SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2311.17138 · 2023

Shadows Don't Lie and Lines Can't Bend! Generative Models don't know Projective Geometry...for now

arXiv · PDF · Open in the Atlas

Code that ran

17 of 18 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; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 17 failures of our instrument; 1 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2311.17138.json · badge: /api/ran/2311.17138.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
hanlinm2/projective-geometry canonical (classifier label)
classifier basis: census
17 ran · 1 no recorded run
FunctionStatusWhere it lives
compute_roc_curve Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
hanlinm2/projective-geometry/prequalifier/prequalifier_test_helpers.py
code served (permissive licence) · get_code("a026c2d9978681cf")
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

# Construct arguments for compute_roc_curve
# The function expects:
# - all_labels: binary labels (0 or 1)
# - all_generated_probs: probability scores for the positive class

# Create small binary classification data
np.random.seed(0)

# Binary labels: array of 0s and 1s
all_labels = np.array([0, 1, 1, 0, 1, 0, 1, 1])

# Generated probabilities: scores between 0 and 1
all_generated_probs = np.array([0.1, 0.9, 0.8, 0.2, 0.7, 0.3, 0.6, 0.85])

ARGS = {
    "all_labels": all_labels,
    "all_generated_probs": all_generated_probs,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — 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
create_model Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_model.py
code served (permissive licence) · get_code("a945396ca0b16803")
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 ClassificationPointNet. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "target_device": torch.device("cpu")
}
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 ClassificationPointNet — a ClassificationPointNet output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ClassificationPointNet 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
create_model Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/perspective_fields/fields_model.py
code served (permissive licence) · get_code("f3c36d02045cfbc2")
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 FieldsClassifier. 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 torch.nn as nn
from torchvision.models import resnet50

torch.manual_seed(0)

target_device = torch.device('cpu')

ARGS = {
    'target_device': target_device
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FieldsClassifier — a FieldsClassifier output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a FieldsClassifier 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
evaluate_on_dataloader Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/prequalifier/prequalifier_test_helpers.py
code served (permissive licence) · get_code("949786004f8eb80c")
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. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn as nn

torch.manual_seed(0)

# Create a simple binary classification model
class SimpleClassifier(nn.Module):
    def __init__(self, input_dim=4, num_classes=2):
        super(SimpleClassifier, self).__init__()
        self.fc = nn.Linear(input_dim, num_classes)
    
    def forward(self, x):
        return self.fc(x)

# Instantiate the model
model = SimpleClassifier(input_dim=4, num_classes=2)

# Create a simple dataloader as a list of (images, labels) tuples
# Each image is a tensor of shape (batch_size, 4)
# Each labels tensor contains class indices (0 or 1)
test_dataloader = [
    (torch.randn(2, 4), torch.tensor([0, 1])),
    (torch.randn(2, 4), torch.tensor([1, 0])),
]

ARGS = {
    'model': model,
    'test_dataloader': test_dataloader,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm — 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
get_recent_timestamp_dataloaders Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/prequalifier/prequalifier_dataset.py
code served (permissive licence) · get_code("eb72ead7b49746e2")
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 torchvision.transforms as transforms
import math
from PIL import Image
import io
import numpy as np
from torch.utils.data import Dataset, DataLoader
import random
import tempfile
import os

torch.manual_seed(0)
np.random.seed(0)

# Create temporary test images
temp_dir = tempfile.mkdtemp()

# Create a simple test image
test_image = Image.new('RGB', (256, 256), color=(73, 109, 137))

# Create directory structure for dataset
real_dir = os.path.join(temp_dir, 'real')
fake_dir = os.path.join(temp_dir, 'fake')
os.makedirs(real_dir, exist_ok=True)
os.makedirs(fake_dir, exist_ok=True)

# Save test images
for i in range(2):
    test_image.save(os.path.join(real_dir, f'img_{i}.jpg'))
    test_image.save(os.path.join(fake_dir, f'img_{i}.jpg'))

