SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.04749 · 2026

COP-Q: Safety-First Reinforcement Learning for Robot Control via Cholesky-Ordered Projection

Julian Kooij, Moritz Zanger, Matthijs Spaan, Guopeng Li

arXiv · PDF · Open in the Atlas

Code that ran

11 of 13 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; 6 failures of our instrument; 2 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/2606.04749.json · badge: /api/ran/2606.04749.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
RomainLITUD/COPQ canonical (classifier label)
classifier basis: census
11 ran · 2 no recorded run
FunctionStatusWhere it lives
cholesky_stable Ran
ran; our contract draft was wrong, not the code
RomainLITUD/COPQ/copq-brax/sac/losses.py
code served (permissive licence) · get_code("d52c599bfaef8d3b")
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 ArrayImpl, ndim 2, dtype family float, shape (2, 2) 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 jax.numpy as jnp
import numpy as np

# Create a small 2x2 covariance matrix
# A covariance matrix should be symmetric and positive semi-definite
cov_data = np.array([[2.0, 0.5], [0.5, 1.5]], dtype=np.float32)
cov = jnp.array(cov_data)

ARGS = {
    "cov": cov,
    "delta_abs": 1e-08,
    "delta_rel": 1e-05,
    "rho_max": 0.9,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 2, dtype family float, shape (2, 2) — shape (2, 2) 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
epistemic_steiner Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
RomainLITUD/COPQ/copq-brax/sac/losses.py
code served (permissive licence) · get_code("f8ffa2d2d7a22222")
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 ArrayImpl, ndim 2, dtype family float, shape (2, 2) 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 jax
import jax.numpy as jnp
import numpy as np

# Analyze the function:
# epistemic_steiner takes:
# - vertices_batch: batched array of vertices, passed to vmap(process)
#   - process expects vertices with shape (2, N) where it computes:
#     - centroid = mean over axis 1, shape (2,)
#     - X_centered shape (2, N)
#     - cov shape (2, 2)
#   - process returns shape (2,) intersection point
# - w: parameter passed to process, shape (2,) based on jnp.matmul(L, w)
#   - L is (2, 2), so w should be (2,)

# Build vertices_batch with shape (batch_size, 2, num_vertices)
# Use batch_size=2, num_vertices=4 for distinct dimensions
batch_size = 2
num_vertices = 4
feat_dim = 2

# Create vertices_batch: shape (2, 2, 4)
# Each item in batch has 2 features and 4 vertices
vertices_batch = jnp.array([
    [[1.0, 2.0, 3.0, 4.0],
     [5.0, 6.0, 7.0, 8.0]],
    [[0.5, 1.5, 2.5, 3.5],
     [4.5, 5.5, 6.5, 7.5]]
])

# Create w: shape (2,)
w = jnp.array([0.1, 0.2])

ARGS = {
    "vertices_batch": vertices_batch,
    "w": w
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 2, dtype family float, shape (2, 2) — shape (2, 2) 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
get_multi_rewards Ran
ran, honoured the contract we drafted
RomainLITUD/COPQ/copq-brax/sac/acting.py
code served (permissive licence) · get_code("b1f7d21dc0cc9c97")
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 ArrayImpl, ndim 1, dtype family float, shape (2,) 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 jax.numpy as jnp

class MockNState:
    def __init__(self, metrics):
        self.metrics = metrics

# For 'hopper' environment
nstate_hopper = MockNState({
    'reward_healthy': 1.0,
    'reward_forward': 2.5,
    'reward_ctrl': 0.5
})

# For 'walker2d' environment
nstate_walker2d = MockNState({
    'reward_healthy': 1.0,
    'reward_forward': 2.5,
    'reward_ctrl': 0.5
})

# For 'ant' environment
nstate_ant = MockNState({
    'reward_survive': 1.0,
    'reward_forward': 2.5,
    'reward_ctrl': 0.5
})

# For other environments
nstate_other = MockNState({
    'reward_alive': 1.0,
    'forward_reward': 2.0,
    'reward_quadctrl': 0.3
})

ARGS = {
    'env_name': 'hopper',
    'nstate': nstate_hopper
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (2,) — shape (2,) 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
get_multi_rewards_easy Ran
ran, honoured the contract we drafted
RomainLITUD/COPQ/copq-brax/sac/acting.py
code served (permissive licence) · get_code("2d87a975db085464")
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 ArrayImpl, ndim 1, dtype family float, shape (2,) 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 jax.numpy as jnp

class NState:
    def __init__(self, env_name):
        self.metrics = {}
        if env_name == 'hopper':
            self.metrics['reward_ctrl'] = 0.1
            self.metrics['reward_healthy'] = 0.5
            self.metrics['reward_forward'] = 1.0
        elif env_name == 'walker2d':
            self.metrics['reward_ctrl'] = 0.2
            self.metrics['reward_healthy'] = 0.3
            self.metrics['reward_forward'] = 0.8
        elif env_name == 'ant':
            self.metrics['reward_ctrl'] = 0.15
            self.metrics['reward_survive'] = 0.4
            self.metrics['reward_forward'] = 1.2
        else:
            self.metrics['reward_quadctrl'] = 0.1
            self.metrics['reward_alive'] = 0.6
            self.metrics['forward_reward'] = 0.9

nstate_hopper = NState('hopper')
nstate_walker2d = NState('walker2d')
nstate_ant = NState('ant')
nstate_other = NState('humanoid')

