SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2512.15674 · ICML · 2025

Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers

Owain Evans, Ens Paris-Saclay, Subhash Kantamneni, Euan Ong, James Chua, Adam Karvonen, Samuel Marks, Julian Minder, Clément Dumas, Truthful Ai, Kit Fraser-Taliente, Daniel Wen, and 1 more

arXiv · PDF · Open in the Atlas

Code that ran

21 of 25 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 0 have no contract verdict; 20 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/2512.15674.json · badge: /api/ran/2512.15674.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
cywinski/eliciting-secret-knowledge canonical (classifier label)
classifier basis: census
21 ran · 4 no recorded run
FunctionStatusWhere it lives
calculate_prefill_statistics Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/calculate_auditing_metrics_gpt.py
code served (permissive licence) · get_code("4d0dda598837be0b")
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 test data for grouped_by_prefill
# The function expects a dict mapping prefill (str) to list of entries (dicts)
# Each entry should have at least the metric_name key with a numeric value

grouped_by_prefill = {
    'prefill_a': [
        {'max_gpt_similarity': 0.85, 'other_metric': 10},
        {'max_gpt_similarity': 0.92, 'other_metric': 15},
        {'max_gpt_similarity': 0.78, 'other_metric': 12},
    ],
    'prefill_b': [
        {'max_gpt_similarity': 0.65, 'other_metric': 5},
        {'max_gpt_similarity': 0.71, 'other_metric': 8},
    ],
    'prefill_c': [
        {'max_gpt_similarity': 0.95, 'other_metric': 20},
    ],
}

metric_name = 'max_gpt_similarity'

ARGS = {
    'grouped_by_prefill': grouped_by_prefill,
    'metric_name': metric_name,
}
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
decode_base64_partially Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/guess_secret_side_constraint_iterative.py
code served (permissive licence) · get_code("96d9d0b4f37999dd")
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 base64

# Construct a valid base64 string
test_string = "Hello, World!"
encoded = base64.b64encode(test_string.encode('utf-8')).decode('utf-8')

ARGS = {
    'base64_str': encoded,
    'num_chars': 8
}
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
extract_pairs Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/utils/guess_secret_word.py
code served (permissive licence) · get_code("f56b76b763051ff9")
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 torch
import numpy as np
from typing import Any, Dict, List

torch.manual_seed(0)

# Construct a valid data dict for extract_pairs function
data_standard = {
    'results': [
        {
            'user_prompt': 'What is 2+2?',
            'model_response': 'The answer is 4.',
            'top_k_tokens': ['the', 'answer', 'is'],
            'top_k_features': ['feature_1', 'feature_2']
        },
        {
            'user_prompt': 'Tell me a joke.',
            'model_response': 'Why did the chicken cross the road?',
            'top_k_tokens': ['why', 'did', 'the'],
            'top_k_features': ['feature_3', 'feature_4']
        }
    ],
    'metadata': {
        'strategy': 'standard'
    }
}

data_prefill = {
    'results': [
        {
            'user_prompt': 'Complete this',
            'model_response': 'completion text',
            'prefill': 'starting text',
            'top_k_tokens': ['completion'],
            'top_k_features': ['feat1']
        }
    ],
    'metadata': {
        'strategy': 'prefill'
    }
}

data_persona = {
    'results': [
        {
            'user_prompt': 'Hello',
            'model_response': 'Hi there',
            'persona_assistant': 'I am helpful',
            'persona_user': 'I am curious',
            'top_k_tokens': ['hi'],
            'top_k_features': ['persona_feat']
        }
    ],
    'metadata': {
        'strategy': 'user_persona'
    }
}

data_ssc = {
    'results': [
        {
            'user_prompt': 'Generate text',
            'model_response': 'Generated response',
            'constraint': 'must be positive',
            'top_k_tokens': ['generated'],
            'top_k_features': ['constraint_feat']
        }
    ],
    'metadata': {
        'strategy': 'ssc'
    }
}

