SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.30003 · 2026

Discovering Cooperative Pipelines: Autoresearch for Sequential Social Dilemmas

Víctor Gallego

arXiv · PDF · Open in the Atlas

Code that ran

12 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 5 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 7 failures of our instrument; 3 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/2605.30003.json · badge: /api/ran/2605.30003.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
vicgalle/autoresearch-social-dilemmas canonical (classifier label)
classifier basis: census
12 ran · 3 no recorded run
FunctionStatusWhere it lives
extract_policy_code Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/llm_self_play.py
code served (permissive licence) · get_code("71812e35b60a8a80")
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 re

# Test case 1: LLM output with ```python``` block containing def policy
text_with_python_block = """
Here is the policy implementation:


def policy(env, agent_id):
    observation = env.observe(agent_id)
    action = env.action_space.sample()
    return action


This policy randomly samples actions.
"""

# Test case 2: LLM output with generic ``` block containing def policy
text_with_generic_block = """
The policy code:


def policy(env, agent_id):
    return 0

"""

# Test case 3: LLM output with def policy not in code blocks
text_with_bare_policy = """
def policy(env, agent_id):
    observation = env.observe(agent_id)
    return env.action_space.sample()

def helper():
    pass
"""

# Test case 4: Text with no policy
text_no_policy = """
This is just some random text about policies in general,
but no code implementation here.
"""

# Use the first test case as the primary argument
ARGS = {
    "text": text_with_python_block
}
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
get_opponents Ran
ran, honoured the contract we drafted
vicgalle/autoresearch-social-dilemmas/llm_self_play.py
code served (permissive licence) · get_code("4e913c04af0e4c4d")
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 2, dtype family int; 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 torch
import numpy as np

class MockEnv:
    def __init__(self):
        # n_agents: number of agents in the environment
        self.n_agents = 4
        
        # agent_pos: position of each agent [agent_id] -> (row, col)
        self.agent_pos = np.array([
            [0.0, 1.0],      # agent 0
            [2.0, 3.0],      # agent 1
            [1.0, 1.0],      # agent 2
            [5.0, 5.0],      # agent 3
        ])
        
        # agent_timeout: timeout counter for each agent
        self.agent_timeout = np.array([0.0, 1.0, 0.0, 0.0])
        
        # agent_beam_hits: beam hit count for each agent
        self.agent_beam_hits = np.array([2.0, 3.0, 1.0, 4.0])

env = MockEnv()
agent_id = 0

ARGS = {
    "env": env,
    "agent_id": agent_id,
}
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 list, ndim 2, dtype family int — 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
gift Ran
ran, honoured the contract we drafted
vicgalle/autoresearch-social-dilemmas/nested_commons_env.py
code served (permissive licence) · get_code("b5d27a2598c26cfb")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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

torch.manual_seed(0)

ARGS = {
    "target_id": 5
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
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_results Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/autoresearch/analyze.py
code served (permissive licence) · get_code("c837be730c58a806")
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
import csv
import os

# Create a temporary CSV file with sample data
temp_dir = tempfile.mkdtemp()
filepath = os.path.join(temp_dir, "results.tsv")

# Write sample TSV data
with open(filepath, 'w', newline='') as f:
    writer = csv.DictWriter(f, delimiter='\t', fieldnames=[
        'experiment', 'efficiency', 'equality', 'sustainability', 'peace', 
        'reward_avg', 'delta_eff', 'run_time_s'
    ])
    writer.writeheader()
    writer.writerow({
        'experiment': '1',
        'efficiency': '0.85',
        'equality': '0.72',
        'sustainability': '0.91',
        'peace': '0.68',
        'reward_avg': '42.5',
        'delta_eff': '+0.15',
        'run_time_s': '123.45'
    })
    writer.writerow({
        'experiment': '2',
        'efficiency': '0.78',
        'equality': '0.65',
        'sustainability': '0.88',
        'peace': '0.70',
        'reward_avg': '39.2',
        'delta_eff': '-0.05',
        'run_time_s': '98.76'
    })
    writer.writerow({
        'experiment': '3',
        'efficiency': '0.92',
        'equality': '',
        'sustainability': '',
        'peace': '',
        'reward_avg': '',
        'delta_eff': 'n/a',
        'run_time_s': ''
    })

