SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.24042 · 2026

Hidden-State Privacy Has an Empty Middle

Alexander Bell

arXiv · PDF

Code that ran

15 of 19 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, 0 violated it and 1 has no contract verdict; 13 failures of our instrument; 4 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/2605.24042.json · badge: /api/ran/2605.24042.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
okezue/tcc-research canonical (classifier label)
classifier basis: census
15 ran · 4 no recorded run
FunctionStatusWhere it lives
alpha_curve Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/analyze_sigma_diag.py
code served (permissive licence) · get_code("af3818432452a307")
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 numpy as np

# Construct rows_ml: a list of dictionaries with 'alpha' and 'worst_mahal' keys
rows_ml = [
    {'alpha': 0.1, 'worst_mahal': 1.5},
    {'alpha': 0.1, 'worst_mahal': 2.3},
    {'alpha': 0.1, 'worst_mahal': 1.8},
    {'alpha': 0.5, 'worst_mahal': 3.2},
    {'alpha': 0.5, 'worst_mahal': 2.9},
    {'alpha': 1.0, 'worst_mahal': 4.1},
    {'alpha': 1.0, 'worst_mahal': 3.9},
    {'alpha': 1.0, 'worst_mahal': 4.5},
]

ARGS = {
    'rows_ml': rows_ml,
    'by': 'worst_mahal'
}
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
by_modellayer Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/analyze_sigma_diag.py
code served (permissive licence) · get_code("49139a918e0add33")
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)

# Analyzing the function:
# - It takes 'rows' as a parameter
# - rows is iterable
# - Each element r in rows must be dict-like with keys 'model' and 'layer'
# - The function groups rows by (model, layer) tuple

# Create sample rows with 'model' and 'layer' keys
rows = [
    {'model': 'bert', 'layer': 0, 'value': 1.5},
    {'model': 'bert', 'layer': 1, 'value': 2.3},
    {'model': 'bert', 'layer': 0, 'value': 1.8},
    {'model': 'gpt2', 'layer': 0, 'value': 3.1},
    {'model': 'gpt2', 'layer': 2, 'value': 2.9},
]

ARGS = {
    'rows': rows
}
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
compute_baseline_stats Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/eval_utility.py
code served (permissive licence) · get_code("73375b11ddfd65e7")
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 torch.nn as nn
import torch.nn.functional as F

torch.manual_seed(0)

# Define a minimal model that has a .logits attribute
class MockModel(nn.Module):
    def __init__(self, vocab_size=256, hidden_dim=4):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_dim)
        self.linear = nn.Linear(hidden_dim, vocab_size)
    
    def forward(self, x):
        # x shape: (batch, seq_len)
        embedded = self.embedding(x)  # (batch, seq_len, hidden_dim)
        logits = self.linear(embedded)  # (batch, seq_len, vocab_size)
        
        # Return an object with .logits attribute
        class Output:
            pass
        out = Output()
        out.logits = logits
        return out

# Create model
model = MockModel(vocab_size=256, hidden_dim=4)

# Create a minimal dataset - list of token tensors
# Each token tensor should have at least prefix_len + 1 tokens
dataset = [
    torch.randint(0, 256, (100,)),  # First sample with 100 tokens
    torch.randint(0, 256, (100,)),  # Second sample with 100 tokens
]

device = 'cpu'
prefix_len = 32

ARGS = {
    'model': model,
    'dataset': dataset,
    'device': device,
    'prefix_len': prefix_len,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_gradient_covariance Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
okezue/tcc-research/two_channel/compute_subspace.py
code served (permissive licence) · get_code("2f5d8bf2db8aad63")
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)

# Build a minimal model that satisfies the function's needs
class MiniTransformer(nn.Module):
    def __init__(self, hidden_size=4, vocab_size=100, n_layers=2):
        super().__init__()
        self.config = type('Config', (), {'hidden_size': hidden_size})()
        self.transformer = type('Transformer', (), {})()
        
        # Create minimal transformer blocks
        self.transformer.h = nn.ModuleList([
            nn.Linear(hidden_size, hidden_size) for _ in range(n_layers)
        ])
        
        self.embedding = nn.Embedding(vocab_size, hidden_size)
        self.lm_head = nn.Linear(hidden_size, vocab_size)
    
    def forward(self, x):
        # x shape: (batch, seq_len)
        h = self.embedding(x)  # (batch, seq_len, hidden_size)
        
