation_variant_weights={'global_var1': 0.5, 'global_var2': 0.5}
)

comp_with_marginal_weights = ExamComposition(
    observation_variant_weights_by_family=None,
    observation_variant_weights={'obs_1': 0.25, 'obs_2': 0.75}
)

comp_with_no_weights = ExamComposition(
    observation_variant_weights_by_family=None,
    observation_variant_weights=None
)

chosen_family_prediction = TaskType.PREDICTION
chosen_family_effect = TaskType.EFFECT_ESTIMATE

ARGS = {
    'comp': comp_with_family_weights,
    'chosen_family': chosen_family_prediction
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_scene_variant_dir Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("bfc09c3b80e6fe08")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned PosixPath. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from pathlib import Path

ARGS = {
    "scene_dir": Path("/tmp/test_scene"),
    "observation_variant": None,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_weighted_choice_from_available Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("3d6d99900a05a2fd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import random
from typing import Any, Dict, List, Optional

# Create a random.Random instance for the rng parameter
rng = random.Random(42)

# Create all_weights as a dictionary with various keys and float values
all_weights = {
    'option_a': 0.5,
    'option_b': 0.3,
    'option_c': 0.2,
    'option_d': 0.0,
}

# Create available as a list of keys that may or may not be in all_weights
available = ['option_a', 'option_c', 'option_e']

# allow_silent_renorm as a boolean
allow_silent_renorm = True

# no_choice as a fallback value
no_choice = None

ARGS = {
    'rng': rng,
    'all_weights': all_weights,
    'available': available,
    'allow_silent_renorm': allow_silent_renorm,
    'no_choice': no_choice,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
list_scene_variants Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("eae3e3a4f970a69c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path

# Create a temporary directory structure to test the function
temp_dir = tempfile.mkdtemp()
scene_dir = Path(temp_dir) / "test_scene"
scene_dir.mkdir(exist_ok=True)

# Create variants directory with some test variants
variants_dir = scene_dir / "variants"
variants_dir.mkdir(exist_ok=True)

# Create variant1 with tasks.json
variant1_dir = variants_dir / "variant1"
variant1_dir.mkdir(exist_ok=True)
(variant1_dir / "tasks.json").touch()

# Create variant2 with data.parquet
variant2_dir = variants_dir / "variant2"
variant2_dir.mkdir(exist_ok=True)
(variant2_dir / "data.parquet").touch()

# Create variant3 without required files (should be skipped)
variant3_dir = variants_dir / "variant3"
variant3_dir.mkdir(exist_ok=True)

# Create a non-directory file in variants (should be skipped)
(variants_dir / "somefile.txt").touch()

ARGS = {
    "scene_dir": scene_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
list_scenes Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("fbf133530644d05a")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path

# Create a temporary directory structure for testing
temp_dir = tempfile.mkdtemp()
base_dir = Path(temp_dir)

# Create scenes directory
scenes_dir = base_dir / 'scenes'
scenes_dir.mkdir(exist_ok=True)

# Create some test scene directories with story.md files
for scene_id in ['scene_001', 'scene_002', 'scene_003']:
    scene_path = scenes_dir / scene_id
    scene_path.mkdir(exist_ok=True)
    story_file = scene_path / 'story.md'
    story_file.write_text(f"# {scene_id}\n\nTest story content")

# Define ARGS
ARGS = {
    'base_dir': base_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_scene_public Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("234aa5fbbbf6b308")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path
import json
import pandas as pd

# Create a temporary directory structure to simulate a scene
with tempfile.TemporaryDirectory() as tmpdir:
    scene_dir = Path(tmpdir) / "scene_001"
    scene_dir.mkdir(parents=True, exist_ok=True)
    
    # Create story.md
    story_path = scene_dir / "story.md"
    story_path.write_text("This is a test story about a scene.", encoding='utf-8')
    
    # Create variants directory
    variants_dir = scene_dir / "variants" / "variant_1"
    variants_dir.mkdir(parents=True, exist_ok=True)
    
    # Create schema.json
    schema_path = variants_dir / "schema.json"
    schema = {
        "version": "1.0",
        "fields": ["id", "name", "value"]
    }
    schema_path.write_text(json.dumps(schema), encoding='utf-8')
    
    # Create data.parquet
    data_path = variants_dir / "data.parquet"
    df_data = pd.DataFrame({
        "id": [1, 2, 3],
        "name": ["alice", "bob", "charlie"],
        "value": [10.5, 20.3, 15.7]
    })
    df_data.to_parquet(data_path)
    
