SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.21716 · 2026

A Defense of the Quadratic Model

Cengiz Pehlevan, Sham Kakade, Depen Morwani, Alex Damian, Alexandru Meterez, Pranav Nair

arXiv · PDF · Open in the Atlas

Code that ran

14 of 14 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 5 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 8 failures of our instrument; 0 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/2607.21716.json · badge: /api/ran/2607.21716.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
alexandrumeterez/QuadraticModel canonical (classifier label)
classifier basis: census
14 ran · 0 no recorded run
FunctionStatusWhere it lives
bool_column Ran
ran, violated the contract we drafted
alexandrumeterez/QuadraticModel/analysis/plot_training_figures.py
code served (permissive licence) · get_code("2639833fe54827fc")
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 Series. 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 pandas as pd
import numpy as np

# Create a pandas Series with mixed string representations of booleans
series_mixed = pd.Series(['true', 'false', 'True', 'False', '1', '0', 'yes', 'no'])

ARGS = {
    'series': series_mixed
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Series — a Series output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Series output the recorded ndim 1 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
combined_positive Ran
ran, honoured the contract we drafted
alexandrumeterez/QuadraticModel/analysis/plot_spectrum_summary.py
code served (permissive licence) · get_code("2fc663a12b313863")
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 numpy as np

# Construct a valid curve dictionary
curve = {
    'L': 3,  # Length of the first part
    'x': np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0]),
    'y': np.array([0.5, 1.5, 2.5, 3.5, 4.5, 5.5]),
    'g': np.array([0.1, 0.2, np.nan, 0.4, 0.5, -0.1]),
    'mid': np.array([10.0, 20.0, 30.0, np.nan, 50.0, 60.0])
}

ARGS = {
    'curve': curve
}
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, 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
curve Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/analysis/plot_paper_panels.py
code served (permissive licence) · get_code("e22c4ee58ab33073")
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 torch.nn as nn

torch.manual_seed(0)

class MockCache:
    def curve(self, batch, pct, curvature, preconditioner):
        return {
            'loss': torch.tensor(0.5),
            'accuracy': torch.tensor(0.9)
        }

class MockMidpoints:
    def __getitem__(self, key):
        return torch.tensor([0.1, 0.2, 0.3, 0.4])

cache = MockCache()
batch = 1
pct = 50
curvature = "fisher"
preconditioner = "identity"
midpoints = MockMidpoints()

ARGS = {
    'cache': cache,
    'batch': batch,
    'pct': pct,
    'curvature': curvature,
    'preconditioner': preconditioner,
    'midpoints': midpoints
}
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 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
displayed_bsz_mults Ran
ran, honoured the contract we drafted
alexandrumeterez/QuadraticModel/analysis/plot_training_figures.py
code served (permissive licence) · get_code("037dff4dbddd7a0e")
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 numpy as np

ARGS = {
    "base": 4
}
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 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
get_max_flops_sec Ran
ran, honoured the contract we drafted
alexandrumeterez/QuadraticModel/util.py
code served (permissive licence) · get_code("dcf154f8d497bc60")
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 jax
import jax.numpy as jnp
import numpy as np

# Mock the jax.devices() call since we're on CPU
# We need to provide arguments that work with the function's logic

ARGS = {
    'dtype': jnp.bfloat16
}
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 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
legend Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/analysis/plot_source_conditions.py
code served (permissive licence) · get_code("983834bb82d5b6b2")
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.

from matplotlib.lines import Line2D
from matplotlib.patches import Patch

ARGS = {
    'fit': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib — 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
lerp Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/opt.py
code served (permissive licence) · get_code("95733eb983e3b0e3")
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

# The lerp function performs linear interpolation between two pytrees
# It uses jax.tree.map to apply the lerp formula element-wise
# Parameters a and b should be pytrees (can be nested structures of arrays)
# Parameter s should be a scalar

# Create simple scalar values as pytrees
a = {
    'x': jnp.array([1.0, 2.0, 3.0, 4.0]),
    'y': jnp.array([[1.0, 2.0], [3.0, 4.0]])
}

b = {
    'x': jnp.array([5.0, 6.0, 7.0, 8.0]),
    'y': jnp.array([[5.0, 6.0], [7.0, 8.0]])
}

s = 0.5

ARGS = {
    'a': a,
    'b': b,
    's': s
}
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
linearization_series Ran
ran, honoured the contract we drafted
alexandrumeterez/QuadraticModel/analysis/plot_training_figures.py
code served (permissive licence) · get_code("00c8304d153c4195")
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 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 pandas as pd
import numpy as np