# Prepare paths and class mapping
recent_timestamp_paths = {
    'timestamp_1': [
        os.path.join(real_dir, 'img_0.jpg'),
        os.path.join(real_dir, 'img_1.jpg'),
    ],
    'timestamp_2': [
        os.path.join(fake_dir, 'img_0.jpg'),
        os.path.join(fake_dir, 'img_1.jpg'),
    ]
}

class_to_idx = {
    'real': 0,
    'fake': 1
}

ARGS = {
    'recent_timestamp_paths': recent_timestamp_paths,
    'class_to_idx': class_to_idx
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch, torchvision — 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
get_test_dataloaders Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_dataset.py
code served (permissive licence) · get_code("7240b9840ca5e27e")
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, 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 torch
import numpy as np
from torch.utils.data import Dataset, DataLoader

torch.manual_seed(0)
np.random.seed(0)

# Create minimal test data
class_to_idx = {'class_a': 0, 'class_b': 1}

# Create image paths with class structure (mimics "path/to/class_name/image.jpg")
test_image_paths = [
    'path/to/class_a/img1.jpg',
    'path/to/class_a/img2.jpg',
    'path/to/class_b/img3.jpg',
    'path/to/class_b/img4.jpg',
    'path/to/class_a/img5.jpg',
]

unconfident_image_paths = [
    'path/to/class_a/img1.jpg',
    'path/to/class_b/img3.jpg',
]

misclassified_image_paths = [
    'path/to/class_b/img4.jpg',
]

# Create image_path_to_lines mapping
# Each entry should have shape (N, 4) where N is number of line segments
image_path_to_lines = {
    'path/to/class_a/img1.jpg': np.random.rand(10, 4).astype(np.float32),
    'path/to/class_a/img2.jpg': np.random.rand(15, 4).astype(np.float32),
    'path/to/class_b/img3.jpg': np.random.rand(8, 4).astype(np.float32),
    'path/to/class_b/img4.jpg': np.random.rand(12, 4).astype(np.float32),
    'path/to/class_a/img5.jpg': np.random.rand(20, 4).astype(np.float32),
}

ARGS = {
    'test_image_paths': test_image_paths,
    'unconfident_image_paths': unconfident_image_paths,
    'misclassified_image_paths': misclassified_image_paths,
    'image_path_to_lines': image_path_to_lines,
    'class_to_idx': class_to_idx,
}
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 tuple, ndim 1, dtype family float — a tuple 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
get_test_dataloaders Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/perspective_fields/fields_dataset.py
code served (permissive licence) · get_code("a52857a6650be07a")
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, 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 torch
from torch.utils.data import Dataset, DataLoader

torch.manual_seed(0)

# Create mock field data that will be "loaded" from paths
mock_fields = {}

def create_mock_field():
    return {
        'pred_latitude_original': torch.randn(4, 3),
        'pred_gravity_original': torch.randn(5, 4, 3)
    }

# Patch torch.load to return mock fields
original_load = torch.load
def mock_load(path):
    if path not in mock_fields:
        mock_fields[path] = create_mock_field()
    return mock_fields[path]

torch.load = mock_load

# Create test image paths with proper structure
test_image_paths = [
    "../dataset/test_Images/dir1/class_a/0.jpg",
    "../dataset/test_Images/dir1/class_a/1.jpg",
    "../dataset/test_Images/dir1/class_b/2.jpg",
    "../dataset/test_Images/dir1/class_b/3.jpg",
    "../dataset/test_Images/dir1/class_a/4.jpg",
]

unconfident_image_paths = [
    "../dataset/test_Images/dir1/class_a/0.jpg",
    "../dataset/test_Images/dir1/class_b/2.jpg",
]

misclassified_image_paths = [
    "../dataset/test_Images/dir1/class_a/4.jpg",
]