        # Pass through transformer blocks
        for block in self.transformer.h:
            h = block(h)
        
        logits = self.lm_head(h)  # (batch, seq_len, vocab_size)
        
        return type('Output', (), {'logits': logits})()

# Build minimal tokenizer
class MiniTokenizer:
    def encode(self, text):
        return torch.tensor([1, 2, 3])

# Create model and tokenizer
model = MiniTransformer(hidden_size=4, vocab_size=100, n_layers=2)
tokenizer = MiniTokenizer()

# Create dataset: list of token tensors
# Each element should have at least prefix_len + 1 tokens
dataset = [
    torch.randint(0, 100, (40,), dtype=torch.long),
    torch.randint(0, 100, (40,), dtype=torch.long),
    torch.randint(0, 100, (40,), dtype=torch.long),
]

ARGS = {
    'model': model,
    'tokenizer': tokenizer,
    'layer_idx': 0,
    'dataset': dataset,
    'prefix_len': 32,
    'device': 'cpu',
    'max_samples': 2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
embed_pool Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
okezue/tcc-research/two_channel/exp_attack_defense.py
code served (permissive licence) · get_code("c0789119cb1c3b0b")
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 2, dtype family float, shape (2, 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
import torch.nn as nn
from unittest.mock import Mock, MagicMock

torch.manual_seed(0)

# Build a minimal model with a layer that can be hooked
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.layer1 = nn.Linear(4, 4)
        self.layer2 = nn.Linear(4, 4)
    
    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        return x

model = SimpleModel()

# Create a small pool of sequences (list of tensors)
# Each sequence should be at least ctx=32 long, but we'll make them shorter
# and the function uses seq[:ctx], so we need at least ctx tokens
# Let's create sequences of length 32 with embedding dim 4
pool = [
    torch.randn(32, 4),
    torch.randn(32, 4),
]

# Layer specification - use 'layer2' as the target layer
layer = 'layer2'

# Device - CPU only
dev = 'cpu'

# U_B - optional projection matrix for subspace
# Should be shape (hidden_dim, subspace_dim) for projection
U_B = torch.randn(4, 2)

# Mode - one of 'full', 'behavior', 'identity'
mode = 'full'

# Context length
ctx = 32

# Mock get_layer_block to return the layer, layer index, and something else
def mock_get_layer_block(model, layer_name):
    if layer_name == 'layer2':
        return model.layer2, 1, None
    return None, -1, None

# Patch the function in the module
import sys
import types
two_channel_mock = types.ModuleType('two_channel')
compute_subspace_mock = types.ModuleType('compute_subspace')
compute_subspace_mock.get_layer_block = mock_get_layer_block
two_channel_mock.compute_subspace = compute_subspace_mock
sys.modules['two_channel'] = two_channel_mock
sys.modules['two_channel.compute_subspace'] = compute_subspace_mock

ARGS = {
    'model': model,
    'pool': pool,
    'layer': layer,
    'dev': dev,
    'U_B': U_B,
    'mode': mode,
    'ctx': ctx,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm, two_channel — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 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
get_layer_block Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/adjacency_builder_v2.py
code served (permissive licence) · get_code("6e8b81f3837f0c7b")
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 Linear. 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)

# Create a simple model with transformer.h structure
class SimpleTransformer(nn.Module):
    def __init__(self, num_layers=3):
        super().__init__()
        self.transformer = nn.Module()
        self.transformer.h = nn.ModuleList([
            nn.Linear(4, 4) for _ in range(num_layers)
        ])

model = SimpleTransformer(num_layers=3)
layer = 1

ARGS = {
    'model': model,
    'layer': layer
}
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 Linear — a Linear output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Linear output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_layer_block Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/compute_subspace.py
code served (permissive licence) · get_code("141f652e9aaad06f")
How it ran

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

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build a model with transformer.h architecture
class TransformerBlock(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.linear = nn.Linear(hidden_size, hidden_size)
    
    def forward(self, x):
        return self.linear(x)

class TransformerModel(nn.Module):
    def __init__(self, num_blocks=5, hidden_size=4):
        super().__init__()
        self.transformer = nn.Module()
        self.transformer.h = nn.ModuleList([
            TransformerBlock(hidden_size) for _ in range(num_blocks)
        ])
    
    def forward(self, x):
        for block in self.transformer.h:
            x = block(x)
        return x

model = TransformerModel(num_blocks=5, hidden_size=4)
layer_idx = 2

ARGS = {
    'model': model,
    'layer_idx': layer_idx
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_subspace Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
okezue/tcc-research/two_channel/exp_attack_defense.py
code served (permissive licence) · get_code("358c50879ad84ef7")
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 torch.nn as nn
import torch.nn.functional as F

torch.manual_seed(0)

