SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.26248 · 2026

Unified Neural Scaling Laws

Irina Rish, Ethan Caballero, David Krueger, Priyank Jaini

arXiv · PDF

Code that ran

18 of 37 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, 4 violated it and 0 have no contract verdict; 13 failures of our instrument; 19 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2605.26248.json · badge: /api/ran/2605.26248.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
google/flax canonical (classifier label)
classifier basis: census
8 ran · 6 no recorded run
jax-ml/jax canonical (classifier label)
classifier basis: census
7 ran · 10 no recorded run
google-deepmind/kfac-jax canonical (classifier label)
classifier basis: census
3 ran · 3 no recorded run
FunctionStatusWhere it lives
add_abstract Ran
ran; our contract draft was wrong, not the code
jax-ml/jax/jax/_src/ad_util.py
code served (permissive licence) · get_code("2084aa5b3266ab8f")
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 ShapedArray. 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
from jax._src import core
from jax._src.core import Primitive

# Create the primitive
add_jaxvals_p = Primitive('add_any')

@add_jaxvals_p.def_abstract_eval
def add_abstract(x, y):
    assert core.typematch(x, y), (x, y)
    return x

# Construct valid arguments for add_abstract
# The function takes two JAX abstract values that must match in type
# We'll use JAX's abstract values

# Create abstract values - these represent the shape/dtype info of JAX arrays
x_abstract = core.ShapedArray(shape=(2, 4), dtype=jnp.float32)
y_abstract = core.ShapedArray(shape=(2, 4), dtype=jnp.float32)

ARGS = {
    'x': x_abstract,
    'y': y_abstract,
}
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 ShapedArray — a ShapedArray output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ShapedArray 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
blocked_fold_in Ran
ran; our contract draft was wrong, not the code
jax-ml/jax/jax/_src/blocked_sampler.py
code served (permissive licence) · get_code("cac8281b61c863de")
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 3, 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 jax
import jax.numpy as jnp

# Set seed for reproducibility
jax.config.update("jax_default_prng_impl", "threefry2x32")

# Create a global PRNGKey
global_key = jax.random.PRNGKey(0)

# Define the arguments for blocked_fold_in
# total_size: shape of the array being generated (16x512)
total_size = (16, 512)

# block_size: shape of an individual block (16x128)
block_size = (16, 128)

# tile_size: smallest unit at which samples are generated (8x128)
tile_size = (8, 128)

# block_index: which block to generate keys for - 2D indices
# Using jnp arrays as ArrayLike objects
block_index = (jnp.array(0), jnp.array(0))

ARGS = {
    "global_key": global_key,
    "total_size": total_size,
    "block_size": block_size,
    "tile_size": tile_size,
    "block_index": block_index,
}
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 list, ndim 3, 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
broadcast_in_dim_equivalent Ran
ran, violated the contract we drafted
google-deepmind/kfac-jax/kfac_jax/_src/tag_graph_matcher.py
code served (permissive licence) · get_code("987e113518340066")
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 bool, 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 jax
import jax.numpy as jnp
import jax.extend as jex
from jax import core

# Create minimal JaxprEqn objects for broadcast_in_dim primitive
def create_broadcast_in_dim_eqn():
    """Create a minimal JaxprEqn for broadcast_in_dim primitive."""
    # Get the broadcast_in_dim primitive
    broadcast_in_dim_prim = core.Primitive.new_name('broadcast_in_dim')
    
    # Create input and output abstractions
    input_shape = (2, 3)
    output_shape = (2, 3, 4)
    
    # Create abstract values
    in_aval = core.ShapedArray(input_shape, jnp.float32)
    out_aval = core.ShapedArray(output_shape, jnp.float32)
    
    # Create variables for the equation
    in_var = core.Var(in_aval)
    out_var = core.Var(out_aval)
    
    # Create JaxprEqn with broadcast_in_dim primitive
    eqn = core.JaxprEqn(
        primitive=core.get_aval(jnp.broadcast_to(jnp.zeros(input_shape), output_shape)).__class__.__bases__[0],
        invars=[in_var],
        outvars=[out_var],
        params={}
    )
    
    return eqn

# Simpler approach: directly access existing primitives and create equations
def create_simple_jaxpr_eqn():
    """Create a JaxprEqn using actual JAX primitives."""
    # Use the actual broadcast_in_dim primitive from JAX
    from jax.interpreters import xla
    from jax.lib import xla_client
    