class_to_idx = {
    "class_a": 0,
    "class_b": 1,
}

ARGS = {
    "test_image_paths": test_image_paths,
    "unconfident_image_paths": unconfident_image_paths,
    "misclassified_image_paths": misclassified_image_paths,
    "class_to_idx": class_to_idx,
}
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 tuple, ndim 1, dtype family float — a tuple 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
get_test_dataloaders Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/prequalifier/prequalifier_dataset.py
code served (permissive licence) · get_code("627f7f54e5d786a4")
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, 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 torch
import numpy as np
from PIL import Image
import io
import random
import tempfile
import os

torch.manual_seed(0)
np.random.seed(0)
random.seed(0)

# Create temporary directory with minimal test images
temp_dir = tempfile.mkdtemp()

# Create minimal RGB images and save them
def create_test_image(path):
    """Create a minimal 256x256 RGB image"""
    img_array = np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8)
    img = Image.fromarray(img_array, 'RGB')
    img.save(path)

# Create directory structure for test images
real_dir = os.path.join(temp_dir, 'real')
fake_dir = os.path.join(temp_dir, 'fake')
os.makedirs(real_dir, exist_ok=True)
os.makedirs(fake_dir, exist_ok=True)

# Create test images
test_image_paths = []
unconfident_image_paths = []
misclassified_image_paths = []

for i in range(20):
    if i < 10:
        path = os.path.join(real_dir, f'img_{i}.jpg')
        test_image_paths.append(path)
    elif i < 15:
        path = os.path.join(real_dir, f'img_{i}.jpg')
        unconfident_image_paths.append(path)
    else:
        path = os.path.join(real_dir, f'img_{i}.jpg')
        misclassified_image_paths.append(path)
    create_test_image(path)

for i in range(20, 25):
    path = os.path.join(fake_dir, f'img_{i}.jpg')
    test_image_paths.append(path)
    create_test_image(path)

# Class mapping
class_to_idx = {
    'real': 0,
    'fake': 1
}

ARGS = {
    'test_image_paths': test_image_paths,
    'unconfident_image_paths': unconfident_image_paths,
    'misclassified_image_paths': misclassified_image_paths,
    'class_to_idx': class_to_idx
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple 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
get_train_dataloaders Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_dataset.py
code served (permissive licence) · get_code("6e612f3f3508395b")
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, 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 torch
import numpy as np

torch.manual_seed(0)
np.random.seed(0)

# Create small mock data
num_train_images = 2
num_val_images = 2
num_classes = 3

# Create image paths with class labels embedded in the path
train_image_paths = [
    "data/class_0/train_img_0.jpg",
    "data/class_1/train_img_1.jpg",
]

val_image_paths = [
    "data/class_2/val_img_0.jpg",
    "data/class_0/val_img_1.jpg",
]

# Create class to index mapping
class_to_idx = {
    "class_0": 0,
    "class_1": 1,
    "class_2": 2,
}

# Create image_path_to_lines dictionary
# Each image maps to a numpy array of line segments (shape: [num_lines, 4])
# 4 represents line endpoints (x1, y1, x2, y2)
image_path_to_lines = {}
for path in train_image_paths:
    num_lines = np.random.randint(100, 200)
    image_path_to_lines[path] = np.random.randn(num_lines, 4).astype(np.float32)

for path in val_image_paths:
    num_lines = np.random.randint(100, 200)
    image_path_to_lines[path] = np.random.randn(num_lines, 4).astype(np.float32)

ARGS = {
    "train_image_paths": train_image_paths,
    "val_image_paths": val_image_paths,
    "image_path_to_lines": image_path_to_lines,
    "class_to_idx": class_to_idx,
}
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 tuple, ndim 1, dtype family float — a tuple 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
get_train_dataloaders Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/perspective_fields/fields_dataset.py
code served (permissive licence) · get_code("59469a924d17ca67")
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, 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 torch
from torch.utils.data import Dataset, DataLoader

torch.manual_seed(0)