# Create a minimal model that satisfies the function's usage
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.config = type('Config', (), {
            'hidden_size': 4,
            'vocab_size': 8
        })()
        self.transformer = nn.Module()
        self.transformer.ln_f = nn.LayerNorm(4)
        self.model = nn.Module()
        self.model.norm = nn.LayerNorm(4)
        self.lm_head = nn.Linear(4, 8)
        self.embedding = nn.Embedding(8, 4)
        self.blocks = nn.ModuleList([
            nn.Linear(4, 4),
            nn.Linear(4, 4)
        ])
    
    def forward(self, x):
        # x shape: (batch, seq_len)
        batch_size, seq_len = x.shape
        # Embedding lookup
        h = self.embedding(x)  # (batch, seq_len, hidden_size)
        # Pass through first block
        for block in self.blocks:
            h = block(h)
        return h

# Create a tokenizer stub (not really used in function)
class MinimalTokenizer:
    def encode(self, text):
        return torch.tensor([1, 2, 3, 4, 5])

# Create minimal dataset - list of token sequences
ds = [
    torch.randint(0, 8, (5,)),
    torch.randint(0, 8, (5,)),
]

# Mock get_layer_block function
import sys
from unittest.mock import MagicMock

class MockModule:
    def register_forward_hook(self, fn):
        class Hook:
            def remove(self):
                pass
        return Hook()

def mock_get_layer_block(model, layer):
    blk = model.blocks[0]
    li = 0
    nl = len(model.blocks)
    return blk, li, nl

# Inject the mock into a fake module
import types
mock_module = types.ModuleType('two_channel.compute_subspace')
mock_module.get_layer_block = mock_get_layer_block
sys.modules['two_channel.compute_subspace'] = mock_module
sys.modules['two_channel'] = types.ModuleType('two_channel')

model = MinimalModel()
tok = MinimalTokenizer()
dev = 'cpu'
layer = 0

ARGS = {
    'model': model,
    'tok': tok,
    'ds': ds,
    'layer': layer,
    'dev': dev,
    'k': 2,
    'n_cal': 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm, two_channel — 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
hidden_at_layer Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/adjacency_builder_v2.py
code served (permissive licence) · get_code("6263d79619c3195f")
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 2, dtype family float, shape (2, 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 torch
import torch.nn as nn

torch.manual_seed(0)

# Create a simple model with a block that we can hook into
class SimpleBlock(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 5)
    
    def forward(self, x):
        return self.linear(x)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = nn.Embedding(10, 4)
        self.blk = SimpleBlock()
    
    def forward(self, ids):
        x = self.embedding(ids)
        x = self.blk(x)
        return x

model = SimpleModel()
blk = model.blk

# Create small input tensor: batch size 2, sequence length 3, token indices 0-9
ids = torch.randint(0, 10, (2, 3))

ARGS = {
    "model": model,
    "blk": blk,
    "ids": ids,
}
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 2, dtype family float, shape (2, 5) — shape (2, 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
hidden_with_logits Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/adjacency_builder_v2.py
code served (permissive licence) · get_code("372287728c9df8ab")
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 torch.nn as nn

torch.manual_seed(0)

# Build a minimal model that:
# 1. Accepts ids as input
# 2. Returns an object with .logits attribute
# 3. Works with forward hooks that capture intermediate outputs

class MinimalBlock(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 5)
    
    def forward(self, x):
        return self.linear(x)

class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.embedding = nn.Embedding(10, 4)
        self.blk = MinimalBlock()
        self.head = nn.Linear(5, 6)
    
    def forward(self, ids):
        x = self.embedding(ids)  # (batch, seq, 4)
        x = self.blk(x)  # (batch, seq, 5)
        logits = self.head(x)  # (batch, seq, 6)
        
        class Output:
            pass
        out = Output()
        out.logits = logits
        return out

model = MinimalModel()
model.eval()