ARGS = {
    'filepath': filepath
}
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
make_gathering Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/gathering_env.py
code served (permissive licence) · get_code("e738fb1b1ffcf0b9")
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 GatheringEnv. 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 numpy as np
from enum import IntEnum
from typing import Dict, List, Optional, Tuple

class Action(IntEnum):
    """Eight agent-centered actions (same in both papers)."""
    FORWARD = 0
    BACKWARD = 1
    STEP_LEFT = 2
    STEP_RIGHT = 3
    ROTATE_LEFT = 4
    ROTATE_RIGHT = 5
    BEAM = 6
    STAND = 7

NUM_ACTIONS = len(Action)

class Orientation(IntEnum):
    NORTH = 0
    EAST = 1
    SOUTH = 2
    WEST = 3

_ROTATIONS = {
    Orientation.NORTH: (-1, 0, 0, 1),
    Orientation.EAST: (0, 1, 1, 0),
    Orientation.SOUTH: (1, 0, 0, -1),
    Orientation.WEST: (0, -1, -1, 0)
}

COLOUR = {
    'background': np.array([0, 0, 0], dtype=np.uint8),
    'wall': np.array([127, 127, 127], dtype=np.uint8),
    'apple': np.array([0, 200, 0], dtype=np.uint8),
    'self': np.array([50, 100, 255], dtype=np.uint8),
    'other': np.array([220, 50, 50], dtype=np.uint8),
    'beam': np.array([255, 255, 60], dtype=np.uint8),
    'timeout': np.array([100, 100, 100], dtype=np.uint8)
}

_DENSITY_RESPAWN_PROBS = {0: 0.0, 1: 0.01, 2: 0.01, 3: 0.05, 4: 0.05}
_DENSITY_RESPAWN_DEFAULT = 0.1

GATHERING_MAP_SMALL = '@@@@@@@@@@@@@@@@@@\n@P.....A.A.....P@\n@......A.A......@\n@......A.A......@\n@......A.A......@\n@......A.A......@\n@......A.A......@\n@......A.A......@\n@................@\n@@@@@@@@@@@@@@@@@@'

GATHERING_MAP = '@@@@@@@@@@@@@@@@@@@@@@@@@\n@P.........AAA.........@\n@..........AAA.........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@...........................@\n@...........................@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA..........@\n@..........AAA.........P@\n@@@@@@@@@@@@@@@@@@@@@@@@@'

ARGS = {
    'ascii_map': GATHERING_MAP_SMALL,
    'n_agents': 2,
    'max_steps': 100,
    'obs_ahead': 15,
    'obs_side': 10,
    'beam_length': 20,
    'beam_width': 1,
    'hits_to_tag': 2,
    'timeout_steps': 25,
    'respawn_mode': 'fixed',
    'apple_respawn_time': 50,
    'density_radius': 2,
    'beam_enabled': True,
    'seed': 42
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GatheringEnv — a GatheringEnv output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a GatheringEnv 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
make_production_economy Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/production_economy_env.py
code served (permissive licence) · get_code("753b8570fab81744")
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 ProductionEconomyEnv. 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 numpy as np
from enum import IntEnum
from typing import Dict, List, Optional, Tuple

class Action(IntEnum):
    """Seventeen discrete actions for the Production Economy game."""
    NOOP = 0
    MOVE_N = 1
    MOVE_S = 2
    MOVE_E = 3
    MOVE_W = 4
    GATHER = 5
    CRAFT = 6
    CRAFT_TOOL = 7
    CRAFT_SHELTER = 8
    DROP_WOOD = 9
    DROP_STONE = 10
    DROP_PLANK = 11
    DROP_BRICK = 12
    PICKUP_WOOD = 13
    PICKUP_STONE = 14
    PICKUP_PLANK = 15
    PICKUP_BRICK = 16

NUM_ACTIONS = len(Action)