# Create mock image paths that match the expected directory structure
# The function expects paths like: ../dataset/{dataset_type}/{directory}/{identifier}/{number}.{ext}
train_image_paths = [
    "../dataset/train/dir1/class_A/image_0.jpg",
    "../dataset/train/dir1/class_A/image_1.jpg",
]

val_image_paths = [
    "../dataset/val/dir1/class_B/image_0.jpg",
    "../dataset/val/dir1/class_B/image_1.jpg",
]

# Class mapping
class_to_idx = {
    "class_A": 0,
    "class_B": 1,
}

# Mock the torch.load to return field data with the expected structure
import os
import tempfile

# Create temporary directory structure and mock field files
temp_dir = tempfile.mkdtemp()

def create_mock_field_file(field_path):
    os.makedirs(os.path.dirname(field_path), exist_ok=True)
    field_data = {
        'pred_latitude_original': torch.randn(1, 4, 3),  # (1, 4, 3)
        'pred_gravity_original': torch.randn(2, 4, 3),   # (2, 4, 3) - 2 gravity maps
    }
    torch.save(field_data, field_path)

# Patch torch.load to return mock data
original_load = torch.load
def mock_load(path, *args, **kwargs):
    if path.endswith('.pt'):
        return {
            'pred_latitude_original': torch.randn(1, 4, 3),
            'pred_gravity_original': torch.randn(2, 4, 3),
        }
    return original_load(path, *args, **kwargs)

torch.load = mock_load

ARGS = {
    "train_image_paths": train_image_paths,
    "val_image_paths": val_image_paths,
    "class_to_idx": class_to_idx,
}
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 tuple, ndim 1, dtype family float — a tuple 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_all_paths Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_path_helpers.py
code served (permissive licence) · get_code("f6bbfd4b67f32c83")
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, ndim 2, 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 os
import tempfile
import glob
from pandas.core.common import flatten

# Create a temporary directory structure that mimics the expected layout
temp_dir = tempfile.mkdtemp()

# Create mock directory structure
base_path = temp_dir + "/"
data_subdir = "dataset1/"
full_data_path = os.path.join(base_path, data_subdir)

# Create train/val/test subdirectories with sample files
os.makedirs(os.path.join(full_data_path, "train", "class1"), exist_ok=True)
os.makedirs(os.path.join(full_data_path, "train", "class2"), exist_ok=True)
os.makedirs(os.path.join(full_data_path, "val", "class1"), exist_ok=True)
os.makedirs(os.path.join(full_data_path, "val", "class2"), exist_ok=True)
os.makedirs(os.path.join(full_data_path, "test", "class1"), exist_ok=True)
os.makedirs(os.path.join(full_data_path, "test", "class2"), exist_ok=True)

# Create dummy files
open(os.path.join(full_data_path, "train", "class1", "file1.txt"), "w").close()
open(os.path.join(full_data_path, "train", "class2", "file2.txt"), "w").close()
open(os.path.join(full_data_path, "val", "class1", "file3.txt"), "w").close()
open(os.path.join(full_data_path, "val", "class2", "file4.txt"), "w").close()
open(os.path.join(full_data_path, "test", "class1", "file5.txt"), "w").close()
open(os.path.join(full_data_path, "test", "class2", "file6.txt"), "w").close()

ARGS = {
    "base_path": base_path,
    "data_paths": [data_subdir]
}
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 tuple, ndim 2, dtype family float — a tuple 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_model Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_model.py
code served (permissive licence) · get_code("6b419f2ecfa7997a")
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 ClassificationPointNet. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn as nn
import tempfile
import os

torch.manual_seed(0)

# Create a temporary checkpoint file with a valid state dict
temp_dir = tempfile.mkdtemp()
checkpoint_path = os.path.join(temp_dir, "model_checkpoint.pt")

