SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.28565 · 2026

Verified Misguidance: Measuring Structural Citation Failures in Search-Augmented LLMs

Dongha Lee, Hyeonseo Jang, Yongsik Seo, Wooseok Jeong, Eunyoung Kim

arXiv · PDF

Code that ran

15 of 20 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 2 honoured the contract we drafted, 2 violated it and 1 has no contract verdict; 10 failures of our instrument; 5 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/2605.28565.json · badge: /api/ran/2605.28565.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
oseoko/verified-misguidance community
Papers with Code archive
15 ran · 5 no recorded run
FunctionStatusWhere it lives
Citation Ran
ran on a synthesized input · constructed an object; no result was computed
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("c5ef2eac1bb9b42a")
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 an instance of its own class, Citation: the run constructed the object and did not call it, so no result was computed. 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
from typing import NamedTuple

class Citation(NamedTuple):
    url: str
    cited_sentence: str
    order: int

ARGS = {
    "url": "https://example.com/article",
    "cited_sentence": "This is a cited sentence from the source material.",
    "order": 1
}
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 Citation — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Citation output the recorded ndim 1 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_clean Ran
ran; our contract draft was wrong, not the code
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("271c479e7c7821ac")
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 re

_CHROME_CHARS = ' \t\n*"\'`()""'''

def _clean(s: str) -> str:
    s = s.strip(_CHROME_CHARS)
    while True:
        m = re.match('^\\s*\\[\\d+\\]\\s*', s)
        if not m:
            break
        s = s[m.end():]
    return s.strip(_CHROME_CHARS)

# The function takes a single parameter 's' which is a string
# It strips chrome characters, removes leading [digit] patterns, and strips again

ARGS = {
    's': '  [123]  hello world  \n'
}
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_last_sentences Ran
ran; our contract draft was wrong, not the code
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("e8489362fb8b5be0")
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 re

_ABBREVIATIONS = ('e.g.', 'i.e.', 'etc.', 'Dr.', 'Mr.', 'Mrs.', 'Ms.', 'U.S.', 'U.K.', 'vs.')

_CHROME_CHARS = ' \t\n*"\'`()""'''

_NUMBERED_LIST_RE = re.compile('\\d+\\.\\s')

_SENTENCE_COUNT = 2

_URL_RE = re.compile('https?://[^\\s)\\]]+')

def _clean(s: str) -> str:
    s = s.strip(_CHROME_CHARS)
    while True:
        m = re.match('^\\s*\\[\\d+\\]\\s*', s)
        if not m:
            break
        s = s[m.end():]
    return s.strip(_CHROME_CHARS)

def _in_spans(pos: int, spans: list[tuple[int, int]]) -> bool:
    return any((s <= pos < e for s, e in spans))

def _is_abbrev_end(text: str, dot_pos: int) -> bool:
    lower = text.lower()
    for abbr in _ABBREVIATIONS:
        a = abbr.lower()
        for offset in range(len(a)):
            if a[offset] != '.':
                continue
            start = dot_pos - offset
            end = start + len(a)
            if 0 <= start and end <= len(text) and (lower[start:end] == a):
                return True
    return False

def _is_md_boundary(text: str, pos: int) -> bool:
    if pos < 0 or pos >= len(text) or text[pos] != '\n':
        return False
    if pos + 1 < len(text) and text[pos + 1] == '\n':
        return True
    j = pos + 1
    while j < len(text) and text[j] in ' \t':
        j += 1
    rest = text[j:j + 4]
    if rest.startswith('- ') or rest.startswith('* ') or rest.startswith('### '):
        return True
    if _NUMBERED_LIST_RE.match(rest):
        return True
    return False

def _is_sentence_end(text: str, pos: int, url_spans: list[tuple[int, int]]) -> bool:
    if pos < 0 or pos >= len(text):
        return False
    ch = text[pos]
    if ch not in '.!?':
        return False
    if _in_spans(pos, url_spans):
        return False
    if ch == '.' and _is_abbrev_end(text, pos):
        return False
    return True