    # Create calibration.parquet
    calibration_path = variants_dir / "calibration.parquet"
    df_calib = pd.DataFrame({
        "id": [1, 2],
        "calibration_value": [0.95, 0.98]
    })
    df_calib.to_parquet(calibration_path)
    
    # Create tasks.json
    tasks_path = variants_dir / "tasks.json"
    tasks = {
        "task_1": {"description": "Task 1", "difficulty": "easy"},
        "task_2": {"description": "Task 2", "difficulty": "hard"}
    }
    tasks_path.write_text(json.dumps(tasks), encoding='utf-8')
    
    # Define ARGS with the scene_dir and observation_variant
    ARGS = {
        "scene_dir": str(scene_dir),
        "observation_variant": "variant_1"
    }
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_outcome_type Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("263573ac913eebcf")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned OutcomeType. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from enum import Enum

class OutcomeType(str, Enum):
    """Outcome variable type."""
    UNKNOWN = 'unknown'
    CONTINUOUS = 'continuous'
    BINARY = 'binary'
    CATEGORICAL = 'categorical'

    def __str__(self) -> str:
        return self.value

def parse_outcome_type(value):
    """Parse an outcome type into an OutcomeType enum."""
    if isinstance(value, OutcomeType):
        return value
    if value is None:
        return OutcomeType.UNKNOWN
    if isinstance(value, str):
        return OutcomeType(value.strip().lower())
    raise ValueError(f'Unrecognized outcome_type: {value!r}')

# Test case 1: OutcomeType enum instance
test_value_1 = OutcomeType.BINARY

# Test case 2: None value
test_value_2 = None

# Test case 3: String value
test_value_3 = "continuous"

# Test case 4: String with whitespace and mixed case
test_value_4 = "  CATEGORICAL  "

ARGS = {
    "value": OutcomeType.CONTINUOUS
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type OutcomeType — a OutcomeType output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a OutcomeType output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_rung Ran
ran, honoured the contract we drafted
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("0986fc61d9da1545")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Rung. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from enum import Enum, IntEnum

class Rung(IntEnum):
    """Pearl hierarchy rung."""
    R1 = 1
    R2 = 2
    R3 = 3

    def __str__(self) -> str:
        return f'R{int(self)}'

def parse_rung(value):
    """Parse a rung value into a Rung enum."""
    if isinstance(value, Rung):
        return value
    if isinstance(value, int):
        return Rung(value)
    if isinstance(value, str):
        v = value.strip().lower()
        if v in {'r1', '1'}:
            return Rung.R1
        if v in {'r2', '2'}:
            return Rung.R2
        if v in {'r3', '3'}:
            return Rung.R3
    raise ValueError(f'Unrecognized rung value: {value!r}')

ARGS = {
    'value': 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Rung — a Rung output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Rung output the recorded ndim 0 and dtype family int describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_task_type Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("c7f4edee0b323065")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned TaskType. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from typing import Any, Dict, List, Optional
from enum import Enum

class TaskType(str, Enum):
    """Question family (semantic intent)."""
    PREDICTION = 'prediction'
    ASSOCIATION = 'association'
    COLLIDER_PHENOMENON = 'collider_phenomenon'
    CAUSAL_SKETCH = 'causal_sketch'
    IDENTIFICATION = 'identification'
    EFFECT_ESTIMATE = 'effect_estimate'
    BIAS_DIAGNOSTIC = 'bias_diagnostic'
    COUNTERFACTUAL_IDENTIFICATION = 'counterfactual_identification'
    COUNTERFACTUAL_EFFECT = 'counterfactual_effect'
    MEDIATION_EFFECT = 'mediation_effect'

    def __str__(self) -> str:
        return self.value

def parse_task_type(value: Any) -> TaskType:
    """Parse a task type into a TaskType enum."""
    if isinstance(value, TaskType):
        return value
    if isinstance(value, str):
        v = value.strip().lower()
        return TaskType(v)
    raise ValueError(f'Unrecognized task_type: {value!r}')


# Test with a string value
ARGS = {
    'value': 'prediction'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TaskType — a TaskType output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a TaskType output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
resolve_structural_label Ran
ran; our contract draft was wrong, not the code
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("0a80820cffeef210")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from typing import Any, Dict, List, Optional

