SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.23940 · 2026

Published at ICLR 2026 Workshop on Reasoning and Planning for LLMs RESIDUAL DRIFT DOMINATES CONTRADICTION IN MULTI-TURN CONSTRAINT REASONING

Sebastien Kawada

arXiv · PDF · Open in the Atlas

Code that ran

17 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; 2 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 14 failures of our instrument; 2 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/2605.23940.json · badge: /api/ran/2605.23940.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
kaons-research/drift-bench canonical (classifier label)
classifier basis: census
17 ran · 2 no recorded run
FunctionStatusWhere it lives
build_answer_retry_messages Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/prompts.py
code served (permissive licence) · get_code("6b3409ba60d6f4b6")
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

# Build setup for build_answer_retry_messages function

# Define test data
prior_assistant_response = "Alice sits in seat 1, Bob in seat 2, Charlie in seat 3."

domain = "seating"

entities = ["Alice", "Bob", "Charlie"]

context = {
    "categories": {
        "color": ["red", "blue", "green"],
        "animal": ["cat", "dog", "bird"],
        "job": ["doctor", "lawyer", "engineer"]
    }
}

ARGS = {
    "prior_assistant_response": prior_assistant_response,
    "domain": domain,
    "entities": entities,
    "context": context
}
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
build_truncation_retry_messages Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/prompts.py
code served (permissive licence) · get_code("a9dc69f03fd54a01")
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

# Build base_messages: a list of dicts with 'role' and 'content' keys
base_messages = [
    {'role': 'user', 'content': 'What is the capital of France?'},
    {'role': 'assistant', 'content': 'The capital of France is Paris.'},
    {'role': 'user', 'content': 'Tell me more about it.'}
]

# Domain parameter: choose a valid domain
domain = 'logic_grid'

# Entities parameter: list of strings representing entities
entities = ['Alice', 'Bob', 'Charlie', 'Diana']

# Context parameter: dict with arbitrary string keys and values
context = {
    'categories': {
        'color': ['red', 'blue', 'green'],
        'age': ['young', 'old'],
        'job': ['teacher', 'doctor']
    },
    'clues': ['Alice is not red', 'Bob likes blue']
}

ARGS = {
    'base_messages': base_messages,
    'domain': domain,
    'entities': entities,
    'context': context
}
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
build_turn_messages Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/prompts.py
code served (permissive licence) · get_code("e08375efdbb3fce2")
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

ARGS = {
    'method': 'ledger_only',
    'conversation_history': [
        {'role': 'user', 'content': 'Alice must sit in seat 1.'},
        {'role': 'assistant', 'content': '{"solution": {"Alice": 1, "Bob": 2, "Charlie": 3}}'}
    ],
    'new_user_message': 'Bob cannot sit adjacent to Alice.',
    'domain': 'seating',
    'entities': ['Alice', 'Bob', 'Charlie'],
    'context': {
        'table_shape': 'round',
        'num_entities': 3
    },
    'ledger': [
        {'entity': 'Alice', 'constraint': 'position', 'value': 1}
    ],
    'repair_signal': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_user_message Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/generate_problems.py
code served (permissive licence) · get_code("87aa5907f0cb48ee")
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
from typing import Any

# Construct a list of constraint dictionaries
# Each constraint dict needs an 'nl' key with a string value
constraints = [
    {
        'nl': 'The output should be valid JSON'
    },
    {
        'nl': 'The response must be under 100 tokens'
    },
    {
        'nl': 'All numbers should be positive integers.'
    }
]

ARGS = {
    'constraints': constraints
}
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
canonicalize_constraint Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/z3_checker.py
code served (permissive licence) · get_code("421bced5ee970b0e")
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

# Construct a valid constraint dictionary argument
constraint = {
    'type': 'adjacent',
    'args': ['entity_1', 'entity_2', 'extra_arg']
}

ARGS = {
    'constraint': constraint
}
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
estimate_tokens_from_text Ran
ran, honoured the contract we drafted
kaons-research/drift-bench/src/run_experiment.py
code served (permissive licence) · get_code("a23de49f4a4e2ca2")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# The function only uses len() and basic arithmetic on a string
# So we just need to provide a text string

text = "This is a sample text to estimate tokens from."