def _url_spans(text: str) -> list[tuple[int, int]]:
    return [(m.start(), m.end()) for m in _URL_RE.finditer(text)]

def _extract_last_sentences(text: str, n: int=_SENTENCE_COUNT) -> list[str]:
    stripped = text.rstrip()
    while stripped and stripped[-1] in _CHROME_CHARS:
        stripped = stripped[:-1].rstrip()
    if not stripped:
        return []
    url_spans = _url_spans(stripped)
    boundaries: list[int] = [len(stripped)]
    i = len(stripped) - 2
    while i >= 0 and len(boundaries) <= n:
        candidate = None
        if _is_sentence_end(stripped, i, url_spans):
            candidate = i + 1
        elif _is_md_boundary(stripped, i):
            candidate = i + (2 if i + 1 < len(stripped) and stripped[i + 1] == '\n' else 1)
        if candidate is not None:
            boundaries.append(candidate)
            if len(boundaries) > n:
                break
        i -= 1
    if len(boundaries) <= n:
        boundaries.append(0)
    out = []
    for a, b in zip(boundaries[1:], boundaries[:-1]):
        piece = _clean(stripped[a:b])
        if piece:
            out.append(piece)
    out.reverse()
    return out


ARGS = {
    "text": "This is the first sentence. This is the second sentence. This is the third sentence.",
    "n": 2
}
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
_find_citations Ran
ran, honoured the contract we drafted
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("8f88f8548a26337e")
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 re

_CITATION_RE = re.compile('\\[(\\d+)\\]')

def _find_citations(text: str) -> list[tuple[int, int, int]]:
    """(N, start, end) tuples for each ``[N]`` marker."""
    return [(int(m.group(1)), m.start(), m.end()) for m in _CITATION_RE.finditer(text)]

ARGS = {
    "text": "This is a reference [1] and another [42] with text in between [7] and more text."
}
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 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
_in_spans Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("1572b4ec6d8b3c2a")
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

def _in_spans(pos: int, spans: list[tuple[int, int]]) -> bool:
    return any((s <= pos < e for s, e in spans))

ARGS = {
    "pos": 5,
    "spans": [(1, 3), (4, 8), (10, 15)]
}
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
_is_abbrev_end Ran
ran, violated the contract we drafted
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("5d601931b54d7a45")
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.

# Analyze the function to understand what arguments are needed
# _is_abbrev_end(text: str, dot_pos: int) -> bool
# - text: a string to search through
# - dot_pos: an integer position in the string

# The function checks if a dot at position dot_pos is part of an abbreviation
# It needs:
# 1. A text string containing some content
# 2. A dot_pos integer that points to a valid position in that string

# Test case 1: Check "Dr." abbreviation
text_1 = "Dr. Smith is here"
dot_pos_1 = 2  # Position of the dot in "Dr."

# Test case 2: Check "U.S." abbreviation
text_2 = "The U.S. is a country"
dot_pos_2 = 6  # Position of the dot in "U.S."

# Test case 3: Regular dot not part of abbreviation
text_3 = "This is a sentence. Another one."
dot_pos_3 = 15  # Position of dot after "sentence"

ARGS = {
    "text": "Dr. Smith works at U.S.A.",
    "dot_pos": 2
}
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 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
_is_md_boundary Ran
ran, violated the contract we drafted
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("1712d836c7143f92")
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 re

_NUMBERED_LIST_RE = re.compile('\\d+\\.\\s')

def _is_md_boundary(text: str, pos: int) -> bool:
    if pos < 0 or pos >= len(text) or text[pos] != '\n':
        return False
    if pos + 1 < len(text) and text[pos + 1] == '\n':
        return True
    j = pos + 1
    while j < len(text) and text[j] in ' \t':
        j += 1
    rest = text[j:j + 4]
    if rest.startswith('- ') or rest.startswith('* ') or rest.startswith('### '):
        return True
    if _NUMBERED_LIST_RE.match(rest):
        return True
    return False