ARGS = {
    'env_name': 'hopper',
    'nstate': nstate_hopper
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (2,) — shape (2,) 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
get_multi_rewards_old Ran
ran, honoured the contract we drafted
RomainLITUD/COPQ/copq-brax/sac/acting.py
code served (permissive licence) · get_code("cb37087dd5fdb979")
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 ArrayImpl, ndim 1, dtype family float, shape (2,) 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 jax.numpy as jnp

class MockState:
    def __init__(self, metrics):
        self.metrics = metrics

# Create mock nstate objects with metrics for different environments
mock_metrics_hopper = {
    'reward_forward': 1.5,
    'reward': 2.3
}

mock_nstate_hopper = MockState(mock_metrics_hopper)

mock_metrics_walker2d = {
    'reward_forward': 1.2,
    'reward': 2.1
}

mock_nstate_walker2d = MockState(mock_metrics_walker2d)

mock_metrics_ant = {
    'velocity_total': 0.8,
    'reward': 1.9
}

mock_nstate_ant = MockState(mock_metrics_ant)

mock_metrics_other = {
    'reward_linvel': 1.1,
    'reward': 2.5
}

mock_nstate_other = MockState(mock_metrics_other)

# Test with hopper environment
ARGS = {
    'env_name': 'hopper',
    'nstate': mock_nstate_hopper
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (2,) — shape (2,) 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
identity_observation_preprocessor Ran
ran, honoured the contract we drafted
RomainLITUD/COPQ/copq-brax/sac/types.py
code served (permissive licence) · get_code("736e12b7afdd2635")
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 ArrayImpl, ndim 2, dtype family float, shape (2, 4) 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 jax.numpy as jnp

# Create a simple observation as a JAX array
observation_array = jnp.array([[1.0, 2.0, 3.0, 4.0],
                                [5.0, 6.0, 7.0, 8.0]])

# Create preprocessor_params as None (or any value, since it's deleted in the function)
preprocessor_params = None

ARGS = {
    "observation": observation_array,
    "preprocessor_params": preprocessor_params
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) 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
loss_and_jacobian Ran
ran; our contract draft was wrong, not the code
RomainLITUD/COPQ/copq-brax/sac/gradient_surgery.py
code served (permissive licence) · get_code("03b2dbb3d9fb7a68")
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 function. 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 jax
import jax.numpy as jnp
import numpy as np

def simple_loss_fn(x, y):
    """A simple loss function that returns loss and auxiliary data."""
    # Compute a scalar loss value
    loss = jnp.sum((x - y) ** 2)
    # Return loss and auxiliary data
    aux_data = {"iteration": 1}
    return loss, aux_data

# Create simple test inputs
x = jnp.array([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]])
y = jnp.array([[1.1, 2.1, 3.1, 4.1], [5.1, 6.1, 7.1, 8.1]])

ARGS = {
    "loss_fn": simple_loss_fn,
    "pmap_axis_name": None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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_gpu_ids Ran
ran, honoured the contract we drafted
RomainLITUD/COPQ/copq-omni/offpolicy_benchmark.py
code served (permissive licence) · get_code("9be66500e00bd3e2")
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 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

# Define ARGS dict with valid arguments for parse_gpu_ids
ARGS = {
    "value": "0,1,2"
}
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 1, 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
stable_cholesky_3d Ran
ran; our contract draft was wrong, not the code
RomainLITUD/COPQ/copq-brax/sac/losses.py
code served (permissive licence) · get_code("d5f9b86af95a5529")
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 ArrayImpl, ndim 2, dtype family float, shape (3, 3) 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 jax.numpy as jnp
import numpy as np

# Construct a valid 3x3 positive definite covariance matrix
# Use a simple symmetric positive definite matrix
np.random.seed(0)
A = np.random.randn(3, 3)
C_np = A @ A.T  # This ensures positive definiteness
C = jnp.array(C_np, dtype=jnp.float64)

ARGS = {
    "C": C,
    "alpha": 0.05,
    "delta_abs": 1e-06,
    "delta_rel": 1e-06,
    "dtype": jnp.float64,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 2, dtype family float, shape (3, 3) — shape (3, 3) 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
value_and_jacobian Ran
ran; our contract draft was wrong, not the code
RomainLITUD/COPQ/copq-brax/sac/gradient_surgery.py
code served (permissive licence) · get_code("5101b6c7d1ec47e9")
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 function. 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 jax
import jax.numpy as jnp