WOOD = 0
STONE = 1
PLANK = 2
BRICK = 3
NUM_ITEM_TYPES = 4

FOREST = 0
QUARRY = 1

_MOVE_OFFSETS = {Action.MOVE_N: (-1, 0), Action.MOVE_S: (1, 0), Action.MOVE_E: (0, 1), Action.MOVE_W: (0, -1)}

_DROP_ITEM = {Action.DROP_WOOD: WOOD, Action.DROP_STONE: STONE, Action.DROP_PLANK: PLANK, Action.DROP_BRICK: BRICK}

_PICKUP_ITEM = {Action.PICKUP_WOOD: WOOD, Action.PICKUP_STONE: STONE, Action.PICKUP_PLANK: PLANK, Action.PICKUP_BRICK: BRICK}

PRODUCTION_ECONOMY_MAP = 'P.F....P....F.P\n....F..........\n....F..........\n.........F.....\n.F.............\n...............\nP..S...G...S..P\nP......G......P\n...M.......M...\n...............\n.....Q.........\n....Q......Q...\n...............\n.Q......Q......\nP......Q......P'

def _parse_map(ascii_map: str):
    """Parse the fixed ASCII map into feature arrays."""
    lines = ascii_map.strip().splitlines()
    height = len(lines)
    width = max((len(ln) for ln in lines))
    lines = [ln.ljust(width) for ln in lines]
    forest: List[Tuple[int, int]] = []
    quarry: List[Tuple[int, int]] = []
    sawmill: List[Tuple[int, int]] = []
    masonry: List[Tuple[int, int]] = []
    forge: List[Tuple[int, int]] = []
    spawn: List[Tuple[int, int]] = []
    for r, ln in enumerate(lines):
        for c, ch in enumerate(ln):
            if ch == 'F':
                forest.append((r, c))
            elif ch == 'Q':
                quarry.append((r, c))
            elif ch == 'S':
                sawmill.append((r, c))
            elif ch == 'M':
                masonry.append((r, c))
            elif ch == 'G':
                forge.append((r, c))
            elif ch == 'P':
                spawn.append((r, c))
    return (height, width, forest, quarry, sawmill, masonry, forge, spawn)

ARGS = {
    'ascii_map': PRODUCTION_ECONOMY_MAP,
    'n_agents': 8,
    'max_steps': 300,
    'winter_step': 200,
    'winter_threshold': 6,
    'winter_reward': 50.0,
    'T_spoil': 80,
    'tool_step_reward': 2.0,
    'tool_gather_bonus': 1,
    'inventory_capacity': 3,
    'cell_drop_capacity': 5,
    'resource_respawn_prob': 0.05,
    'initial_stocked_fraction': 0.8,
    'seed': 42,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ProductionEconomyEnv — a ProductionEconomyEnv output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ProductionEconomyEnv 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_cleanup_map Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/cleanup_env.py
code served (permissive licence) · get_code("ec5f25200b4fe84a")
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 tuple. 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 numpy as np
from typing import Dict, List, Optional, Tuple

# Create a simple ASCII map string for testing
ascii_map = """
@@@@@@@@@@
@P  A   @
@   H   @
@   R   @
@   S   @
@@@@@@@@@@
"""

ARGS = {
    "ascii_map": ascii_map
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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_map Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/gathering_env.py
code served (permissive licence) · get_code("0ee896b5cd691fee")
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 tuple. 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 numpy as np

ARGS = {
    "ascii_map": """@@@@@@@@@@
@P  A   @
@       @
@   A   @
@@@@@@@@@@"""
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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_mining_map Ran
ran; our contract draft was wrong, not the code
vicgalle/autoresearch-social-dilemmas/coop_mining_env.py
code served (permissive licence) · get_code("33eaa2604e2d78e8")
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 tuple. 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 numpy as np
from typing import Dict, List, Optional, Tuple

# Construct a valid ASCII map string for the parse_mining_map function
ascii_map = """
@@@@@@@@@@
@P  I   @
@   G   @
@       @
@@@@@@@@@@
"""

ARGS = {
    "ascii_map": ascii_map
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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
raid Ran
ran, honoured the contract we drafted
vicgalle/autoresearch-social-dilemmas/nested_commons_env.py
code served (permissive licence) · get_code("5553e251f5b1acce")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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