    # Create abstract input/output values
    in_aval = core.ShapedArray((2, 3), jnp.float32)
    out_aval = core.ShapedArray((2, 3, 4), jnp.float32)
    
    in_var = core.Var(in_aval)
    out_var = core.Var(out_aval)
    
    # Get the actual broadcast_in_dim primitive
    from jax.experimental.key_reuse import _check_key_reuse_count
    from jax._src.dtypes import preinit
    from jax.experimental import io_callback
    
    # Access primitive directly through core module
    prim = core.Primitive.new_name('broadcast_in_dim')
    prim.multiple_results = False
    
    eqn = core.JaxprEqn(
        primitive=prim,
        invars=[in_var],
        outvars=[out_var],
        params={'shape': (2, 3, 4), 'broadcast_dimensions': (0, 1)}
    )
    
    return eqn

# Most direct approach: construct minimal mock objects
class MockPrimitive:
    def __init__(self, name):
        self.name = name

class MockJaxprEqn:
    def __init__(self, prim_name):
        self.primitive = MockPrimitive(prim_name)

# Create the arguments
equation1 = MockJaxprEqn('broadcast_in_dim')
equation2 = MockJaxprEqn('broadcast_in_dim')

ARGS = {
    'equation1': equation1,
    'equation2': equation2
}
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 bool, 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
buffer_callback Ran
ran; our contract draft was wrong, not the code
jax-ml/jax/jax/_src/buffer_callback.py
code served (permissive licence) · get_code("29226c63dfc413f0")
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 numpy as np
import jax
import jax.numpy as jnp
from jax._src import core
from jax._src import tree_util


def simple_callback(ctx, out, x):
    """A simple callback that adds 1 to the input."""
    np.asarray(out)[...] = np.asarray(x) + 1


# Create a small input array
x = jnp.array([1, 2, 3, 4], dtype=jnp.int32)

# Define the output shape and dtype
result_shape_dtypes = jax.ShapeDtypeStruct(shape=(4,), dtype=jnp.int32)

ARGS = {
    'callback': simple_callback,
    'result_shape_dtypes': result_shape_dtypes,
    'has_side_effect': False,
    'vmap_method': None,
    'input_output_aliases': None,
    'command_buffer_compatible': False,
}
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
build_shaped_array Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
google/flax/flax/core/axes_scan.py
code served (permissive licence) · get_code("3a55efb472f943e3")
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 ShapedArray. 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
from jax import core

# Create a simple JAX array to pass as input
x = jnp.array([[1.0, 2.0, 3.0, 4.0],
               [5.0, 6.0, 7.0, 8.0]])

# Define ARGS dictionary with all parameters
ARGS = {
    'x': x,
    'batch_dim': False
}
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 ShapedArray — a ShapedArray output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ShapedArray 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
compatible_sum Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
google-deepmind/kfac-jax/kfac_jax/_src/curvature_blocks/utils.py
code served (permissive licence) · get_code("dc1b986b1be917d4")
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 4, dtype family float, shape (2, 1, 3, 5) 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 test tensors and parameters for compatible_sum function
tensor = jnp.ones((2, 4, 3, 5))  # 4D tensor with distinct dimensions
target_shape = (1, 3, 5)  # Target shape compatible with tensor
skip_axes = [0]  # Axes to skip during summation

ARGS = {
    "tensor": tensor,
    "target_shape": target_shape,
    "skip_axes": skip_axes
}
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 4, dtype family float, shape (2, 1, 3, 5) — shape (2, 1, 3, 5) 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
eval_jaxpr_eqn Ran
ran; our contract draft was wrong, not the code
google-deepmind/kfac-jax/kfac_jax/_src/tag_graph_matcher.py
code served (permissive licence) · get_code("42474a82097165d1")
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 jax
import jax.numpy as jnp
from jax import core
from jax.extend import core as jex_core
import jax.extend as jex