ARGS = {
    "text": text
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
extract_json_block Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/extraction.py
code served (permissive licence) · get_code("5727f4a3df316a5a")
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 re
from typing import Any

ARGS = {
    "text": """
    <think>
    Let me analyze this JSON extraction problem.
    </think>
    {
        "name": "test",
        "values": [1, 2, 3],
        "nested": {
            "key": "value"
        }
    }
    """
}
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
ledger_token_count Ran
ran, honoured the contract we drafted
kaons-research/drift-bench/src/run_experiment.py
code served (permissive licence) · get_code("fa850eb8d5f90663")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
from typing import Any

def estimate_tokens_from_text(text: str) -> int:
    return max(1, len(text) // 4)

def ledger_token_count(ledger: list[dict[str, Any]]) -> int:
    if not ledger:
        return 0
    payload = json.dumps({'constraints': ledger}, ensure_ascii=True)
    return estimate_tokens_from_text(payload)

# Construct a valid ledger argument
# The function uses ledger as a list of dicts that gets JSON serialized
ledger_value = [
    {
        "constraint_id": 1,
        "type": "budget",
        "value": 100
    },
    {
        "constraint_id": 2,
        "type": "deadline",
        "value": "2024-12-31"
    },
    {
        "constraint_id": 3,
        "parameters": [1, 2, 3],
        "metadata": {"source": "test"}
    }
]

ARGS = {
    "ledger": ledger_value
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_problem_gold Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/analyze.py
code served (permissive licence) · get_code("e1dd1a7c6320d923")
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 json
import tempfile
from pathlib import Path

# Create a temporary directory with test problem files
temp_dir = tempfile.TemporaryDirectory()
problems_dir = Path(temp_dir.name)

# Create first problem file
problem1 = {
    "problem_id": "prob_001",
    "split": "train",
    "turns": [
        {
            "turn_number": "1",
            "new_constraints": [
                {"constraint_type": "linear", "value": 5}
            ]
        },
        {
            "turn_number": "2",
            "new_constraints": [
                {"constraint_type": "nonlinear", "value": 10}
            ]
        }
    ]
}

# Create second problem file
problem2 = {
    "problem_id": "prob_002",
    "split": "test",
    "turns": [
        {
            "turn_number": "1",
            "new_constraints": []
        }
    ]
}

# Write problem files
with open(problems_dir / "problem_001.json", "w") as f:
    json.dump(problem1, f)

with open(problems_dir / "problem_002.json", "w") as f:
    json.dump(problem2, f)

ARGS = {
    "problems_dir": problems_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 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
load_problems Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/run_experiment.py
code served (permissive licence) · get_code("4737d24ad6135c21")
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 tempfile
from pathlib import Path
from typing import Any

# Create temporary directory structure with test JSON files
temp_dir = tempfile.TemporaryDirectory()
problems_dir = Path(temp_dir.name)

# Create subdirectory for the split
split_name = "train"
split_dir = problems_dir / split_name
split_dir.mkdir(parents=True, exist_ok=True)

# Create test JSON files
test_problem_1 = {
    "split": "train",
    "id": "problem_1",
    "data": "test data 1"
}

test_problem_2 = {
    "split": "train",
    "id": "problem_2",
    "data": "test data 2"
}

test_problem_wrong_split = {
    "split": "test",
    "id": "problem_3",
    "data": "wrong split"
}

# Write JSON files to the split directory
with open(split_dir / "problem_1.json", "w", encoding="utf-8") as f:
    json.dump(test_problem_1, f)

with open(split_dir / "problem_2.json", "w", encoding="utf-8") as f:
    json.dump(test_problem_2, f)

with open(split_dir / "problem_3.json", "w", encoding="utf-8") as f:
    json.dump(test_problem_wrong_split, f)