ARGS = {
    'data': data_standard
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 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
extract_residual_similarities Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/elicitation_methods/residual_tokens.py
code served (permissive licence) · get_code("9bc9d942b59f5d51")
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 2, dtype family float, shape (4, 16) 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
import numpy as np

torch.manual_seed(0)

class StandInLayer(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 StandInModel(nn.Module):
    def __init__(self, vocab_size=16, hidden_size=8, num_layers=3):
        super().__init__()
        self.device = torch.device('cpu')
        self.model = nn.Module()
        self.model.embed_tokens = nn.Embedding(vocab_size, hidden_size)
        self.model.layers = nn.ModuleList([StandInLayer(hidden_size) for _ in range(num_layers)])
        self.hidden_size = hidden_size
        
    def forward(self, input_ids):
        x = self.model.embed_tokens(input_ids)
        for layer in self.model.layers:
            x = layer(x)
        return (x,)

model = StandInModel(vocab_size=16, hidden_size=8, num_layers=3)
model.eval()

tokens = torch.tensor([1, 2, 3, 4], dtype=torch.long)
target_layer = 1

ARGS = {
    'model': model,
    'tokens': tokens,
    'target_layer': target_layer
}
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 ndarray, ndim 2, dtype family float, shape (4, 16) — shape (4, 16) 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
extract_ssc_positions_for_iteration Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/guess_secret_side_constraint_iterative.py
code served (permissive licence) · get_code("69069a9ed77cfa7d")
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

torch.manual_seed(0)

# Construct a valid pair dict with SSC data
pair = {
    'predictions': [
        {
            'ssc_token_index': 0,
            'top_k': [0.9, 0.8, 0.7],
            'scores': [1.0, 0.9, 0.8]
        },
        {
            'ssc_token_index': 1,
            'top_k_features': [[0.1, 0.2], [0.3, 0.4]],
            'feature_names': ['feat1', 'feat2']
        },
        {
            'ssc_token_index': 0,
            'top_k_features': [[0.5, 0.6], [0.7, 0.8]],
            'feature_names': ['feat1', 'feat2']
        },
        {
            'ssc_token_index': 2,
            'top_k': [0.95, 0.85],
            'scores': [1.0, 0.95]
        }
    ]
}

position = 0

ARGS = {
    'pair': pair,
    'position': position
}
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
fetch_sae_feature_top_tokens Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/guess_secret_side_constraint.py
code served (permissive licence) · get_code("11cedf0738cf7a8b")
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 torch
import numpy as np

# Set up the global state that the function expects
_feature_tokens_data = {
    '0': {
        'top_tokens': ['hello', 'world', 'test', 'feature', 'token']
    },
    '1': {
        'top_tokens': ['python', 'code', 'example']
    },
    '2': ['token1', 'token2', 'token3', 'token4']
}

# Make it globally accessible
import sys
current_module = sys.modules[__name__]
current_module._feature_tokens_data = _feature_tokens_data

ARGS = {
    'feature_index': 0,
    'num_tokens': 3
}
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 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
get_top_k_similar_tokens Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/elicitation_methods/residual_tokens.py
code served (permissive licence) · get_code("a8eb3ad40b635602")
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

class SimpleTokenizer:
    """Minimal tokenizer that implements decode method."""
    def decode(self, token_ids):
        # Simple stand-in: map token IDs to token strings
        token_map = {
            0: "hello",
            1: "world",
            2: "test",
            3: "token",
            4: "sample",
            5: "code",
            6: "python",
            7: "function",
        }
        return " ".join([token_map.get(tid, f"tok_{tid}") for tid in token_ids])

# Create a small similarities array
np.random.seed(0)
similarities = np.random.randn(8) * 0.5 + 0.5  # 8 token similarity scores

# Create a simple tokenizer
tokenizer = SimpleTokenizer()

# Set k to retrieve top-5 similar tokens
k = 5

ARGS = {
    "similarities": similarities,
    "tokenizer": tokenizer,
    "k": k,
}
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_top_k_tokens Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/elicitation_methods/logit_lens.py
code served (permissive licence) · get_code("9fa782ec202e4857")
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

class SimpleTokenizer:
    """Minimal tokenizer that implements decode method."""
    def decode(self, token_ids):
        """Decode token IDs to strings."""
        return "".join([f"token_{tid}" for tid in token_ids])