# Test case 1: text with newline at pos, followed by another newline (boundary)
text1 = "hello\n\nworld"
pos1 = 5

# Test case 2: text with newline at pos, followed by list marker
text2 = "hello\n- item"
pos2 = 5

# Test case 3: text with newline at pos, followed by spaces and numbered list
text3 = "hello\n  1. item"
pos3 = 5

# Test case 4: text with valid position but no boundary
text4 = "hello\nworld"
pos4 = 5

ARGS = {
    "text": text1,
    "pos": pos1
}
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 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
_is_sentence_end Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("b2fbd8bbc8166dfb")
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
import numpy as np

# Analyze the function _is_sentence_end to understand what it needs:
# - text: str - a text string
# - pos: int - a position in the text
# - url_spans: list[tuple[int, int]] - list of (start, end) tuples for URL positions

# Create test data
text = "Hello world. This is a test. Visit https://example.com for info. Dr. Smith works here! What?"
pos = 12  # Position of the first period in "world."
url_spans = [(26, 45)]  # Span covering the URL "https://example.com"

ARGS = {
    "text": text,
    "pos": pos,
    "url_spans": url_spans
}
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 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
_join_sentences Ran
ran; our contract draft was wrong, not the code
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("00242b550dae52c8")
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

torch.manual_seed(0)

ARGS = {
    "sents": ["Hello world", "This is a test", "Python is great"]
}
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 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
_replace_gpt_links Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("4645ba7e21bad41d")
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 re

_GPT_LINK_RE = re.compile(r'\[([^\]]+)\]\((https?://[^)]+)\)')

def _replace_gpt_links(text: str, annotations: list[dict]) -> str:
    """OpenAI / xAI markdown links ``[text](url)`` → ``[N]`` numbered."""
    url_to_num: dict[str, int] = {}
    for idx, ann in enumerate(annotations, start=1):
        url = ann.get('url', '')
        if url and url not in url_to_num:
            url_to_num[url] = idx
    next_num = max(url_to_num.values(), default=0)

    def _r(m: re.Match) -> str:
        nonlocal next_num
        url = m.group(2)
        if url not in url_to_num:
            next_num += 1
            url_to_num[url] = next_num
        return f'[{url_to_num[url]}]'
    return (_GPT_LINK_RE.sub(_r, text), url_to_num)


ARGS = {
    'text': 'Check out [this article](https://example.com/article) and [another link](https://example.com/blog)',
    'annotations': [
        {'url': 'https://example.com/article', 'text': 'this article'},
        {'url': 'https://example.com/docs', 'text': 'documentation'},
    ]
}
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
_replace_grok_links Ran
ran; our contract draft was wrong, not the code
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("6b9ec5188dae0d14")
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 re

_GROK_LINK_RE = re.compile('\\[\\[(\\d+)\\]\\]\\((https?://[^)]+)\\)')

def _replace_grok_links(text: str) -> tuple[str, dict[str, int]]:
    url_to_num: dict[str, int] = {}

    def _r(m: re.Match) -> str:
        n = int(m.group(1))
        url_to_num.setdefault(m.group(2), n)
        return f'[{n}]'
    return (_GROK_LINK_RE.sub(_r, text), url_to_num)


# Setup code
text_with_grok_links = "Check this out [[1]](https://example.com) and [[2]](https://another.org) for more info."

ARGS = {
    "text": text_with_grok_links
}
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
_strip_sonar_url_list Ran
ran; our contract draft was wrong, not the code
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("05945f69447a3289")
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 re

def _strip_sonar_url_list(text: str) -> str:
    return re.split('\\n\\s*\\[\\d+\\]\\s+https?://', text, maxsplit=1)[0]

# Setup code to construct valid arguments
text_with_url = "This is some content\n  [1]  https://example.com/page\nMore content after the URL"

ARGS = {
    "text": text_with_url
}
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
_url_spans Ran
ran, honoured the contract we drafted
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("d0fe19e201db3635")
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 re

_URL_RE = re.compile('https?://[^\\s)\\]]+')

def _url_spans(text: str) -> list[tuple[int, int]]:
    return [(m.start(), m.end()) for m in _URL_RE.finditer(text)]

