SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.11940 · 2026

Exhaustive Circuit Mapping of a Single-Cell Foundation Model Reveals Massive Redundancy, Heavy-Tailed Hub Architecture, and Layer-Dependent Differentiation Control

Ihor Kendiukhov

arXiv · PDF · Open in the Atlas

Code that ran

17 of 23 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; 7 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 10 failures of our instrument; 6 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/2603.11940.json · badge: /api/ran/2603.11940.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
Biodyn-AI/sae-biological-map canonical (classifier label)
classifier basis: census
17 ran · 6 no recorded run
FunctionStatusWhere it lives
build_cells Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/p2_steering.py
code served (permissive licence) · get_code("b406eab057b704b0")
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
import os
import tempfile
import h5py
import pickle


class MockOrigObject:
    def __init__(self):
        self.TOKEN_DICTS_DIR = tempfile.gettempdir()
        self.IMMUNE_H5AD = os.path.join(tempfile.gettempdir(), "immune.h5ad")
        self.HEMATOPOIETIC_TISSUES = ["blood", "bone marrow"]
        self.MYELOID_TYPES = ["monocyte", "macrophage", "neutrophil"]
        self.LYMPHOID_TYPES = ["t cell", "b cell", "nk cell"]
        self.MAX_SEQ_LEN = 256
        
        self._setup_files()
    
    def _setup_files(self):
        # Create token dictionary
        token_dict = {"ENSG001": 1, "ENSG002": 2, "ENSG003": 3}
        with open(os.path.join(self.TOKEN_DICTS_DIR, 'token_dictionary_gc104M.pkl'), 'wb') as f:
            pickle.dump(token_dict, f)
        
        # Create median dictionary
        median_dict = {"ENSG001": 1.5, "ENSG002": 2.0}
        with open(os.path.join(self.TOKEN_DICTS_DIR, 'gene_median_dictionary_gc104M.pkl'), 'wb') as f:
            pickle.dump(median_dict, f)
        
        # Create gene name id dictionary
        name_id = {"GENE1": "ENSG001", "GENE2": "ENSG002", "GENE3": "ENSG003"}
        with open(os.path.join(self.TOKEN_DICTS_DIR, 'gene_name_id_dict_gc104M.pkl'), 'wb') as f:
            pickle.dump(name_id, f)
        
        # Create mock h5ad file
        with h5py.File(self.IMMUNE_H5AD, 'w') as f:
            # Create X (sparse matrix) - using CSR format
            n_cells = 10
            n_genes = 3
            nnz = 15
            
            data = f.create_dataset('X/data', data=np.random.randn(nnz).astype(np.float32))
            indices = f.create_dataset('X/indices', data=np.array([0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2], dtype=np.int32))
            indptr = f.create_dataset('X/indptr', data=np.array([0, 2, 4, 6, 8, 10, 12, 14, 15, 15, 15], dtype=np.int32))
            
            # Create var (genes)
            var_group = f.create_group('var')
            var_group.create_dataset('_index', data=np.array(['GENE1', 'GENE2', 'GENE3'], dtype='S10'))
            
            # Create obs (cells)
            obs_group = f.create_group('obs')
            cell_types_data = np.array(['Monocyte', 'T Cell', 'B Cell', 'Macrophage', 'NK Cell', 'Neutrophil', 'Monocyte', 'T Cell', 'B Cell', 'Monocyte'], dtype='S20')
            obs_group.create_dataset('cell_type', data=cell_types_data)
            
            tissues_data = np.array(['blood', 'blood', 'bone marrow', 'bone marrow', 'blood', 'blood', 'bone marrow', 'blood', 'blood', 'bone marrow'], dtype='S20')
            obs_group.create_dataset('tissue_in_publication', data=tissues_data)
    
    def load_categorical_column(self, group, col_name):
        """Mock implementation to load categorical column from h5 group."""
        data = group[col_name][:]
        if isinstance(data[0], bytes):
            return np.array([d.decode('utf-8') for d in data])
        return data
    
    def load_sparse_row(self, X_group, idx, n_genes):
        """Mock implementation to load a sparse row."""
        indptr = X_group['indptr'][:]
        indices = X_group['indices'][:]
        data = X_group['data'][:]
        
        row_start = indptr[idx]
        row_end = indptr[idx + 1]
        
        row = np.zeros(n_genes, dtype=np.float32)
        row[indices[row_start:row_end]] = data[row_start:row_end]
        return row
    
    def tokenize_cell(self, row, mvi, mti, mmd, max_seq_len):
        """Mock implementation of tokenization."""
        # Simple tokenization: return non-zero indices as tokens
        nonzero = np.where(row[mvi] > 0)[0]
        if len(nonzero) == 0:
            return None
        tokens = mti[nonzero[:max_seq_len]]
        return tokens.tolist()


orig = MockOrigObject()