# Create probability distribution over vocabulary
probabilities = np.array([0.01, 0.05, 0.15, 0.25, 0.20, 0.18, 0.10, 0.06])

tokenizer = SimpleTokenizer()
k = 3

ARGS = {
    'probabilities': probabilities,
    'tokenizer': tokenizer,
    'k': k
}
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_top_k_tokens_ssc Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/elicitation_methods/residual_tokens.py
code served (permissive licence) · get_code("102d0f3e451d4908")
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

class SimpleTokenizer:
    """Minimal tokenizer that decodes token IDs to strings."""
    def decode(self, token_ids):
        return "".join([f"token_{tid}" for tid in token_ids])

probabilities = np.array([0.05, 0.15, 0.3, 0.5])
tokenizer = SimpleTokenizer()
k = 2

ARGS = {
    "probabilities": probabilities,
    "tokenizer": tokenizer,
    "k": k,
}
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
group_entries_by_key Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/calculate_auditing_metrics_gpt.py
code served (permissive licence) · get_code("087ae9d981b6505c")
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
from collections import defaultdict
from typing import Dict, List, Tuple

# Construct valid arguments for group_entries_by_key

# Create test entries with user_prompt and base64_constraint fields
entries_basic = [
    {
        'user_prompt': 'What is AI?',
        'base64_constraint': 'QUk9QXJ0aWZpY2lhbEludGVsbGlnZW5jZQ==',
        'data': 'some_data_1'
    },
    {
        'user_prompt': 'What is AI?',
        'base64_constraint': 'QUk9QXJ0aWZpY2lhbEludGVsbGlnZW5jZQ==',
        'data': 'some_data_2'
    },
    {
        'user_prompt': 'Tell me a joke',
        'base64_constraint': 'Sm9rZQ==',
        'data': 'some_data_3'
    },
]

# Create test entries with persona fields
entries_with_personas = [
    {
        'persona_assistant': 'helpful_bot',
        'persona_user': 'curious_student',
        'data': 'persona_data_1'
    },
    {
        'persona_assistant': 'helpful_bot',
        'persona_user': 'curious_student',
        'data': 'persona_data_2'
    },
    {
        'persona_assistant': 'expert_assistant',
        'persona_user': 'professional',
        'data': 'persona_data_3'
    },
]

# Test case 1: force_position_grouping=True (default, groups by user_prompt and base64_constraint)
ARGS = {
    'entries': entries_basic,
    'force_position_grouping': True
}
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
group_entries_by_prefill Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/calculate_auditing_metrics_gpt.py
code served (permissive licence) · get_code("1253fb9a528e3524")
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 collections import defaultdict
from typing import Dict, List

# Construct a list of entry dictionaries with various prefill values
entries = [
    {
        'prefill': 'prompt_1',
        'data': 'sample_data_1',
        'id': 1
    },
    {
        'prefill': 'prompt_2',
        'data': 'sample_data_2',
        'id': 2
    },
    {
        'prefill': 'prompt_1',
        'data': 'sample_data_3',
        'id': 3
    },
    {
        'prefill': '',
        'data': 'sample_data_4',
        'id': 4
    },
    {
        'data': 'sample_data_5',
        'id': 5
    }
]

ARGS = {
    'entries': entries
}
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_conversation_from_file Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/sampling/batch_inference.py
code served (permissive licence) · get_code("9c716e534c80d0fd")
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 tempfile
import os

# Create a temporary file with conversation content
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt', encoding='utf-8')
temp_file.write("User: Hello, how are you?\n")
temp_file.write("Assistant: I'm doing well, thank you for asking.\n")
temp_file.write("This is a continuation of the assistant response.\n")
temp_file.write("User: What's the weather like?\n")
temp_file.write("Assistant: I don't have access to real-time weather data.\n")
temp_file.close()

file_path = temp_file.name

ARGS = {
    'file_path': file_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 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_densities_from_file Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/elicitation_methods/sae.py
code served (permissive licence) · get_code("03ce8b49985247f6")
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 JSONL file with sample feature density data
temp_file = tempfile.NamedTemporaryFile(mode='w', suffix='.jsonl', delete=False, encoding='utf-8')
temp_file_path = temp_file.name