# Create a minimal Jaxpr equation for testing
# We'll use a simple primitive operation like addition

# Get the add primitive
add_prim = jax.lax.add_p

# Create minimal source info
class MockSourceInfo:
    def __init__(self):
        self.traceback = None

# Create a minimal JaxprEqn-like object
class MockJaxprEqn:
    def __init__(self):
        self.primitive = add_prim
        self.params = {}
        self.source_info = MockSourceInfo()
        self.invars = []
        self.outvars = []

# Construct the equation
eqn = MockJaxprEqn()

# Create input values - simple JAX arrays for addition
in_values = [
    jnp.array([1.0, 2.0, 3.0, 4.0]),
    jnp.array([5.0, 6.0, 7.0, 8.0])
]

ARGS = {
    'eqn': eqn,
    'in_values': in_values
}
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 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
identity Ran
ran, honoured the contract we drafted
jax-ml/jax/jax/_src/ad_checkpoint.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from dholzmueller/universal_double_descent, is pointer only (licence: NONE) · get_code("9910e2fc297f8665")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from dholzmueller/universal_double_descent, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

is_axis_metadata Ran
ran, violated the contract we drafted
google/flax/flax/core/meta.py
code served (permissive licence) · get_code("a1fcbca5bacade77")
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 bool, 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 abc
from typing import Any, Generic, TypeVar

A = TypeVar('A')
B = TypeVar('B')
TAxisMetadata = TypeVar('TAxisMetadata', bound='AxisMetadata[Any]')

class AxisMetadata(Generic[A], metaclass=abc.ABCMeta):
    """Abstract base class for boxed Metadata."""

    @abc.abstractmethod
    def unbox(self) -> A:
        pass

    @abc.abstractmethod
    def replace_boxed(self, val: B) -> 'AxisMetadata[B]':
        pass

    @abc.abstractmethod
    def add_axis(self: TAxisMetadata, index: int, params: dict[Any, Any]) -> TAxisMetadata:
        pass

    @abc.abstractmethod
    def remove_axis(self: TAxisMetadata, index: int, params: dict[Any, Any]) -> TAxisMetadata:
        pass

class ConcreteAxisMetadata(AxisMetadata):
    """Concrete implementation of AxisMetadata for testing."""
    
    def __init__(self, value: Any):
        self._value = value
    
    def unbox(self) -> Any:
        return self._value
    
    def replace_boxed(self, val: Any) -> 'ConcreteAxisMetadata':
        return ConcreteAxisMetadata(val)
    
    def add_axis(self, index: int, params: dict[Any, Any]) -> 'ConcreteAxisMetadata':
        return ConcreteAxisMetadata(self._value)
    
    def remove_axis(self, index: int, params: dict[Any, Any]) -> 'ConcreteAxisMetadata':
        return ConcreteAxisMetadata(self._value)

def is_axis_metadata(val: Any) -> bool:
    """Returns whether the argument is an instance of AxisMetadata."""
    return isinstance(val, AxisMetadata)

# Create test instances
metadata_instance = ConcreteAxisMetadata(value=[1, 2, 3])
non_metadata_instance = "not metadata"

ARGS = {
    'val': metadata_instance
}
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 bool, 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
is_named_tuple Ran
ran, violated the contract we drafted
google/flax/flax/cursor.py
code served (permissive licence) · get_code("77bb0fc60c449e7d")
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 bool, 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.

from collections import namedtuple

# Create a valid named tuple type
TestNamedTuple = namedtuple('TestNamedTuple', ['field1', 'field2', 'field3'])

# Create an instance of the named tuple
obj = TestNamedTuple(field1=1, field2='test', field3=[1, 2, 3])

ARGS = {
    'obj': obj
}
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 bool, 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
lazy_init Ran
ran; our contract draft was wrong, not the code
google/flax/flax/core/partial_eval.py
code served (permissive licence) · get_code("3f76f2c44d58d7ce")
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 functools

def simple_function(x, y):
    """A simple function to be wrapped with lazy_init."""
    return x + y

def another_function(a, b, c=None):
    """Another function with keyword arguments."""
    if c is None:
        return a * b
    return a * b + c