ARGS = {
    "orig": orig,
    "n_total": 500
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
h5py, 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
feature_pool Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/p3_triplets_extended.py
code served (permissive licence) · get_code("01b61380fdf5912e")
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 numpy as np

class MockFeatureCatalog:
    def __init__(self):
        self.features = {
            0: {
                0: {'activation_freq': 0.05},
                1: {'activation_freq': 0.01},
                2: {'activation_freq': 0.03},
            },
            5: {
                10: {'activation_freq': 0.04},
                11: {'activation_freq': 0.015},
                12: {'activation_freq': 0.06},
                13: {'activation_freq': 0.02},
            },
            11: {
                20: {'activation_freq': 0.025},
                21: {'activation_freq': 0.008},
                22: {'activation_freq': 0.035},
            }
        }
        self.labels = {
            0: {
                0: 'DNA Damage Response',
                1: 'Mitotic Spindle',
                2: 'Chromatin Remodeling',
            },
            5: {
                10: 'Translation Initiation',
                11: 'Ribosomal Assembly',
                12: 'mRNA Splicing',
                13: 'Vesicle Transport',
            },
            11: {
                20: 'Oxidative Phosphorylation',
                21: 'Histone Modification',
                22: 'Double-strand Break Repair',
            }
        }
    
    def load_feature_catalog(self, layer):
        active = list(self.features[layer].keys())
        info = self.features[layer]
        return active, info
    
    def load_feature_labels(self, layer):
        return self.labels[layer]

torch.manual_seed(0)

orig = MockFeatureCatalog()

ARGS = {
    'orig': orig
}
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 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
hedges_g Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/revision/p1b_null_fdr_thresholds.py
code served (permissive licence) · get_code("673181d7723a0f76")
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 torch
import numpy as np

torch.manual_seed(0)

# hedges_g takes two parameters: d and n
# d appears to be a numeric effect size value
# n appears to be a sample size (used in comparison and arithmetic)

d_value = 0.75
n_value = 10

ARGS = {
    "d": d_value,
    "n": n_value,
}
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 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
jload Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/make_figures.py
code served (permissive licence) · get_code("8c0c8031ac6f0fc4")
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 json
import tempfile
import os

# Create a temporary JSON file with test data
temp_dir = tempfile.mkdtemp()
test_file = os.path.join(temp_dir, "test.json")

test_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "object"}
}

with open(test_file, "w") as f:
    json.dump(test_data, f)

ARGS = {
    "p": 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 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_categorical_column Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/exhaustive_feature_tracing.py
code served (permissive licence) · get_code("2f11e78bedd3d2ac")
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 ndarray, ndim 1, dtype family float, shape (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 h5py
import numpy as np
import tempfile
import os

# Create a temporary HDF5 file with test data
temp_dir = tempfile.mkdtemp()
h5_path = os.path.join(temp_dir, "test.h5")

with h5py.File(h5_path, "w") as f:
    # Create a group with categorical data (categories + codes structure)
    cat_group = f.create_group("categorical_col")
    categories = np.array([b"cat", b"dog", b"bird"], dtype='S10')
    codes = np.array([0, 1, 2, 0, 1], dtype=np.int32)
    cat_group.create_dataset("categories", data=categories)
    cat_group.create_dataset("codes", data=codes)
    
    # Create a simple dataset column
    f.create_dataset("simple_col", data=np.array([1.0, 2.0, 3.0, 4.0, 5.0]))

# Open the HDF5 file and prepare arguments
h5file = h5py.File(h5_path, "r")
h5group = h5file

ARGS = {
    "h5group": h5group,
    "col_name": "categorical_col"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
h5py, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (5,) — shape (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
load_features Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/revision/make_figures.py
code served (permissive licence) · get_code("4d8ed6301a3c8094")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import json
import tempfile
import shutil