# Create a block reference - should be the same as model.blk
blk = model.blk

# Create input ids: shape (batch_size=2, seq_len=3)
ids = torch.randint(0, 10, (2, 3), dtype=torch.long)

ARGS = {
    'model': model,
    'blk': blk,
    'ids': ids
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — 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
load_all Ran
ran; our contract draft was wrong, not the code
okezue/tcc-research/two_channel/analyze_sigma_diag.py
code served (permissive licence) · get_code("9aafed64160f8319")
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 json
import os
import tempfile
import shutil

# Create a temporary directory for test data
temp_dir = tempfile.mkdtemp()

# Create sample JSON files matching the expected format
sigma_data_1 = {
    'model': 'model_v1',
    'layer': 'layer_0',
    'd': 128,
    'F_diag_sum': 42.5,
    'rows': [
        {'sigma': 0.1, 'value': 1.0},
        {'sigma': 0.2, 'value': 2.0}
    ]
}

sigma_data_2 = {
    'model': 'model_v2',
    'layer': 'layer_1',
    'd': 256,
    'F_diag_sum': 85.3,
    'rows': [
        {'sigma': 0.15, 'value': 1.5}
    ]
}

# Write test JSON files
with open(os.path.join(temp_dir, 'sigma_diag_0.json'), 'w') as f:
    json.dump(sigma_data_1, f)

with open(os.path.join(temp_dir, 'sigma_diag_1.json'), 'w') as f:
    json.dump(sigma_data_2, f)

ARGS = {
    'out_dir': temp_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_model Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
okezue/tcc-research/two_channel/compute_subspace.py
code served (permissive licence) · get_code("5148fb5c61d1921e")
How it ran

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

import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutputWithPast

torch.manual_seed(0)

# Create a minimal stand-in tokenizer
class MinimalTokenizer:
    def __init__(self):
        self.pad_token = "<pad>"
        self.eos_token = "<eos>"
        self.vocab_size = 100
        
    def __call__(self, text, **kwargs):
        return {"input_ids": torch.tensor([[1, 2, 3]]), "attention_mask": torch.tensor([[1, 1, 1]])}
    
    @classmethod
    def from_pretrained(cls, model_id):
        return cls()

# Create a minimal stand-in causal LM model
class MinimalCausalLM(PreTrainedModel):
    def __init__(self, config=None):
        if config is None:
            config = PretrainedConfig()
        super().__init__(config)
        self.hidden_size = 64
        self.embedding = nn.Embedding(100, 64)
        self.linear = nn.Linear(64, 100)
        self.output_hidden_states = True
    
    def forward(self, input_ids, attention_mask=None, **kwargs):
        hidden = self.embedding(input_ids)
        logits = self.linear(hidden)
        return CausalLMOutputWithPast(
            loss=None,
            logits=logits,
            hidden_states=(hidden,)
        )
    
    def eval(self):
        return super().eval()
    
    def to(self, device):
        return super().to(device)
    
    @classmethod
    def from_pretrained(cls, model_id, **kwargs):
        model = cls()
        model.output_hidden_states = kwargs.get('output_hidden_states', False)
        return model

# Monkey patch AutoModelForCausalLM and AutoTokenizer
import transformers
transformers.AutoModelForCausalLM.from_pretrained = MinimalCausalLM.from_pretrained
transformers.AutoTokenizer.from_pretrained = MinimalTokenizer.from_pretrained

ARGS = {
    "model_id": "gpt2",
    "device": "cpu"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_ds Ran
ran, honoured the contract we drafted
okezue/tcc-research/two_channel/eval_quotient_batch.py
code served (permissive licence) · get_code("c36ccd0250aec47f")
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 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 torch
from unittest.mock import MagicMock

torch.manual_seed(0)

class MockTokenizer:
    def __call__(self, text, add_special_tokens=False, truncation=True, max_length=512):
        # Return a dict with 'input_ids' key containing token IDs
        # Make it long enough to satisfy the sequence length requirement
        token_ids = [101, 2054, 2003, 1037, 3231, 1029, 102] * 100
        return {'input_ids': token_ids[:max_length]}

class MockDataset:
    def __init__(self):
        self.data = [
            {'text': 'This is a sample text with enough content to pass the length check. ' * 3},
            {'text': 'Another example of text that is long enough for processing. ' * 3},
            {'text': 'Yet another text sample with sufficient length requirements. ' * 3},
        ]
    
    def __iter__(self):
        return iter(self.data)

def mock_load_dataset(name, config, split):
    return MockDataset()