# Test case 1: Simple function with concrete arrays
concrete_x = jnp.array([1.0, 2.0, 3.0])
concrete_y = jnp.array([4.0, 5.0, 6.0])

# Test case 2: Function with ShapeDtypeStruct
sds_x = jax.ShapeDtypeStruct(shape=(2, 4), dtype=jnp.float32)
sds_y = jax.ShapeDtypeStruct(shape=(2, 4), dtype=jnp.float32)

# Test case 3: Function with mixed concrete and ShapeDtypeStruct
mixed_concrete = jnp.array([1.0, 2.0])
mixed_sds = jax.ShapeDtypeStruct(shape=(3,), dtype=jnp.float32)

# Test case 4: Function with keyword arguments
sds_a = jax.ShapeDtypeStruct(shape=(5, 3), dtype=jnp.int32)
sds_b = jax.ShapeDtypeStruct(shape=(5, 3), dtype=jnp.int32)
sds_c = jax.ShapeDtypeStruct(shape=(5, 3), dtype=jnp.float32)

ARGS = {
    "fn": simple_function
}
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
map_axis_meta Ran
ran; our contract draft was wrong, not the code
google/flax/flax/core/meta.py
code served (permissive licence) · get_code("ebaf2e0326f0ee3d")
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 jax
import jax.numpy as jnp
from typing import Any

class ConcreteAxisMetadata:
    """Concrete implementation of AxisMetadata for testing."""
    
    def __init__(self, value: Any):
        self._value = value
        self._axes = []
    
    def unbox(self) -> Any:
        return self._value
    
    def replace_boxed(self, val: Any) -> 'ConcreteAxisMetadata':
        return ConcreteAxisMetadata(val)
    
    def add_axis(self, index: int, params: dict[Any, Any]) -> 'ConcreteAxisMetadata':
        new_meta = ConcreteAxisMetadata(self._value)
        new_meta._axes = self._axes.copy()
        new_meta._axes.insert(index, ('add', params))
        return new_meta
    
    def remove_axis(self, index: int, params: dict[Any, Any]) -> 'ConcreteAxisMetadata':
        new_meta = ConcreteAxisMetadata(self._value)
        new_meta._axes = self._axes.copy()
        new_meta._axes.pop(index)
        return new_meta

def test_function(fn: Any, tree: Any) -> Any:
    """Test function that increments each AxisMetadata value by 1."""
    def increment_meta(meta):
        val = meta.unbox()
        if isinstance(val, (int, float)):
            return meta.replace_boxed(val + 1)
        return meta
    return increment_meta(fn) if isinstance(fn, ConcreteAxisMetadata) else fn

# Create test data
metadata_obj = ConcreteAxisMetadata(42)
nested_tree = {
    'a': ConcreteAxisMetadata(10),
    'b': {
        'c': ConcreteAxisMetadata(20),
        'd': jnp.array([1, 2, 3, 4])
    },
    'e': [ConcreteAxisMetadata(5), 'string', 3.14]
}

def example_fn(meta):
    """Example function to map over AxisMetadata instances."""
    if isinstance(meta, ConcreteAxisMetadata):
        val = meta.unbox()
        if isinstance(val, int):
            return val * 2
    return None

ARGS = {
    'fn': example_fn,
    'tree': nested_tree
}
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 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
offload_dot_with_no_batch_dims Ran
ran; our contract draft was wrong, not the code
jax-ml/jax/jax/_src/ad_checkpoint.py
code served (permissive licence) · get_code("810b7a396afb59d3")
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 OffloadDotWithNoBatchDims. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from jax._src.interpreters import partial_eval as pe
from jax._src.lax import lax as lax_internal

torch.manual_seed(0)

# The function offload_dot_with_no_batch_dims takes two string parameters:
# offload_src and offload_dst. These are used to construct an OffloadDotWithNoBatchDims
# instance, which stores them as attributes and uses them when creating a 
# pe.Offloadable object.