# Create a temporary directory structure to mimic the expected project layout
temp_dir = tempfile.mkdtemp()
proj_root = temp_dir
eXH_dir = os.path.join(proj_root, 'experiments', 'phase8_exhaustive_tracing')
os.makedirs(eXH_dir, exist_ok=True)

# Set the environment variable
os.environ['SAE_PROJ_ROOT'] = proj_root

# Create sample feature JSON files
feature_data_1 = {
    'feature_idx': 0,
    'total_significant_edges': 5,
    'label': 'test_label_1',
    'activation_freq': 0.75,
    'downstream_edges': {
        'layer_1': {'n_significant': 3},
        'layer_2': {'n_significant': 2}
    }
}

feature_data_2 = {
    'feature_idx': 1,
    'total_significant_edges': 8,
    'label': 'unannotated',
    'activation_freq': 0.45,
    'downstream_edges': {
        'layer_1': {'n_significant': 4},
        'layer_2': {'n_significant': 4}
    }
}

feature_data_3 = {
    'feature_idx': 2,
    'total_significant_edges': 3,
    'label': 'test_label_2',
    'activation_freq': 0.92,
    'downstream_edges': {
        'layer_1': {'n_significant': 1},
        'layer_2': {'n_significant': 2}
    }
}

# Write feature files
with open(os.path.join(eXH_dir, 'feature_0.json'), 'w') as f:
    json.dump(feature_data_1, f)

with open(os.path.join(eXH_dir, 'feature_1.json'), 'w') as f:
    json.dump(feature_data_2, f)

with open(os.path.join(eXH_dir, 'feature_2.json'), 'w') as f:
    json.dump(feature_data_3, f)

# Define arguments for the load_features function
ARGS = {
    'edge_override': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_sparse_row Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/trajectory_steering.py
code served (permissive licence) · get_code("559b022277f2b71e")
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 ndarray, ndim 1, dtype family float, shape (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 numpy as np
from unittest.mock import MagicMock

# Create a mock f_group object that behaves like an HDF5 group
f_group = MagicMock()

# Set up sparse matrix data in CSR (Compressed Sparse Row) format
# For a sparse matrix with 3 rows and 5 columns
n_rows = 3
n_cols = 5

# indptr: row pointers (4 elements for 3 rows)
indptr = np.array([0, 2, 3, 5], dtype=np.int32)

# indices: column indices for non-zero elements
indices = np.array([0, 3, 1, 2, 4], dtype=np.int32)

# data: non-zero values
data = np.array([1.5, 2.3, 4.1, 3.2, 5.0], dtype=np.float32)

# Configure the mock to return these arrays when indexed
f_group.__getitem__ = MagicMock(side_effect=lambda key: {
    'indptr': indptr,
    'indices': indices,
    'data': data
}[key])

# Also support direct attribute access
f_group.indptr = indptr
f_group.indices = indices
f_group.data = data

row_idx = 1
n_cols = 5

ARGS = {
    'f_group': f_group,
    'row_idx': row_idx,
    'n_cols': n_cols
}
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 ndarray, ndim 1, dtype family float, shape (5,) — shape (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
maybe Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/make_figures.py
code served (permissive licence) · get_code("351f56a11eaf34fa")
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 json
import os
import tempfile

# Create a temporary directory and a test JSON file
temp_dir = tempfile.mkdtemp()
test_json_path = os.path.join(temp_dir, "test.json")

# Write a valid JSON file
test_data = {"key": "value", "number": 42, "list": [1, 2, 3]}
with open(test_json_path, "w") as f:
    json.dump(test_data, f)

# Define ARGS with the path parameter
ARGS = {
    "p": test_json_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 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
mean_pool Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Biodyn-AI/sae-biological-map/src/revision/p2_steering.py
code served (permissive licence) · get_code("e39b2842a3fb120c")
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 Tensor, ndim 1, dtype family float, shape (4,) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# hidden: tensor that will be indexed by gene_positions
# The function does hidden[gene_positions].mean(dim=0)
# So hidden needs to be indexable and support mean(dim=0)
# gene_positions will be used as an index into hidden

hidden = torch.randn(5, 4)  # 5 genes, 4 dimensions

# gene_positions: indices to select from hidden
# Should be a tensor or list of indices
gene_positions = torch.tensor([0, 2, 4])  # Select 3 genes out of 5

ARGS = {
    "hidden": hidden,
    "gene_positions": gene_positions,
}
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 Tensor, ndim 1, dtype family float, shape (4,) — shape (4,) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
pathway_of Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/p3_triplets_extended.py
code served (permissive licence) · get_code("07fd78908105f548")
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.