# Case 1: metadata with structural_label
metadata_with_label = {
    'structural_label': 'arch1',
    'motif': 'fallback_motif'
}

# Case 2: graph_info with applied_grafts > 0
graph_info_with_grafts = {
    'augmentation': {
        'applied_grafts': 3
    },
    'main_graph': {
        'motif': 'main_motif'
    }
}

# Case 3: graph_info with auxiliary_graph_grafts list
graph_info_with_auxiliary = {
    'auxiliary_graph_grafts': ['graft1', 'graft2'],
    'main_graph': {
        'motif': 'backup_motif'
    }
}

# Case 4: metadata with motif
metadata_with_motif = {
    'motif': 'scene_motif'
}

# Case 5: graph_info with main_graph.motif
graph_info_with_main = {
    'main_graph': {
        'motif': 'main_scene_motif'
    }
}

# Case 6: empty/None inputs
empty_metadata = {}
empty_graph_info = {}

# Primary test case: metadata with structural_label takes precedence
ARGS = {
    'metadata': metadata_with_label,
    'graph_info': graph_info_with_grafts
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
Exam Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("0188381c0a0c1981")
How it ran

no invocation recorded.

ExamItem Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("ae97603a464ef027")
How it ran

no invocation recorded.

OutcomeType Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("f8896f6aa98efe68")
How it ran

no invocation recorded.

OutputVariant Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("010c920e7dcd9a49")
How it ran

no invocation recorded.

Rung Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("51a4180de1afbc91")
How it ran

no invocation recorded.

TaskInputMode Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("fa486f10ca92e91f")
How it ran

no invocation recorded.

_apply_rung_tilt Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("321ae7d5014ac05c")
How it ran

no invocation recorded.

_assign_opaque_answer_file_stems Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("8e248affd3e201fe")
How it ran

no invocation recorded.

_normalize_composition Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("e795869ed09af8ac")
How it ran

no invocation recorded.

_resolve_variant_weights Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("26c8d8b2a34e4fbc")
How it ran

no invocation recorded.

build_exam Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("8b6c01925666d449")
How it ran

no invocation recorded.

get_output_variant_difficulty_order Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("4c28f9830b3a77fe")
How it ran

no invocation recorded.

get_task_input_mode Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("cffe0153b32ba4e9")
How it ran

no invocation recorded.

infer_rung_from_task_type Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("c1f83096d08ad8bb")
How it ran

no invocation recorded.

normalize_task_fields Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("9112deda4fa73ebe")
How it ran

no invocation recorded.

parse_output_variant Not yet run
unverified (harvested, no recorded run)
andleb/causalds/causalds/exam_builder.py
code served (permissive licence) · get_code("e709ab15df436195")
How it ran

no invocation recorded.

Also present in other papers' code

Of 33 samples lifted from this paper's linked repositories, 33 carry a function identity and 0 do not. Of the 33 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 33 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Large language models (LLMs) increasingly act as integrated data-science agents, combining abstract reasoning with advanced tool use. Yet the relevant benchmark landscape largely divides into symbolic causal reasoning benchmarks without realistic data analysis or data analysis benchmarks without a principled causal data-generating structure. Furthermore, existing causal evaluation datasets are often restricted to curated examples from existing sources, with diversity coming from limited templatized variations rather than from systematic generation of novel synthetic causal structures. We introduce CausalDS, a benchmark for evaluating causal reasoning in agentic data-science workflows. Each benchmark instance is a scene consisting of a sampled structural causal model (SCM) with generated observational data and an accompanying synthetic natural-language story grounded in a realistic domain. We optionally ground the composition of the benchmark components in empirical distributions obtained from real-world datasets, thus retaining empirical structure while reducing the "causal parrot" risk through completely synthetic generation. From each scene, we then derive tasks spanning all three of Pearl's rungs, with typical data-science prediction tasks appearing as Rung 1. Most tasks include a data science coding component, where the model typically needs to use several tools to arrive at the final answer due to the frequent presence of imperfect observations, which are generated by an observation model. Additionally, recognizing when a question admits no warranted answer and abstaining is treated as a first-class scored outcome. The benchmark thus jointly evaluates symbolic causal reasoning, data science, uncertainty quantification, abstention, and tool use/coding.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2607.08093")
get_code_for_paper("2607.08093")
have("2607.08093")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2607.08093.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2607.08093.svg)](https://syntology.ai/paper/2607.08093)

Connect an agent — have() is free.