ARGS = {
    "offload_src": "cpu",
    "offload_dst": "gpu"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type OffloadDotWithNoBatchDims — a OffloadDotWithNoBatchDims output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a OffloadDotWithNoBatchDims 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
sample_block Ran
ran; our contract draft was wrong, not the code
jax-ml/jax/jax/_src/blocked_sampler.py
code served (permissive licence) · get_code("90d492383f675799")
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 (4, 6) 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 jax.random as jrandom

# Define a simple sampler function that matches the SampleFn protocol
def mock_sampler_fn(key, shape=None):
    """A mock sampling function that returns random uniform samples."""
    return jrandom.uniform(key, shape=shape)

# Create a random key for the keys parameter
# The keys should be a nested structure/grid of JAX random keys
# For a 2D block, we need a 2D grid of keys
seed = 42
master_key = jrandom.PRNGKey(seed)

# Generate a grid of keys matching the size_in_tiles structure
# If block_size=(4, 6) and tile_size=(2, 3), size_in_tiles=(2, 2)
# So we need a 2x2 grid of keys
keys_2d = []
for i in range(2):
    row = []
    for j in range(2):
        master_key, subkey = jrandom.split(master_key)
        row.append(subkey)
    keys_2d.append(row)

ARGS = {
    "sampler_fn": mock_sampler_fn,
    "keys": keys_2d,
    "block_size": (4, 6),
    "tile_size": (2, 3),
}
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 (4, 6) — shape (4, 6) 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
save_from_both_policies Ran
ran; our contract draft was wrong, not the code
jax-ml/jax/jax/_src/ad_checkpoint.py
code served (permissive licence) · get_code("2e107afaca8a349d")
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 torch
import torch.nn as nn

torch.manual_seed(0)

def policy_1(prim, *args, **params):
    """Simple policy that returns True for even primitives."""
    return True

def policy_2(prim, *args, **params):
    """Simple policy that returns False for odd primitives."""
    return False

ARGS = {
    "policy_1": policy_1,
    "policy_2": policy_2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
scan Ran
ran; our contract draft was wrong, not the code
google/flax/flax/core/axes_scan.py
code served (permissive licence) · get_code("c0375204f39d6824")
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
from collections.abc import Callable

# Define the broadcast sentinel from the module
class _Broadcast:
    pass

broadcast = _Broadcast()

# Define a simple body function that matches the expected signature
def body_fn(broadcast_in, carry, x):
    """
    Simple body function matching signature:
    (broadcast_in, carry, *args) -> (broadcast_out, carry, scan_out)
    """
    broadcast_out = broadcast_in + 1
    new_carry = carry + x
    scan_out = x * 2
    return broadcast_out, new_carry, scan_out

# Create simple test data
# xs should be a pytree with scan axes specified
xs = jnp.arange(12).reshape(3, 4)  # shape (3, 4) - batch of 3, feature dim 4

# in_axes specifies which arguments get scanned
# broadcast for broadcast_in (doesn't scan)
# broadcast for carry (doesn't scan)  
# 0 for xs (scan along axis 0)
in_axes = (broadcast, broadcast, 0)

# out_axes specifies which outputs get stacked
# broadcast for broadcast_out (doesn't stack)
# broadcast for carry (doesn't stack)
# 0 for scan_out (stack along axis 0)
out_axes = (broadcast, broadcast, 0)

# length can be derived from xs, but we can specify it explicitly
length = 3

ARGS = {
    'fn': body_fn,
    'in_axes': in_axes,
    'out_axes': out_axes,
    'length': length,
    'reverse': False,
    'unroll': 1,
    '_split_transpose': False,
    'check_constancy_invariants': True,
}
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
static_bool_env Ran
ran, violated the contract we drafted
google/flax/flax/configurations.py
code served (permissive licence) · get_code("59ef236289eabe8e")
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 bool, 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 os

# Set up environment variables for testing
os.environ['TEST_VAR_TRUE'] = 'yes'
os.environ['TEST_VAR_FALSE'] = 'no'
os.environ['TEST_VAR_UNSET'] = ''

# Test case 1: varname that exists with a true value
ARGS = {
    'varname': 'TEST_VAR_TRUE',
    'default': False
}
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 bool, 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
unfreeze Ran
ran; our contract draft was wrong, not the code
google/flax/flax/core/frozen_dict.py
code served (permissive licence) · get_code("b4dc618ebcbc9bb7")
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 jax
import jax.tree_util