ARGS = {
    'text': 'Check out https://example.com and http://test.org for more info!'
}
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 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
extract_gemini Ran
ran; our contract draft was wrong, not the code
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("546e642176fa9e86")
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 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 re
from typing import NamedTuple

class Citation(NamedTuple):
    url: str
    cited_sentence: str
    order: int

# Construct valid arguments for extract_gemini function
ARGS = {
    'raw': 'This is some raw text that will be ignored.',
    'meta': {
        'grounding_chunks': [
            {
                'uri': 'https://example.com/page1'
            },
            {
                'uri': 'https://example.com/page2'
            },
            {
                'uri': 'https://example.com/page3'
            }
        ],
        'grounding_supports': [
            {
                'text': 'This is the first supporting sentence.',
                'chunk_indices': [0, 1]
            },
            {
                'text': 'This is the second supporting sentence with more details.',
                'chunk_indices': [1, 2]
            },
            {
                'text': 'Another support here.',
                'grounding_chunk_indices': [0]
            }
        ]
    }
}
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 2, 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_openai Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("d93fe0f5924dd593")
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 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 re

ARGS = {
    'raw': 'This is a great resource [1]. The study shows [2] that results are positive. Visit [my link](https://example.com) for more info.',
    'meta': {
        'annotations': [
            {'url': 'https://example.com/paper1'},
            {'url': 'https://example.com/paper2'}
        ]
    }
}
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 2, 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
_rows_from_numbered_text Not yet run
unverified (harvested, no recorded run)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("5393bd7051513ef3")
How it ran

no invocation recorded.

extract Not yet run
unverified (harvested, no recorded run)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("70d67c6d4214e9c2")
How it ran

no invocation recorded.

extract_anthropic Not yet run
unverified (harvested, no recorded run)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("1111af6476477c6a")
How it ran

no invocation recorded.

extract_grok Not yet run
unverified (harvested, no recorded run)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("968d8dc9a367cd43")
How it ran

no invocation recorded.

extract_perplexity Not yet run
unverified (harvested, no recorded run)
oseoko/verified-misguidance/src/search/extractors.py
code served (permissive licence) · get_code("fdd03cb80b0aa173")
How it ran

no invocation recorded.

Also present in other papers' code

Of 20 samples lifted from this paper's linked repositories, 20 carry a function identity and 0 do not. Of the 20 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 20 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

Users of search-augmented LLMs rely on citations as evidence that responses are grounded in real sources, and rarely verify the cited pages themselves. Millions of queries per day now pass through these systems, making citation quality a silent determinant of whether users are informed or misled-yet existing benchmarks each address one facet in isolation, leaving the joint structure that determines citation trustworthiness unmeasured. We construct CITETRACE, a large-scale dataset that traces the full citation chain from user query through retrieved source to generated answer: 11,200 real-world queries from 28 communities paired with 112,000 responses from ten models across five providers, yielding 761,495 evaluable citation pairs. We design a three-dimension evaluation framework that scores each citation on intent-purpose alignment, source suitability, and answer-source fidelity, using expert-validated predefined matrices and a five-level fidelity rubric; the framework applies to any system that produces citation-bearing responses. Applying this framework at scale, we identify a systematic pattern we call VERIFIED MISGUIDANCE (VM): models cite real, accessible sources yet fail along one or more dimensions, producing a fidelity-suitability trade-off in which faithful models select inappropriate sources and vice versa. Across our pool, 30.6% of citations distort their sources and 27.1% originate from domain-inappropriate sources; at the response level, up to 96% of users encounter at least one structurally misleading citation. Provider-level differences explain 88-96% of citation-quality variance, suggesting that source selection is governed more by factors beyond individual model capability than by the LLMs themselves. Together, CITETRACE and its evaluation framework provide the first resource for diagnosing structural citation failures in deployed search-augmented systems.

For agents

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

get_harvested_code_for_paper("2605.28565")
get_code_for_paper("2605.28565")
have("2605.28565")

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

curl https://syntology.ai/api/ran/2605.28565.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.28565.svg)](https://syntology.ai/paper/2605.28565)

Connect an agent — have() is free.