# Define a simple function that returns a value and auxiliary data
def simple_fun(x, y):
    """A test function that returns (value, aux) tuple."""
    # value is sum of x and y
    value = jnp.sum(x) + jnp.sum(y)
    # aux is some auxiliary data
    aux = {"info": "test"}
    return value, aux

# Create small test arrays
x = jnp.array([[1.0, 2.0, 3.0, 4.0],
               [5.0, 6.0, 7.0, 8.0]])  # shape (2, 4)

y = jnp.array([[0.1, 0.2, 0.3, 0.4],
               [0.5, 0.6, 0.7, 0.8]])  # shape (2, 4)

ARGS = {
    "fun": simple_fun
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
visualize Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
RomainLITUD/COPQ/copq-brax/vis_utils.py
code served (permissive licence) · get_code("9656d45158b8bcfa")
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 Figure. 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
import matplotlib.pyplot as plt
from unittest.mock import patch, MagicMock
import os
import tempfile

# Create temporary directory structure for results
temp_dir = tempfile.mkdtemp()
results_dir = os.path.join(temp_dir, 'results')
os.makedirs(results_dir, exist_ok=True)

env_list = ['hopper', 'walker2d', 'ant', 'humanoid']
model_list = ['model_a', 'model_b']

# Create mock .npz files for each environment and model
for env in env_list:
    env_dir = os.path.join(results_dir, env)
    os.makedirs(env_dir, exist_ok=True)
    
    for model in model_list:
        # Create sample data with shape (num_seeds, num_steps, 2)
        # x: (1, num_steps) - environment steps
        # y: (num_seeds, num_steps, 2) - [cost, reward]
        num_steps = 100
        num_seeds = 3
        
        x = np.array([np.linspace(0, 1e7, num_steps)])  # (1, num_steps)
        y = np.random.rand(num_seeds, num_steps, 2) * 1000  # (num_seeds, num_steps, 2)
        
        filepath = os.path.join(env_dir, f'{model}.npz')
        np.savez(filepath, x=x, y=y)

# Patch the file loading to use our temp directory
original_load = np.load

def patched_load(path, **kwargs):
    # Replace './results/' with our temp results path
    if './results/' in path:
        path = path.replace('./results/', results_dir + '/')
    return original_load(path, **kwargs)

ARGS = {
    'model_list': ['model_a', 'model_b'],
    'label_list': ['Algorithm A', 'Algorithm B'],
    'colors': ['#1f77b4', '#ff7f0e'],
    'figsize': (10, 2),
    'anchor': (0, 0)
}

# Patch np.load before the function is called
np.load = patched_load
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Figure — a Figure output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Figure 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
compare_onpolicy Not yet run
unverified (harvested, no recorded run)
RomainLITUD/COPQ/copq-brax/vis_utils.py
code served (permissive licence) · get_code("e5ca9d58f091c3ac")
How it ran

no invocation recorded.

ratio_visualize Not yet run
unverified (harvested, no recorded run)
RomainLITUD/COPQ/copq-brax/vis_utils.py
code served (permissive licence) · get_code("a15967d4f38731d1")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 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 13 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

Safe robot control requires maximizing return while satisfying safety constraints. In off-policy safe reinforcement learning, reward and safety Q-values are commonly learned by separate critic ensembles, with uncertainty handled independently for each objective. This objective-wise treatment neglects inter-objective correlation and can lead to overly conservative value estimates, thereby reducing sample efficiency. To address this issue, we propose Cholesky-Ordered Projection Q-learning (COP-Q), a safety-first method that incorporates inter-objective covariance into vector-valued Q-value estimation. COP-Q constructs a generalized confidence bound in the joint Q-value space and uses Cholesky factorization to encode objective priority in a sequential form. This preserves conservatism on safety while adaptively reducing excessive conservatism on the reward objective. The resulting estimate is used in both temporal-difference target computation and actor optimization. COP-Q incurs minimal computational overhead and is readily compatible with most existing deep Q-learning frameworks. Experiments on robot locomotion in Brax and safe navigation in Safety-Gymnasium, covering both hard-and softsafety settings, demonstrate that COP-Q achieves strong safety performance together with competitive or improved sample efficiency relative to representative baselines.

For agents

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

get_harvested_code_for_paper("2606.04749")
get_code_for_paper("2606.04749")
have("2606.04749")

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

curl https://syntology.ai/api/ran/2606.04749.json

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

[![Syntology run record](https://syntology.ai/api/ran/2606.04749.svg)](https://syntology.ai/paper/2606.04749)

Connect an agent — have() is free.