# The unfreeze function takes a single parameter 'x' which should be either
# a FrozenDict or a dict[str, Any]

# Test case 1: FrozenDict input
frozen_dict_input = {
    'params': {
        'layer1': {
            'weight': 0.5,
            'bias': 0.1
        },
        'layer2': {
            'weight': 0.3
        }
    },
    'metadata': {
        'version': 1,
        'name': 'test'
    }
}

class FrozenKeysView:
    def __init__(self, d):
        self._dict = d
    def __repr__(self):
        return f'frozen_dict_keys({list(self._dict.keys())})'

class FrozenValuesView:
    def __init__(self, d):
        self._dict = d
    def __repr__(self):
        return f'frozen_dict_values({list(self._dict.values())})'

def _prepare_freeze(xs):
    if isinstance(xs, dict):
        return {key: _prepare_freeze(val) for key, val in xs.items()}
    return xs

class FrozenDict(dict):
    def __init__(self, *args, __unsafe_skip_copy__=False, **kwargs):
        xs = dict(*args, **kwargs)
        if __unsafe_skip_copy__:
            self._dict = xs
        else:
            self._dict = _prepare_freeze(xs)
        self._hash = None
        super().__init__(self._dict)
    
    def unfreeze(self):
        return dict(self._dict)

# Create a FrozenDict instance
frozen_dict = FrozenDict(frozen_dict_input)

# Test case 2: regular dict input
regular_dict = {
    'config': {
        'learning_rate': 0.001,
        'batch_size': 32
    },
    'state': {
        'step': 100
    }
}

ARGS = {
    'x': frozen_dict
}
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 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
add_axis Not yet run
unverified (harvested, no recorded run)
google/flax/flax/core/meta.py
code served (permissive licence) · get_code("218f103466769b26")
How it ran

no invocation recorded.

add_impl Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/ad_util.py
code served (permissive licence) · get_code("7dc8234c8244b402")
How it ran

no invocation recorded.

add_jaxvals Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/ad_util.py
code served (permissive licence) · get_code("c4008ce797ce88b7")
How it ran

no invocation recorded.

argnums_partial Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/api_util.py
code served (permissive licence) · get_code("0e74585505a2eea7")
How it ran

no invocation recorded.

copy Not yet run
unverified (harvested, no recorded run)
google/flax/flax/core/frozen_dict.py
code served (permissive licence) · get_code("1ff52699dd9f95b3")
How it ran

no invocation recorded.

flatten_fun Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/api_util.py
code served (permissive licence) · get_code("eb71d064139c0c1e")
How it ran

no invocation recorded.

flatten_fun_nokwargs Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/api_util.py
code served (permissive licence) · get_code("3ccc19e76d1ee880")
How it ran

no invocation recorded.

freeze Not yet run
unverified (harvested, no recorded run)
google/flax/flax/core/frozen_dict.py
code served (permissive licence) · get_code("cda663e4a9e350d5")
How it ran

no invocation recorded.

get Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/cache_key.py
code served (permissive licence) · get_code("9fc0e9452ec7806b")
How it ran

no invocation recorded.

get_and_verify_loss_meta Not yet run
unverified (harvested, no recorded run)
google-deepmind/kfac-jax/kfac_jax/_src/layers_and_loss_tags.py
code served (permissive licence) · get_code("14e10e479569a361")
How it ran

no invocation recorded.

get_loss_outputs Not yet run
unverified (harvested, no recorded run)
google-deepmind/kfac-jax/kfac_jax/_src/layers_and_loss_tags.py
code served (permissive licence) · get_code("b4a5e9703e6605e1")
How it ran

no invocation recorded.

get_mesh Not yet run
unverified (harvested, no recorded run)
google/flax/flax/core/spmd.py
code served (permissive licence) · get_code("0d00372c02321358")
How it ran

no invocation recorded.

io_callback_batching_rule Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/callback.py
code served (permissive licence) · get_code("a3183105bcb38a23")
How it ran

no invocation recorded.

jit Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/api.py
code served (permissive licence) · get_code("1dc34cecff86ce05")
How it ran

no invocation recorded.