# Create a dummy model and save its state dict
dummy_model = nn.Linear(256, 10)
torch.save(dummy_model.state_dict(), checkpoint_path)

# However, we need to save a ClassificationPointNet state dict instead
# Let's create the correct model and save it
model_to_save = nn.Module()  # Create a minimal state dict compatible with load_model expectations
# Actually, let's just create an empty dict that won't cause issues
torch.save({}, checkpoint_path)

ARGS = {
    'target_device': 'cpu',
    'path_to_checkpoint': checkpoint_path
}
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 ClassificationPointNet — a ClassificationPointNet output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ClassificationPointNet 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
load_model Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/perspective_fields/fields_model.py
code served (permissive licence) · get_code("bde9121fab9d6344")
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 FieldsClassifier. 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 torch.nn as nn
from torchvision.models import resnet50
import tempfile
import os

torch.manual_seed(0)

# Create a temporary checkpoint file
temp_dir = tempfile.gettempdir()
checkpoint_path = os.path.join(temp_dir, "test_checkpoint.pth")

# Create and save a minimal model state dict
model = nn.Module()
model_state = {
    'resnet.conv1.weight': torch.randn(64, 3, 7, 7),
    'resnet.bn1.weight': torch.randn(64),
    'resnet.bn1.bias': torch.randn(64),
    'resnet.bn1.running_mean': torch.randn(64),
    'resnet.bn1.running_var': torch.randn(64),
    'resnet.bn1.num_batches_tracked': torch.tensor(0),
    'resnet.fc.weight': torch.randn(2, 2048),
    'resnet.fc.bias': torch.randn(2),
}

torch.save(model_state, checkpoint_path)

ARGS = {
    'target_device': 'cpu',
    'path_to_checkpoint': checkpoint_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, torchvision — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type FieldsClassifier — a FieldsClassifier output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a FieldsClassifier 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
load_outdoor_problem_paths Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_path_helpers.py
code served (permissive licence) · get_code("c56c6c235242c726")
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 os
import pickle
import tempfile

# Create a temporary directory and a mock pickle file
temp_dir = tempfile.mkdtemp()
mock_problem_paths = [
    '/path/to/problem1.txt',
    '/path/to/problem2.txt',
    '/path/to/problem3.txt'
]

pickle_file_path = os.path.join(temp_dir, 'all_outdoor_problem_paths.pkl')
with open(pickle_file_path, 'wb') as f:
    pickle.dump(mock_problem_paths, f)

ARGS = {
    'base_path': temp_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_recent_timestamp_paths Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/prequalifier/prequalifier_path_helpers.py
code served (permissive licence) · get_code("f372edd1805aaa57")
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 glob
import tempfile
import os

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

# Create mock directory structure for indoor data
indoor_dir = os.path.join(temp_dir, 'indoor_dataset1', 'test', 'subdir1', 'subdir2')
os.makedirs(indoor_dir, exist_ok=True)
# Create a dummy file
open(os.path.join(indoor_dir, 'file1.txt'), 'w').close()

# Create mock directory structure for outdoor data
outdoor_dir = os.path.join(temp_dir, 'outdoor_dataset1', 'test', 'subdir1', 'subdir2')
os.makedirs(outdoor_dir, exist_ok=True)
# Create a dummy file
open(os.path.join(outdoor_dir, 'file2.txt'), 'w').close()

# Set up arguments for the function
ARGS = {
    'base_path': temp_dir,
    'recent_timestamp_data_paths': ['indoor_dataset1', 'outdoor_dataset1'],
    'category': 'indoor'
}
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
remove_problem_paths Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/line_segment/lines_path_helpers.py
code served (permissive licence) · get_code("45dd68d819a5dbf6")
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. 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 os
import pickle
import tempfile