ARGS = {
    'label': 'DNA damage and repair mechanisms in double-strand breaks'
}
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 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
perm_stats Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/revision/p1b_null_fdr_thresholds.py
code served (permissive licence) · get_code("2b8bf9b0966d7a1e")
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 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 numpy as np

# Construct valid arguments for perm_stats function
# mat: (n, T) array of values
# signs: (B, n) array in {-1, +1}

np.random.seed(0)

n = 4  # number of samples
T = 3  # number of features/tasks
B = 2  # batch size

# Create mat as (n, T) numpy array with random values
mat = np.random.randn(n, T)

# Create signs as (B, n) array with values in {-1, +1}
signs = np.random.choice([-1, 1], size=(B, n)).astype(np.float64)

ARGS = {
    "mat": mat,
    "signs": signs,
}
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 3, 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
pl_cdf Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Biodyn-AI/sae-biological-map/src/revision/p4_tail_fitting.py
code served (permissive licence) · get_code("5f6c3504b3a44c1b")
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 ndarray, ndim 1, dtype family float, shape (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 numpy as np

# Analyze the function:
# - vals: array-like, used with .max() and passed to np.searchsorted
# - alpha: numeric scalar, used in exponentiation
# - xmin: numeric scalar, used as start of arange

# Create valid test arguments
vals = np.array([1, 2, 3, 5, 8], dtype=float)
alpha = 2.5
xmin = 1

ARGS = {
    "vals": vals,
    "alpha": alpha,
    "xmin": xmin,
}
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 ndarray, ndim 1, dtype family float, shape (5,) — shape (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
pl_logpmf Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/revision/p4_tail_fitting.py
code served (permissive licence) · get_code("76e4898286b59af6")
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 float64, ndim 0, 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

# Analyzing the function:
# - pl_logpmf takes parameters: x, alpha, xmin
# - x: used with np.log(), so should be a numeric value or array
# - alpha: used with multiplication and passed to _zeta_tail
# - xmin: passed to _zeta_tail as starting point for np.arange
# All parameters are numeric (int or float)

ARGS = {
    "x": 5,
    "alpha": 1.5,
    "xmin": 1
}
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 float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py could not observe it (no_result)
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
precompute_clean Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/p1a_retrace_with_deltas.py
code served (permissive licence) · get_code("af97fac5a8f58bde")
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 torch
import numpy as np
from torch import nn

torch.manual_seed(0)

# Mock SAE (Sparse Autoencoder) class
class MockSAE(nn.Module):
    def __init__(self, input_dim, latent_dim):
        super().__init__()
        self.input_dim = input_dim
        self.latent_dim = latent_dim
        self.encoder = nn.Linear(input_dim, latent_dim)
    
    def encode(self, x):
        # Returns sparse encoding and topk indices
        encoded = self.encoder(x)
        # Simulate topk selection
        topk_vals, topk_indices = torch.topk(torch.abs(encoded), k=min(2, encoded.shape[-1]), dim=-1)
        return encoded, topk_indices
    
    def to(self, device):
        return super().to(device)

# Mock Model with hidden_states output
class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.hidden_dim = 4
        self.num_layers = 18
        # Create embeddings and layers
        self.embedding = nn.Embedding(1024, self.hidden_dim)
        self.layers = nn.ModuleList([nn.Linear(self.hidden_dim, self.hidden_dim) for _ in range(self.num_layers)])
    
    def forward(self, input_ids, attention_mask):
        # input_ids: (batch, seq_len)
        x = self.embedding(input_ids)  # (batch, seq_len, hidden_dim)
        hidden_states = [x]
        
        for layer in self.layers:
            x = layer(x)
            hidden_states.append(x)
        
        # Create a mock output object
        class Output:
            pass
        
        output = Output()
        output.hidden_states = hidden_states
        return output

# Mock SAE cache
class MockSAECache:
    def __init__(self):
        self.hidden_dim = 4
        self.latent_dim = 8
        self.saes = {}
        self.means = {}
        
