SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.04199 · 2026

Which Leakage Types Matter? A Quantitative Landscape Across 2,047 Benchmark Datasets

Simon Roth

arXiv · PDF · Open in the Atlas

Code that ran

9 of 10 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; 1 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 7 failures of our instrument; 1 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/2604.04199.json · badge: /api/ran/2604.04199.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
epagogy/ml canonical (classifier label)
classifier basis: census
9 ran · 1 no recorded run
FunctionStatusWhere it lives
acc Ran
ran, honoured the contract we drafted
epagogy/ml/al/benchmarks/bench_vs_sklearn.py
code served (permissive licence) · get_code("8a2312120b633148")
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 float, ndim 0, dtype family float, 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 numpy as np

# The function acc(p, y) compares predictions p with labels y
# It returns the mean accuracy as a float
# Both p and y should be array-like with comparable shapes

# Create small numpy arrays for predictions and labels
p = np.array([0, 1, 1, 0, 1])  # predictions
y = np.array([0, 1, 0, 0, 1])  # ground truth labels

ARGS = {
    "p": p,
    "y": y,
}
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 float, ndim 0, dtype family float, 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
get_configs Ran
ran; our contract draft was wrong, not the code
epagogy/ml/al/benchmarks/bench_openml.py
code served (permissive licence) · get_code("f0a6a45b398e88f7")
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 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 typing import Tuple, Dict, List

# The function signature shows it takes:
# - algorithm: str

# Looking at the function body, it checks the algorithm parameter
# against specific string values and returns a list of tuples containing
# (config_label: str, hyperparams: dict)

# Valid algorithm names based on the function logic:
# 'svm', 'decision_tree', 'random_forest', 'gradient_boosting', 'linear'
# or any other string (which would just get the default empty dict config)

ARGS = {
    'algorithm': 'decision_tree'
}
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 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
get_grid_for_task Ran
ran; our contract draft was wrong, not the code
epagogy/ml/al/benchmarks/bench_openml.py
code served (permissive licence) · get_code("e9058d9374fde690")
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 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.

# Analysis of get_grid_for_task function:
# - Takes a single parameter 'task' which is a string
# - The function checks if task == 'clf' or task == 'reg'
# - Returns a filtered list based on the task type

# Setup code
ARGS = {
    'task': 'clf'
}
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 2, 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
init_db Ran
ran; our contract draft was wrong, not the code
epagogy/ml/al/benchmarks/bench_landscape.py
code served (permissive licence) · get_code("8d8ea566cb1a5b9b")
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 Connection. 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 sqlite3
import tempfile
import os

# Create a temporary database file path
temp_dir = tempfile.gettempdir()
db_path = os.path.join(temp_dir, "test_results.db")

ARGS = {
    "path": db_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Connection — a Connection output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Connection 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_noisy_xor Ran
ran; our contract draft was wrong, not the code
epagogy/ml/al/benchmarks/bench_cart.py
code served (permissive licence) · get_code("eb31fed70ac561a5")
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 = {
    "n": 100,
    "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 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
make_overlapping Ran
ran; our contract draft was wrong, not the code
epagogy/ml/al/benchmarks/bench_cart.py
code served (permissive licence) · get_code("ef4ecea7ca003697")
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 = {
    "n": 100,
    "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 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
make_wide_sparse Ran
ran; our contract draft was wrong, not the code
epagogy/ml/al/benchmarks/bench_cart.py
code served (permissive licence) · get_code("6d818ec25a1b1c4c")
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 = {
    "n": 100,
    "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 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
r2 Ran
ran, violated the contract we drafted
epagogy/ml/al/benchmarks/bench_vs_sklearn.py
code served (permissive licence) · get_code("8057ced46afec5dc")
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 float, ndim 0, dtype family float, 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 numpy as np

# The r2 function computes R² coefficient of determination
# It takes:
# - p: predicted values (array-like)
# - y: actual/true values (array-like)
# Both should be 1D or compatible arrays of the same shape