# Create a minimal DataFrame that satisfies the function's requirements
data = pd.DataFrame({
    'schedule': ['linear', 'linear', 'linear', 'linear', 'cosine', 'cosine'],
    'base_batch_size': [32, 32, 32, 32, 64, 64],
    'checkpoint_run_id': [1, 1, 2, 2, 1, 1],
    'checkpoint_step': [100, 100, 200, 200, 100, 100],
    'eval_index': [0, 1, 0, 1, 0, 1],
    'tokens_raw': [1000.0, 2000.0, 1500.0, 2500.0, 3000.0, 4000.0],
    'L0_raw': [5.0, 4.5, 5.2, 4.8, 5.1, 4.7],
    'L_minus_L0': [0.5, 0.3, 0.4, 0.2, 0.6, 0.4],
    'tokens_prox': [1100.0, 2100.0, 1600.0, 2600.0, 3100.0, 4100.0],
    'L0_prox': [5.1, 4.6, 5.3, 4.9, 5.2, 4.8],
    'L_prox_minus_L0': [0.55, 0.35, 0.45, 0.25, 0.65, 0.45],
    'tokens_quad': [1050.0, 2050.0, 1550.0, 2550.0, 3050.0, 4050.0],
    'L0_quad': [5.05, 4.55, 5.25, 4.85, 5.15, 4.75],
    'L_quad_minus_L0': [0.52, 0.32, 0.42, 0.22, 0.62, 0.42],
})

ARGS = {
    'data': data,
    'schedule': 'linear',
    'batch': 32,
    'transform': 'raw'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — 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 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 Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/analysis/plot_sample_size_comparison.py
code served (permissive licence) · get_code("be89a8d6d1fe6ca0")
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 tempfile
from pathlib import Path
import numpy as np

# Create a temporary directory to serve as cache_dir
temp_dir = tempfile.TemporaryDirectory()
cache_dir = Path(temp_dir.name)

# Create a valid cache file matching the expected pattern
# The function looks for files matching: gn_adam_ntok{n_tokens}_r0_*_{slot}_*.npz
# For batch=64, n_tokens=10000000, slot=4
batch = 64
n_tokens = 10000000
slot = 4

# Create sample cache data with required keys
cache_data = {
    'L': np.array(5),
    'R': np.array(10),
    'cut': np.array(3),
    'pos': np.array(2),
    'some_array': np.array([1.0, 2.0, 3.0])
}

# Save the cache file with the expected naming convention
cache_filename = f'gn_adam_ntok{n_tokens}_r0_test_{slot}_extra.npz'
cache_path = cache_dir / cache_filename
np.savez(cache_path, **cache_data)

ARGS = {
    'cache_dir': cache_dir,
    'batch': batch,
    'n_tokens': n_tokens
}
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 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_curves Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/analysis/plot_source_conditions.py
code served (permissive licence) · get_code("1dd8832729f8b64d")
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.

from pathlib import Path
import numpy as np
import tempfile
import os

# Create a temporary directory and a valid npz file
temp_dir = tempfile.mkdtemp()
temp_path = Path(temp_dir) / "test_curves.npz"

# Generate sample data matching the expected structure
# The function expects arrays for each combination of batch, pct, kind, and field
data_dict = {}

BATCHES = (1, 64, 1024)
CHECKPOINTS = (10, 50, 100)
CURVES = {'gn': ('GN', '#1f5f8b', '#6da4d8'), 'hessian': ('Hessian', '#a4550d', '#f4a261')}

schedule = "test_schedule"

# Create dummy arrays for all expected keys
for batch in BATCHES:
    for pct in CHECKPOINTS:
        for kind in CURVES.keys():
            # Create small arrays for x, lo, hi, mid fields
            data_dict[f'{schedule}_B{batch}_P{pct}_{kind}_x'] = np.array([1.0, 2.0, 3.0], dtype=np.float32)
            data_dict[f'{schedule}_B{batch}_P{pct}_{kind}_lo'] = np.array([0.5, 1.5, 2.5], dtype=np.float32)
            data_dict[f'{schedule}_B{batch}_P{pct}_{kind}_hi'] = np.array([1.5, 2.5, 3.5], dtype=np.float32)
            data_dict[f'{schedule}_B{batch}_P{pct}_{kind}_mid'] = np.array([1.0, 2.0, 3.0], dtype=np.float32)

# Save to npz file
np.savez(temp_path, **data_dict)

ARGS = {
    'path': temp_path,
    'schedule': schedule,
}
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 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_legacy Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/analysis/plot_source_conditions.py
code served (permissive licence) · get_code("f0b5b1f90652a9a7")
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 tempfile
from pathlib import Path
import numpy as np