map_sharding Not yet run
unverified (harvested, no recorded run)
google/flax/flax/core/spmd.py
code served (permissive licence) · get_code("10ee7620c9ca5e4b")
How it ran

no invocation recorded.

pure_callback Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/callback.py
code served (permissive licence) · get_code("a860c143f6dbe657")
How it ran

no invocation recorded.

pure_callback_lowering Not yet run
unverified (harvested, no recorded run)
jax-ml/jax/jax/_src/callback.py
code served (permissive licence) · get_code("55ac98a7c46161c0")
How it ran

no invocation recorded.

reshape_equivalent Not yet run
unverified (harvested, no recorded run)
google-deepmind/kfac-jax/kfac_jax/_src/tag_graph_matcher.py
code served (permissive licence) · get_code("ee339b16cfba7ff0")
How it ran

no invocation recorded.

tree_map_rngs Not yet run
unverified (harvested, no recorded run)
google/flax/flax/core/lift.py
code served (permissive licence) · get_code("cb753af998425f9a")
How it ran

no invocation recorded.

Also present in other papers' code

Of 37 samples lifted from this paper's linked repositories, 37 carry a function identity and 0 do not. Of the 37 distinct functions with an identity, 23 are also present in other papers whose code we hold (23 with at least one paper dated earlier, 1 the same month, 1 later) and 14 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

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

FunctionAlso present inEarliest-datedOthers named
identity 69 other papers
67 dated earlier, 1 the same month, 1 later · 5 only as an assembly brick
1706.07296 (2017-06)
A Minimal Developmental Model Can Increase Evolvability in Soft Robots
1802.09419 (2018-02)
Stochastic Hyperparameter Optimization through Hypernetworks
1804.06055 (2018-04)
Co-occurrence Feature Learning from Skeleton Data for Action Recognition and Detection with Hierarchical Aggregation
and 66 more not named here
broadcast_in_dim_equivalent 2 other papers
2 dated earlier
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2203.08120 (2022-03)
Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers
compatible_sum 2 other papers
2 dated earlier
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2203.08120 (2022-03)
Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers
eval_jaxpr_eqn 2 other papers
2 dated earlier
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2203.08120 (2022-03)
Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers
get_and_verify_loss_meta 2 other papers
2 dated earlier
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2203.08120 (2022-03)
Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers
get_loss_outputs 2 other papers
2 dated earlier
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2203.08120 (2022-03)
Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers
reshape_equivalent 2 other papers
2 dated earlier
1503.05671 (2015-03)
Optimizing Neural Networks with Kronecker-factored Approximate Curvature
2203.08120 (2022-03)
Deep Learning without Shortcuts: Shaping the Kernel with Tailored Rectifiers
add_abstract 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
add_impl 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
add_jaxvals 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
argnums_partial 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
blocked_fold_in 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
buffer_callback 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
flatten_fun 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
flatten_fun_nokwargs 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
get 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
io_callback_batching_rule 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
jit 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
offload_dot_with_no_batch_dims 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals
pure_callback 1 other paper
1 dated earlier
2411.03604 (2024-11)
Temporal-Difference Learning Using Distributed Error Signals

Shown: the 20 functions present in the most other papers. 3 more shared functions are not listed here.

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

Repositories linked to this paper

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

Abstract

We present a functional form (that we refer to as a Unified Neural Scaling Law (UNSL)) that accurately models and extrapolates the scaling behaviors of deep neural networks as multiple dimensions all vary simultaneously (i.e. how the evaluation metric of interest varies as one simultaneously varies the number of model parameters, training dataset size, number of training steps, number of inference steps, and various hyperparameters) for various architectures and for each of various tasks within a varied set of upstream and downstream tasks. When compared to other functional forms for neural scaling, this functional form yields extrapolations of scaling behavior that are considerably more accurate on this set.

For agents

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

get_harvested_code_for_paper("2605.26248")
get_code_for_paper("2605.26248")
have("2605.26248")

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

curl https://syntology.ai/api/ran/2605.26248.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.26248.svg)](https://syntology.ai/paper/2605.26248)

Connect an agent — have() is free.