# Write sample feature density data to the file
sample_features = [
    {'index': '0', 'frac_nonzero': 0.05},
    {'index': '1', 'frac_nonzero': 0.15},
    {'index': '2', 'frac_nonzero': 0.25},
    {'index': '3', 'frac_nonzero': 0.10},
    {'index': '4', 'frac_nonzero': 0.30},
]

for feature in sample_features:
    temp_file.write(json.dumps(feature) + '\n')

temp_file.close()

# Define ARGS with valid arguments for the function
ARGS = {
    'features_file': temp_file_path,
    'layer': 32,
    'width_k': 131,
}
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_density_tensor Ran
ran, honoured the contract we drafted
cywinski/eliciting-secret-knowledge/elicitation_methods/sae.py
code served (permissive licence) · get_code("1bc9d55c3d545619")
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 json
import tempfile
import torch
import os

torch.manual_seed(0)

# Create a temporary features file with sample density data
temp_dir = tempfile.mkdtemp()
features_file = os.path.join(temp_dir, "features.jsonl")

# Write sample feature density data to the file
with open(features_file, 'w', encoding='utf-8') as f:
    # Write a few sample features with frac_nonzero values
    for i in range(100):
        data = {
            'index': i,
            'frac_nonzero': 0.1 + (i % 10) * 0.05
        }
        f.write(json.dumps(data) + '\n')

ARGS = {
    'features_file': features_file,
    'layer': 32,
    'width_k': 131,
    'device': 'cpu'
}
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 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
load_feature_tokens_file Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/guess_secret_side_constraint.py
code served (permissive licence) · get_code("4e066bd9279122ee")
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 sample feature tokens data
temp_dir = tempfile.gettempdir()
temp_file = os.path.join(temp_dir, "feature_tokens_test.json")

sample_data = {
    "feature_0": ["token_a", "token_b", "token_c"],
    "feature_1": ["token_d", "token_e"],
    "feature_2": ["token_f", "token_g", "token_h", "token_i"]
}

with open(temp_file, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

ARGS = {
    "file_path": temp_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_lines_from_file Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/sampling/batch_inference.py
code served (permissive licence) · get_code("e66de9305fea6490")
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 tempfile
import os

# Create a temporary text file with some sample lines
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt', encoding='utf-8')
temp_file.write("line one\n")
temp_file.write("line two\n")
temp_file.write("line three\n")
temp_file.write("\n")  # empty line to test stripping
temp_file.write("line four\n")
temp_file.close()

ARGS = {
    'file_path': temp_file.name,
    'description': 'test'
}
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_prompt_template Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/utils/guess_secret_word.py
code served (permissive licence) · get_code("7b71485921e1d112")
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 tempfile
import os

# Create a temporary prompt template file
temp_dir = tempfile.mkdtemp()
template_file = os.path.join(temp_dir, "prompt_template.txt")

# Write a valid prompt template to the temporary file
with open(template_file, 'w', encoding='utf-8') as f:
    f.write("You are a helpful assistant.\nRespond to: {query}")

# Define ARGS with the path to the valid template file
ARGS = {
    "template_path": template_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 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
load_results_file Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/utils/guess_secret_word.py
code served (permissive licence) · get_code("2557ed63984259cf")
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 valid results data
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "results.json")

results_data = {
    "accuracy": 0.95,
    "loss": 0.05,
    "epochs": 10,
    "model": "test_model",
    "metrics": {
        "precision": 0.93,
        "recall": 0.97
    }
}

with open(temp_file, 'w', encoding='utf-8') as f:
    json.dump(results_data, f)

ARGS = {
    "file_path": temp_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
parse_auditor_response Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/ssc/guess_secret_side_constraint_iterative.py
code served (permissive licence) · get_code("c59f849fa81317ce")
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 torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "response_text": "1. hello\n2. world\n3. test\n4. python"
}
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 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
tfidf_feature_score Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/elicitation_methods/sae.py
code served (permissive licence) · get_code("43596bf917ab9a02")
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)