# Create a temporary npz file with the expected structure
temp_dir = tempfile.mkdtemp()
temp_path = Path(temp_dir) / "test_data.npz"

# Create sample data matching the expected keys
data_dict = {}
BATCHES = (1, 64, 1024)
CURVES = {'gn': ('GN', '#1f5f8b', '#6da4d8'), 'hessian': ('Hessian', '#a4550d', '#f4a261')}

for batch in BATCHES:
    for kind in CURVES:
        for field in ('x', 'lo', 'hi', 'mid'):
            key = f'B{batch}_{kind}_{field}'
            # Create small sample arrays
            data_dict[key] = np.random.randn(10, 5).astype(np.float32)

# Save the npz file
np.savez(temp_path, **data_dict)

ARGS = {
    'path': temp_path,
}
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 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
positive_scale Ran
ran, honoured the contract we drafted
alexandrumeterez/QuadraticModel/analysis/plot_spectrum_summary.py
code served (permissive licence) · get_code("d6520e0f242ae372")
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

# Create a valid curve dict with 'g' and 'mid' keys
# The function expects arrays where:
# - g should be positive values (growth rates)
# - mid should be positive values (midpoints)
# - Both should be finite
# - There should be at least some finite positive values to interpolate

curve = {
    'g': np.array([1.5, 2.0, 2.5, 3.0, 3.5]),
    'mid': np.array([100.0, 500.0, 1000.0, 5000.0, 10000.0])
}

ARGS = {
    'curve': curve,
    'anchor': 1000000.0
}
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
tree_to_dict Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/util.py
code served (permissive licence) · get_code("cff72abe2634a6bc")
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

# Create a simple pytree structure with nested dictionaries and arrays
pytree = {
    'layer1': {
        'weights': jnp.array([[1.0, 2.0, 3.0, 4.0],
                               [5.0, 6.0, 7.0, 8.0]]),
        'bias': jnp.array([0.1, 0.2, 0.3, 0.4])
    },
    'layer2': {
        'weights': jnp.array([[9.0, 10.0, 11.0, 12.0],
                               [13.0, 14.0, 15.0, 16.0]]),
        'bias': jnp.array([0.5, 0.6, 0.7, 0.8])
    },
    'metadata': jnp.array([1.0, 2.0, 3.0, 4.0])
}

ARGS = {
    'pytree': pytree,
    'sep': '.'
}
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
yfmt Ran
ran; our contract draft was wrong, not the code
alexandrumeterez/QuadraticModel/analysis/plot_paper_panels.py
code served (permissive licence) · get_code("a1ad40101802c3be")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

def yfmt(value, _):
    if value == 0:
        return '0'
    sign = '-' if value < 0 else ''
    return f'${sign}10^{{{int(np.round(np.log10(abs(value))))}}}$'

# Setup valid arguments
value = 1000.0
unused_arg = None

ARGS = {
    'value': value,
    '_': unused_arg
}
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 str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 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 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.

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

Due to the complexity of neural network loss landscapes, optimization theory is forced to rely on idealized models, and there is generally a tradeoff between how theoretically tractable the model is, and how accurately it describes the true optimization dynamics. In this work, we stress test the simplest possible model of optimization -the quadratic model -and show that it can be surprisingly predictive in an LLM setting with 150M parameters and 3B training tokens. Specifically, we show that Taylor expanding the model and the loss function at intermediate checkpoints through training can accurately predict the optimization dynamics over windows that can last up to 10% of training. Having established this agreement, we then turn to analyzing the structure of these local quadratic optimization problems through two lenses: the Hessian spectrum and local stability. Using Lanczos quadrature with extremely deep probes, we are able to estimate the Hessian spectrum deep into the tail, and we find a surprising amount of structure in both the eigenvalues and eigenvectors, which depends on the batch size, preconditioner, and training time. We also empirically test local linear stability at intermediate checkpoints and compare it to theoretical predictions to demonstrate that optimization in LLMs typically occurs at a stochastic edge of stability, whose nature is also determined by batch size. Our results indicate the quadratic model may be a theoretically tractable proxy for pretraining optimization dynamics.

For agents

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

get_harvested_code_for_paper("2607.21716")
get_code_for_paper("2607.21716")
have("2607.21716")

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

curl https://syntology.ai/api/ran/2607.21716.json

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

[![Syntology run record](https://syntology.ai/api/ran/2607.21716.svg)](https://syntology.ai/paper/2607.21716)

Connect an agent — have() is free.