        # Create SAEs and means for all needed layers
        for layer_idx in [5, 6, 11, 17]:
            self.saes[layer_idx] = MockSAE(self.hidden_dim, self.latent_dim)
            self.means[layer_idx] = np.zeros(self.hidden_dim, dtype=np.float32)
    
    def get(self, layer_idx):
        return self.saes[layer_idx], self.means[layer_idx]

# Create test data
model = MockModel()
device = torch.device('cpu')
sae_cache = MockSAECache()

# Create token sequences with some valid tokens (not 2 or 3)
all_tokens = [
    np.array([0, 1, 4, 5, 6], dtype=np.int64),  # sequence 1 with valid tokens
    np.array([2, 3, 2, 3], dtype=np.int64),      # sequence 2 with all masked tokens
    np.array([1, 4, 5], dtype=np.int64),         # sequence 3 with valid tokens
]

ARGS = {
    'all_tokens': all_tokens,
    'model': model,
    'device': device,
    'sae_cache': sae_cache,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — 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
select_triplets Ran
ran; our contract draft was wrong, not the code
Biodyn-AI/sae-biological-map/src/revision/p3_triplets_extended.py
code served (permissive licence) · get_code("66af41ef506b2dcc")
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

# Build the pool structure that the function expects
# pool[L] is a dict where keys are feature IDs and values are dicts with 'pathway' and 'freq'
pool = {}

# Define feature data for each layer
feature_data = {
    0: {
        'f0': {'pathway': 'mitosis', 'freq': 100, 'label': 'mitosis_feature_0'},
        'f1': {'pathway': 'vesicle', 'freq': 90, 'label': 'vesicle_feature_1'},
        'f2': {'pathway': 'translation', 'freq': 80, 'label': 'translation_feature_2'},
        'f3': {'pathway': 'chromatin', 'freq': 70, 'label': 'chromatin_feature_3'},
        'f4': {'pathway': 'rna', 'freq': 60, 'label': 'rna_feature_4'},
        'f5': {'pathway': 'metabolism', 'freq': 50, 'label': 'metabolism_feature_5'},
        'f6': {'pathway': 'ddr', 'freq': 40, 'label': 'ddr_feature_6'},
    },
    5: {
        'f10': {'pathway': 'mitosis', 'freq': 95, 'label': 'mitosis_feature_10'},
        'f11': {'pathway': 'vesicle', 'freq': 85, 'label': 'vesicle_feature_11'},
        'f12': {'pathway': 'translation', 'freq': 75, 'label': 'translation_feature_12'},
        'f13': {'pathway': 'chromatin', 'freq': 65, 'label': 'chromatin_feature_13'},
        'f14': {'pathway': 'rna', 'freq': 55, 'label': 'rna_feature_14'},
        'f15': {'pathway': 'metabolism', 'freq': 45, 'label': 'metabolism_feature_15'},
        'f16': {'pathway': 'ddr', 'freq': 35, 'label': 'ddr_feature_16'},
    },
    11: {
        'f20': {'pathway': 'mitosis', 'freq': 98, 'label': 'mitosis_feature_20'},
        'f21': {'pathway': 'vesicle', 'freq': 88, 'label': 'vesicle_feature_21'},
        'f22': {'pathway': 'translation', 'freq': 78, 'label': 'translation_feature_22'},
        'f23': {'pathway': 'chromatin', 'freq': 68, 'label': 'chromatin_feature_23'},
        'f24': {'pathway': 'rna', 'freq': 58, 'label': 'rna_feature_24'},
        'f25': {'pathway': 'metabolism', 'freq': 48, 'label': 'metabolism_feature_25'},
        'f26': {'pathway': 'ddr', 'freq': 38, 'label': 'ddr_feature_26'},
    },
}

for layer_id, features in feature_data.items():
    pool[layer_id] = features

ARGS = {
    'pool': pool,
    'n_same': 4,
    'n_cross': 6,
    'n_random': 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
stats_from Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/revision/p1b_null_fdr_thresholds.py
code served (permissive licence) · get_code("664f812244c7bac5")
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

# Create a small numpy matrix with shape (n_samples, n_features)
# Using distinct sizes: 5 samples, 3 features
mat = np.array([
    [1.0, 2.0, 3.0],
    [2.0, 4.0, 6.0],
    [3.0, 6.0, 9.0],
    [-1.0, 0.0, 2.0],
    [0.0, 1.0, -1.0]
], dtype=np.float32)