ARGS = {
    "problems_dir": problems_dir,
    "split": "train"
}
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_rows Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/analyze.py
code served (permissive licence) · get_code("9f1ee38270e3cbda")
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 sqlite3
import tempfile

# Create a temporary in-memory SQLite database
conn = sqlite3.connect(':memory:')

# Create the results table with the required columns
conn.execute('''
    CREATE TABLE results (
        model TEXT,
        method TEXT,
        domain TEXT,
        problem_id TEXT,
        turn_number INTEGER,
        answer_correct INTEGER,
        extraction_null INTEGER,
        ledger_json TEXT,
        constraint_extraction_success INTEGER,
        z3_sat INTEGER,
        mus_size INTEGER,
        mus_json TEXT,
        repair_attempted INTEGER,
        repair_trigger TEXT,
        repair_trigger_count INTEGER,
        repair_reason_json TEXT
    )
''')

# Insert some sample rows
conn.execute('''
    INSERT INTO results VALUES 
    ('model1', 'method1', 'domain1', 'prob1', 1, 1, 0, '{}', 1, 1, 5, '[]', 1, 'trigger1', 2, '{}'),
    ('model1', 'method1', 'domain1', 'prob2', 1, 0, 1, '{}', 0, 0, 0, '[]', 0, NULL, NULL, NULL),
    ('model2', 'method2', 'domain2', 'prob3', 2, 1, 0, '{}', 1, 1, 3, '[]', 1, 'trigger2', 1, '{}')
''')
conn.commit()

# Create a problems_split_map
problems_split_map = {
    'prob1': 'train',
    'prob2': 'val',
    'prob3': 'train'
}

# Define the ARGS dictionary with all required parameters
ARGS = {
    'conn': conn,
    'split': 'all',
    'problems_split_map': problems_split_map
}
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
ordered_unique Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/repair_policy.py
code served (permissive licence) · get_code("8bead903b69ffed0")
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 List

# The function takes a list of strings and returns a list of strings
# It filters out empty strings and duplicates, preserving order

ARGS = {
    "values": ["apple", "banana", "apple", "", "cherry", "banana", "date"]
}
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
pick_distinct_pair Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/generate_problems.py
code served (permissive licence) · get_code("2added5f4145a1d2")
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 random

ARGS = {
    "rng": random.Random(42),
    "items": ["apple", "banana", "cherry", "date", "elderberry"]
}
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 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
policy_uses_sat_triggers Ran
ran, violated the contract we drafted
kaons-research/drift-bench/src/repair_policy.py
code served (permissive licence) · get_code("1a41a1bb65acebd1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "policy": "expanded"
}
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 bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sample_seating_constraint Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/generate_problems.py
code served (permissive licence) · get_code("d33b30864c53fb64")
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 random

def pick_distinct_pair(rng: random.Random, items: list[str]) -> tuple[str, str]:
    a, b = rng.sample(items, 2)
    return (a, b)

def sample_seating_constraint(rng: random.Random, entities: list[str], num_entities: int, table_shape: str) -> dict:
    kinds = ['adjacent', 'not_adjacent', 'at_position', 'left_of', 'separated_by']
    if table_shape == 'rectangular':
        kinds.extend(['same_side', 'opposite_side'])
    ctype = rng.choice(kinds)
    if ctype == 'adjacent':
        a, b = pick_distinct_pair(rng, entities)
        return {'type': ctype, 'args': [a, b], 'nl': f'{a} must sit next to {b}'}
    if ctype == 'not_adjacent':
        a, b = pick_distinct_pair(rng, entities)
        return {'type': ctype, 'args': [a, b], 'nl': f'{a} cannot sit next to {b}'}
    if ctype == 'at_position':
        a = rng.choice(entities)
        pos = rng.randint(1, num_entities)
        return {'type': ctype, 'args': [a, pos], 'nl': f'{a} must sit at position {pos}'}
    if ctype == 'left_of':
        a, b = pick_distinct_pair(rng, entities)
        return {'type': ctype, 'args': [a, b], 'nl': f'{a} sits immediately left of {b}'}
    if ctype == 'separated_by':
        a, b = pick_distinct_pair(rng, entities)
        n_sep = rng.randint(1, max(1, num_entities // 2 - 1))
        return {'type': ctype, 'args': [a, b, n_sep], 'nl': f'{a} and {b} must have at least {n_sep} seats between them'}
    if ctype == 'same_side':
        a, b = pick_distinct_pair(rng, entities)
        return {'type': ctype, 'args': [a, b], 'nl': f'{a} and {b} must sit on the same side of the table'}
    if ctype == 'opposite_side':
        a, b = pick_distinct_pair(rng, entities)
        return {'type': ctype, 'args': [a, b], 'nl': f'{a} and {b} must sit on opposite sides of the table'}
    raise ValueError(f'Unsupported seating constraint type: {ctype}')