# Patch the load_dataset function
import datasets
original_load_dataset = datasets.load_dataset
datasets.load_dataset = mock_load_dataset

tok = MockTokenizer()
n = 2
sl = 64

ARGS = {
    'tok': tok,
    'n': n,
    'sl': sl,
    'seed': 42
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, dtype family 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
make_ds Ran
ran on a synthesized input
okezue/tcc-research/two_channel/exp_attack_defense.py
code served (permissive licence) · get_code("7157763e28dc845c")
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 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 torch

class MockTokenizer:
    """Stand-in tokenizer that mimics HuggingFace tokenizer behavior"""
    def __call__(self, text, add_special_tokens=False, truncation=True, max_length=512):
        # Simulate tokenization by returning a dict with 'input_ids' key
        # Split text into words and convert to token IDs
        words = text.split()
        token_ids = [hash(word) % 1000 for word in words[:max_length]]
        return {'input_ids': token_ids}

# Create a mock tokenizer
tok = MockTokenizer()

ARGS = {
    'tok': tok,
    'n': 10,
    'sl': 5,
    'seed': 42
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, dtype family 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
random_neighbors Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
okezue/tcc-research/two_channel/adjacency_builder.py
code served (permissive licence) · get_code("62c2bb4f1efb997d")
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 2, dtype family float, shape (1024, 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)

# H is used as:
# - H.shape[0] to get n (number of rows)
# - H[i:i + 1] to slice a single row and repeat it
# - H[j] to index multiple rows
# So H needs to be a 2D tensor where we can index and slice

H = torch.randn(2, 4)

ARGS = {
    "H": H,
    "k": 512,
    "seed": 0,
}
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 2, dtype family float, shape (1024, 4) — shape (1024, 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
calibrate_quant_stats Not yet run
unverified (harvested, no recorded run)
okezue/tcc-research/two_channel/eval_utility.py
code served (permissive licence) · get_code("1d8e979edfb64149")
How it ran

no invocation recorded.

make_test Not yet run
unverified (harvested, no recorded run)
okezue/tcc-research/two_channel/eval_inv_direct.py
code served (permissive licence) · get_code("edec64212e4621cf")
How it ran

no invocation recorded.

nearest_neighbors Not yet run
unverified (harvested, no recorded run)
okezue/tcc-research/two_channel/adjacency_builder.py
code served (permissive licence) · get_code("7383c4c2299cfe66")
How it ran

no invocation recorded.

top_prob_neighbors Not yet run
unverified (harvested, no recorded run)
okezue/tcc-research/two_channel/adjacency_builder.py
code served (permissive licence) · get_code("9e3d1668d28583e5")
How it ran

no invocation recorded.

Also present in other papers' code

Of 19 samples lifted from this paper's linked repositories, 19 carry a function identity and 0 do not. Of the 19 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 19 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

Of 1,536 Gaussian release covariances we tested for single-layer hidden-state privacy, zero achieve both moderate utility and moderate privacy against an adaptive retrieval attacker. We prove a complementary Fisher-ball lower bound: every fullrank Gaussian release at O(1) Fisher utility admits a direction whose Mahalanobis signal grows linearly in hidden width, ruling out uniform Gaussian safety in the class and matching the empirical empty middle. The diagonal inverse-Fisher release is the unique minimax-optimal diagonal mechanism at first-order KL budget K and the only release with worst-attacker top-1 ≤ 0.001 at every point of a 32 model-layer grid, but it sits on a privacy/utility edge rather than filling the middle. A generalized-eigen mechanism reaching 13× Pareto reduction under Euclidean retrieval collapses to 100% top-1 under the adaptive Mahalanobis attacker, and a full-trajectory sequence inverter recovers 94% of clean GPT-2 prefixes but 0% under Σ diag . A split-memory transformer trained from scratch reaches G Mah ∈ [20, 33] at 90M and maintains a 6-24× advantage over same-budget GPT baselines from 30M to 1B at a fixed-token language-modeling loss penalty; pretrained models top out at 9.3. These results reframe hidden-state release from mechanism-design within the Gaussian class to architecture or release co-design.

For agents

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

get_harvested_code_for_paper("2605.24042")
get_code_for_paper("2605.24042")
have("2605.24042")

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

curl https://syntology.ai/api/ran/2605.24042.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.24042.svg)](https://syntology.ai/paper/2605.24042)

Connect an agent — have() is free.