# Create small numpy arrays for predictions and true values
np.random.seed(0)
predictions = np.array([1.0, 2.5, 3.1, 4.9, 5.2])
true_values = np.array([1.1, 2.3, 3.0, 5.0, 5.1])

ARGS = {
    "p": predictions,
    "y": true_values,
}
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 float, ndim 0, dtype family float, 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
run_file Ran
ran; our contract draft was wrong, not the code
epagogy/ml/python/safe_test_runner.py
code served (permissive licence) · get_code("c9a3608c0ecaf020")
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 os
import sys

# The function `run_file` takes a single parameter: test_file (a string)
# Looking at the function body:
# - It expects test_file to be a filename that will be joined with TEST_DIR
# - It checks if the file exists using os.path.exists()
# - It passes the filepath to subprocess.run() as part of a pytest command
# - The function doesn't actually need the file to exist for valid argument construction
#   (it handles the 'not found' case gracefully)

# Create a simple test_file argument
test_file = "test_example.py"

ARGS = {
    "test_file": test_file
}
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 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
fetch_suite Not yet run
unverified (harvested, no recorded run)
epagogy/ml/al/benchmarks/bench_openml.py
code served (permissive licence) · get_code("6865a9d7d8471e71")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 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 10 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

Twenty-nine experiments (twenty-eight core, plus a boundary experiment on 129 temporal datasets) across 2,047 iid tabular datasets support a four-class taxonomy of data leakage organized by causal mechanism. Class I (estimation leakage -fitting scalers or encoders on full data) is negligible: nine conditions all produce |ΔAUC| ≤ 0.005. Class II (selection leakage -peeking, seed cherry-picking, early stopping) is substantial at practical dataset sizes (corpus median 𝑛 ≈ 1,900): the measured effect is consistent with an approximately 90% noise-exploitation share (under the assumed zero-diversity decomposition for seeds; see Limitation 11) inflating reported scores. Seed inflation vanishes by 𝑛 = 5,000; peeking retains a residual at 𝑛 = 100,000 that reflects genuine algorithm diversity, not persistent leakage. Class III (memorization) is amplified by model capacity: at 10% duplication, six algorithms span 𝑑 𝑧 = 0.37 (NB) to 1.11 (DT), reaching 𝑑 𝑧 = 1.38 for DT at 30% duplication. Class IV (boundary leakage) is invisible under random CV: a boundary experiment across 129 temporal datasets (92 FOREX as null control, 14 with verified genuine timestamps, 23 with spurious time columns) shows that random CV censors structural contamination. The pure temporal effect averages +0.023 on genuine temporal datasets but near zero on benchmarks without real drift. Feature selection leakage is negligible at typical dimensionality but reaches +0.018 mean at high 𝑝/𝑛 ratios, confirming Ambroise and McLachlan's (2002) finding at scale. Metric selection flips model rankings on 31% of datasets. Cross-validation confidence intervals achieve only 55% actual coverage at nominal 95% under the naive 𝑧-based default (fold SE divided by √ 𝑘, 𝑧 critical value), the convention practitioners report by default; Conservative-Z (fold SD without ÷ √ 𝑘) closes most of the gap to approximately 87%, but is not in standard practice. Within the OpenML-class iid tabular regime studied here, the textbook emphasis is inverted: the leakage type most prominent in standard references (normalization) matters least; selection leakage at practical dataset sizes matters most. Translation to non-tabular (image, text) and other non-iid (group, spatial) domains is an open question (see Limitations).

For agents

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

get_harvested_code_for_paper("2604.04199")
get_code_for_paper("2604.04199")
have("2604.04199")

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

curl https://syntology.ai/api/ran/2604.04199.json

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

[![Syntology run record](https://syntology.ai/api/ran/2604.04199.svg)](https://syntology.ai/paper/2604.04199)

Connect an agent — have() is free.