ARGS = {
    'rng': random.Random(42),
    'entities': ['Alice', 'Bob', 'Charlie', 'Diana', 'Eve'],
    'num_entities': 5,
    'table_shape': 'rectangular'
}
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
table_columns Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/analyze.py
code served (permissive licence) · get_code("82a316cc22bc1015")
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 set. 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 sqlite3
import tempfile
import os

# Create a temporary database file
temp_db = tempfile.NamedTemporaryFile(delete=False, suffix='.db')
temp_db_path = temp_db.name
temp_db.close()

try:
    # Create a connection and a test table
    conn = sqlite3.connect(temp_db_path)
    conn.execute('''
        CREATE TABLE test_table (
            id INTEGER PRIMARY KEY,
            name TEXT,
            age INTEGER,
            email TEXT
        )
    ''')
    conn.commit()
    
    # Define the arguments
    ARGS = {
        'conn': conn,
        'table': 'test_table'
    }
finally:
    # Clean up will happen after function execution
    pass
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 set — a set output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a set 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
unique_trigger_codes Ran
ran; our contract draft was wrong, not the code
kaons-research/drift-bench/src/repair_policy.py
code served (permissive licence) · get_code("26f1b08bdc9578f4")
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

# Construct arguments for unique_trigger_codes function
# The function expects: reasons: list[dict[str, Any]]

ARGS = {
    "reasons": [
        {"code": "ERROR_001"},
        {"code": "  WARNING_002  "},
        {"code": ""},
        {"code": "ERROR_001"},  # duplicate, should be filtered
        {"other_field": "value"},  # missing 'code' key
        {"code": None},
        {"code": "INFO_003"},
    ]
}
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_answer Not yet run
unverified (harvested, no recorded run)
kaons-research/drift-bench/src/extraction.py
code served (permissive licence) · get_code("2859b0d4274869b4")
How it ran

no invocation recorded.

extract_answer_with_retry Not yet run
unverified (harvested, no recorded run)
kaons-research/drift-bench/src/extraction.py
code served (permissive licence) · get_code("94d9078851c10b86")
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

How do multi-turn reasoning systems fail? The expected answer is logical contradiction, in which the system's maintained state becomes unsatisfiable. We show that the dominant mode is instead satisfiable drift, where the internal state stays consistent while the returned answer silently violates prior commitments. We build DRIFT-BENCH (Decomposing Reasoning Into Failure Types), a solver-instrumented benchmark of 816 test problems across three constraint domains, and evaluate four methods on it across four open-weight models (8B-120B parameters). MUS-Repair, which feeds minimal unsatisfiable subsets back to the generator, is strongest in every setting (+1.8 to +15.0 pp over the best non-MUS baseline). But the central finding is what repair leaves behind. After structured feedback, models rarely contradict themselves. They forget. Residual errors are 98-100% satisfiable drift across all settings, while contradiction drops to near zero. Reliable multi-turn systems must separately validate that the returned answer respects the maintained state. Code is available at https://github.com/kaons-research/drift-bench.

For agents

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

get_harvested_code_for_paper("2605.23940")
get_code_for_paper("2605.23940")
have("2605.23940")

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

curl https://syntology.ai/api/ran/2605.23940.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.23940.svg)](https://syntology.ai/paper/2605.23940)

Connect an agent — have() is free.