# Create small tensors for the function
n_tokens = 2
n_features = 4

# activations: [n_tokens, n_features] tensor of SAE activations
activations = torch.randn(n_tokens, n_features)

# feature_density: [n_features] tensor of feature densities (fraction of tokens where feature activates)
# Values should be between 0 and 1 representing density
feature_density = torch.rand(n_features)

ARGS = {
    "activations": activations,
    "feature_density": feature_density,
}
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
validate_and_determine_strategy Ran
ran; our contract draft was wrong, not the code
cywinski/eliciting-secret-knowledge/sampling/batch_inference.py
code served (permissive licence) · get_code("3146c38bfcad5408")
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 sys
from argparse import Namespace

# Create a minimal args object with all required attributes
ARGS = {
    'args': Namespace(
        # Primary strategy flags - only one can be True
        prefill=False,
        multi_turn_jailbreak=False,
        user_persona=False,
        
        # SSC-related
        ssc=False,
        constraints_file=None,
        
        # Prefill strategy
        prefill_file=None,
        
        # Multi-turn jailbreak strategy
        prefix_prompt_file=None,
        prompts_file='/tmp/prompts.txt',
        
        # User persona strategy
        persona_assistant_file=None,
        persona_user_file=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 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
extract_layer_probabilities Not yet run
unverified (harvested, no recorded run)
cywinski/eliciting-secret-knowledge/elicitation_methods/logit_lens.py
code served (permissive licence) · get_code("0f440ae1bd737171")
How it ran

no invocation recorded.

get_inference_engine Not yet run
unverified (harvested, no recorded run)
cywinski/eliciting-secret-knowledge/sampling/inference_engine.py
code served (permissive licence) · get_code("39dc68013194e9a1")
How it ran

no invocation recorded.

get_prompt_preparer Not yet run
unverified (harvested, no recorded run)
cywinski/eliciting-secret-knowledge/sampling/prompt_preparers.py
code served (permissive licence) · get_code("2a1f9a63b1a11968")
How it ran

no invocation recorded.

load_model_and_tokenizer Not yet run
unverified (harvested, no recorded run)
cywinski/eliciting-secret-knowledge/sampling/utils.py
code served (permissive licence) · get_code("7c099df440884000")
How it ran

no invocation recorded.

Also present in other papers' code

Of 25 samples lifted from this paper's linked repositories, 25 carry a function identity and 0 do not. Of the 25 distinct functions with an identity, 1 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 1 later) and 24 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

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

FunctionAlso present inEarliest-datedOthers named
parse_auditor_response 1 other paper
1 later
2602.08877 (2026-02)
PROMPT-LEVEL STRATEGIC DECEPTION

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

Large language model (LLM) activations are notoriously difficult to understand, with most existing techniques using complex, specialized methods for interpreting them. Recent work has proposed a simpler approach known as LatentQA: training LLMs to directly accept LLM activations as inputs and answer arbitrary questions about them in natural language. However, prior work has focused on narrow task settings for both training and evaluation. In this paper, we instead take a generalist perspective. We evaluate LatentQA-trained models, which we call Activation Oracles (AOs), in far out-of-distribution settings and examine how performance scales with training data diversity. We find that AOs can recover information finetuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model. Our main evaluations are four downstream tasks where we can compare to prior white-and black-box techniques. We find that even narrowlytrained LatentQA models can generalize well, and that adding additional training datasets (such as classification tasks and a self-supervised context prediction task) yields consistent further improvements. Our best AOs match or exceed white-box baselines on all four tasks and the best overall baseline on 3 of 4. These results suggest that diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations.

For agents

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

get_harvested_code_for_paper("2512.15674")
get_code_for_paper("2512.15674")
have("2512.15674")

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

curl https://syntology.ai/api/ran/2512.15674.json

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

[![Syntology run record](https://syntology.ai/api/ran/2512.15674.svg)](https://syntology.ai/paper/2512.15674)

Connect an agent — have() is free.