ARGS = {
    "target_id": 5
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
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
travel Ran
ran, honoured the contract we drafted
vicgalle/autoresearch-social-dilemmas/nested_commons_env.py
code served (permissive licence) · get_code("243f9e3eee0378ed")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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

# The function travel takes a quadrant parameter (int) and returns an int
# quadrant should be in range 0..3 (representing A, B, C, D)
# The function simply adds quadrant to TRAVEL_BASE (38)

ARGS = {
    "quadrant": 0
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
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
validate_code_safety Ran
ran, honoured the contract we drafted
vicgalle/autoresearch-social-dilemmas/llm_self_play.py
code served (permissive licence) · get_code("32fe42a6f676d009")
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 ast

# Test case 1: Safe code with no violations
safe_code = """
x = 5
y = x + 10
print(y)
"""

# Test case 2: Code with forbidden import
forbidden_import_code = """
import os
print(os.getcwd())
"""

# Test case 3: Code with forbidden call
forbidden_call_code = """
code = "x = 1"
eval(code)
"""

# Test case 4: Code with syntax error
syntax_error_code = """
def foo(
    x = 1
"""

# Test case 5: Code with multiple violations
multi_violation_code = """
import sys
import subprocess
exec("print('hello')")
"""

# Test case 6: Code with forbidden attribute call
forbidden_attr_code = """
obj = object()
obj.__import__('os')
"""

ARGS = {
    "code": safe_code
}
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
make_coop_mining Not yet run
unverified (harvested, no recorded run)
vicgalle/autoresearch-social-dilemmas/coop_mining_env.py
code served (permissive licence) · get_code("90fdef34ff30e86d")
How it ran

no invocation recorded.

make_coop_mining_large Not yet run
unverified (harvested, no recorded run)
vicgalle/autoresearch-social-dilemmas/coop_mining_env.py
code served (permissive licence) · get_code("e8c883c639e16ccb")
How it ran

no invocation recorded.

make_gathering_large Not yet run
unverified (harvested, no recorded run)
vicgalle/autoresearch-social-dilemmas/gathering_env.py
code served (permissive licence) · get_code("bd5dc108a5391524")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 7 are also present in other papers whose code we hold (7 with at least one paper dated earlier, 0 the same month, 0 later) and 8 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.

FunctionAlso present inEarliest-datedOthers named
extract_policy_code 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas
get_opponents 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas
make_gathering 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas
make_gathering_large 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas
parse_cleanup_map 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas
parse_map 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas
validate_code_safety 1 other paper
1 dated earlier
2603.19453 (2026-03)
Beyond Scalar Rewards: Dense Feedback for LLM Policy Synthesis in Sequential Social Dilemmas

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

We study two-level autoresearch for cooperation: an outer-loop AI agent autonomously redesigns the inner-loop pipeline of an LLM policy-synthesis system for multi-agent Sequential Social Dilemmas (SSDs). A researcher agent R (run as a coding agent) reads the inner-loop source code, edits system prompts, feedback functions, helper libraries, and iteration logic, runs evaluations, and decides what to keep, following the autoresearch paradigm. Across two games (Cleanup and Gathering), two policy-synthesizer LLMs, and two welfare objectives (utilitarian efficiency and Rawlsian maximin), the researcher reliably exceeds hand-designed baselines, sharply tightens run-to-run variance, and outperforms prompt-only optimization. The discovered pipelines are objective-dependent: only under maximin does the researcher inject an explicit fairness mechanism into synthesizer pipelines, a class of mechanism that is absent from its own objective-agnostic system prompt and from every efficiency-optimized pipeline. This supports an information-design reading in which the researcher chooses what to reveal to the boundedly rational synthesizer as a function of the welfare objective. Code at https://github.com/vicgalle/autoresearch-social-dilemmas.

For agents

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

get_harvested_code_for_paper("2605.30003")
get_code_for_paper("2605.30003")
have("2605.30003")

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

curl https://syntology.ai/api/ran/2605.30003.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.30003.svg)](https://syntology.ai/paper/2605.30003)

Connect an agent — have() is free.