# Create a temporary directory and pickle file for testing
temp_dir = tempfile.mkdtemp()
all_problem_paths = [
    'path/to/outdoor/problem/1.txt',
    'path/to/outdoor/problem/2.txt',
    'path/to/outdoor/problem/3.txt'
]
pkl_path = os.path.join(temp_dir, 'all_outdoor_problem_paths.pkl')
with open(pkl_path, 'wb') as f:
    pickle.dump(all_problem_paths, f)

# Define test arguments
list_of_paths = [
    ['path/to/outdoor/problem/1.txt', 'path/to/problem/4.txt', 'path/to/problem/5.txt'],
    ['path/to/outdoor/problem/2.txt', 'path/to/problem/6.txt'],
    ['path/to/problem/7.txt', 'path/to/outdoor/problem/3.txt']
]

ARGS = {
    'list_of_paths': list_of_paths,
    'base_path': temp_dir,
    'category': 'outdoor'
}
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 — a list 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
timestamp_paths_from_data_paths Ran
ran; our contract draft was wrong, not the code
hanlinm2/projective-geometry/prequalifier/prequalifier_path_helpers.py
code served (permissive licence) · get_code("4ba2962f643eb701")
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 glob
import tempfile
import os
from pathlib import Path

# Create temporary directory structure that mimics what the function expects
temp_dir = tempfile.mkdtemp()
base_path = temp_dir + "/"

# Create mock directory structure
# The function expects: base_path + data_path to be globbable
# Then it globs path + '/test/*/*' 

data_path_1 = "recent1_dataset1"
data_path_2 = "recent2_dataset2"

# Create directories that will be found by glob
dir1 = os.path.join(temp_dir, data_path_1)
dir2 = os.path.join(temp_dir, data_path_2)

os.makedirs(dir1, exist_ok=True)
os.makedirs(dir2, exist_ok=True)

# Create subdirectories for the test/*/* pattern
test_subdir1 = os.path.join(dir1, "test", "subdir1", "file1.txt")
test_subdir2 = os.path.join(dir1, "test", "subdir2", "file2.txt")
test_subdir3 = os.path.join(dir2, "test", "subdir3", "file3.txt")

os.makedirs(os.path.dirname(test_subdir1), exist_ok=True)
os.makedirs(os.path.dirname(test_subdir2), exist_ok=True)
os.makedirs(os.path.dirname(test_subdir3), exist_ok=True)

Path(test_subdir1).touch()
Path(test_subdir2).touch()
Path(test_subdir3).touch()

# Define arguments
ARGS = {
    'base_path': base_path,
    'data_paths': [data_path_1, data_path_2]
}
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
load_model Not yet run
unverified (harvested, no recorded run)
hanlinm2/projective-geometry/prequalifier/prequalifier_model.py
code served (permissive licence) · get_code("f9e6838b930e3168")
How it ran

no invocation recorded.

Also present in other papers' code

Of 18 samples lifted from this paper's linked repositories, 18 carry a function identity and 0 do not. Of the 18 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 18 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

Generative models can produce impressively realistic images. This paper demonstrates that generated images have geometric features different from those of real images. We build a set of collections of generated images, prequalified to fool simple, signal-based classifiers into believing they are real. We then show that prequalified generated images can be identified reliably by classifiers that only look at geometric properties. We use three such classifiers. All three classifiers are denied access to image pixels, and look only at derived geometric features. The first classifier looks at the perspective field of the image, the second looks at lines detected in the image, and the third looks at relations between detected objects and shadows. Our procedure detects generated images more reliably than SOTA local signal based detectors, for images from a number of distinct generators. Saliency maps suggest that the classifiers can identify geometric problems reliably. We conclude that current generators cannot reliably reproduce geometric properties of real images.

For agents

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

get_harvested_code_for_paper("2311.17138")
get_code_for_paper("2311.17138")
have("2311.17138")

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

curl https://syntology.ai/api/ran/2311.17138.json

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

[![Syntology run record](https://syntology.ai/api/ran/2311.17138.svg)](https://syntology.ai/paper/2311.17138)

Connect an agent — have() is free.