ARGS = {
    "mat": mat
}
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
stats_from Ran
ran, honoured the contract we drafted
Biodyn-AI/sae-biological-map/src/revision/p1c_normmatched_control.py
code served (permissive licence) · get_code("d4f1a29c33f25447")
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

# The function stats_from takes a single parameter 'mat' which should be a numpy array
# Looking at the function body:
# - mat.shape[0] is accessed (number of rows)
# - mat.mean(0) is called (mean along axis 0)
# - mat.std(0, ddof=1) is called (standard deviation along axis 0)
# - mat > 0 is used (element-wise comparison)
# - (mat > 0).sum(0) is called (sum along axis 0)
# - N_FEATURES = 4608, so the second dimension should be 4608

# Create a small test matrix with shape (n_samples, N_FEATURES)
# We'll use a batch of 5 samples with 4608 features
N_FEATURES = 4608
np.random.seed(0)

# Create a matrix with some variation to test the statistics calculation
mat = np.random.randn(5, N_FEATURES) * 10 + np.random.randn(1, N_FEATURES)

ARGS = {
    "mat": mat
}
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
build_model Not yet run
unverified (harvested, no recorded run)
Biodyn-AI/sae-biological-map/src/revision/p1a_retrace_with_deltas.py
code served (permissive licence) · get_code("9242b130d7b1594b")
How it ran

no invocation recorded.

fit_powerlaw_discrete Not yet run
unverified (harvested, no recorded run)
Biodyn-AI/sae-biological-map/src/revision/p4_tail_fitting.py
code served (permissive licence) · get_code("fc105b6a966ef35a")
How it ran

no invocation recorded.

load_and_tokenize_cells Not yet run
unverified (harvested, no recorded run)
Biodyn-AI/sae-biological-map/src/exhaustive_feature_tracing.py
code served (permissive licence) · get_code("03d7d79f35178aba")
How it ran

no invocation recorded.

n_edges Not yet run
unverified (harvested, no recorded run)
Biodyn-AI/sae-biological-map/src/revision/p1c_normmatched_control.py
code served (permissive licence) · get_code("b76ae80605127006")
How it ran

no invocation recorded.

run_tail Not yet run
unverified (harvested, no recorded run)
Biodyn-AI/sae-biological-map/src/revision/p1a_retrace_with_deltas.py
code served (permissive licence) · get_code("695d0d66410f71ee")
How it ran

no invocation recorded.

tokenize_cell Not yet run
unverified (harvested, no recorded run)
Biodyn-AI/sae-biological-map/src/exhaustive_feature_tracing.py
code served (permissive licence) · get_code("5e5d73ae58cd79fd")
How it ran

no invocation recorded.

Also present in other papers' code

Of 23 samples lifted from this paper's linked repositories, 23 carry a function identity and 0 do not. Of the 23 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 23 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

Mechanistic interpretability of biological foundation models has relied on selective feature sampling, pairwise interaction testing, and observational trajectory analysis-each introducing systematic biases. Here we present three experiments that address these limitations through exhaustive circuit tracing, higher-order combinatorial ablation, and causal trajectory steering in Geneformer, a transformer-based single-cell foundation model. First, exhaustive tracing of all 4,065 active sparse autoencoder features at layer 5 yields 1,393,850 significant downstream edges-a 27-fold expansion over selective sampling-revealing a heavy-tailed hub distribution where 1.8% of features account for disproportionate connectivity and 40% of the top-20 hubs lack biological annotation, demonstrating systematic annotation bias in prior selective analyses. Second, three-way combinatorial ablation across 8 feature triplets shows that redundancy deepens monotonically with interaction order (three-way ratio 0.59 vs. pairwise 0.74) with zero synergy, confirming that the model's circuit architecture is fundamentally subadditive at all tested orders. Third, trajectory-guided feature steering establishes a causal link between layer position and differentiation directionality: late-layer features (L17) universally push cell states toward maturity (fraction positive = 1.0), while early-and mid-layer features (L0, L11) predominantly push away from maturity (fraction positive = 0.00-0.58), transitioning from correlation to causal evidence for layer-dependent cell state control.

For agents

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

get_harvested_code_for_paper("2603.11940")
get_code_for_paper("2603.11940")
have("2603.11940")

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

curl https://syntology.ai/api/ran/2603.11940.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.11940.svg)](https://syntology.ai/paper/2603.11940)

Connect an agent — have() is free.