SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.13124 · 2026

ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

Xiang Li, Ming Xu, Jiarui Li, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Qingyu Zhang, Qianhao Yuan, Xiuying Zhao

arXiv · PDF · Open in the Atlas

Code that ran

21 of 24 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, 2 violated it and 0 have no contract verdict; 18 failures of our instrument; 3 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/2607.13124.json · badge: /api/ran/2607.13124.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
VisionOPD/Vision-OPD canonical (classifier label)
classifier basis: census
11 ran · 3 no recorded run
icip-cas/ShortX canonical (classifier label)
classifier basis: census
10 ran · 0 no recorded run
FunctionStatusWhere it lives
acc_text Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/cal_acc.py
code served (permissive licence) · get_code("2fb398873e75361f")
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 = {
    'correct': 42,
    'total': 100
}
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
block_influence Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
icip-cas/ShortX/shortx/shortgpt/metrics.py
code served (permissive licence) · get_code("5949cb967d9e2f26")
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 (6,) 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 block_influence function
# in_hidden: shape (batch, seq_len, hidden_dim)
# out_hidden: shape (batch, seq_len, hidden_dim)
batch_size = 2
seq_len = 3
hidden_dim = 4

in_hidden = torch.randn(batch_size, seq_len, hidden_dim)
out_hidden = torch.randn(batch_size, seq_len, hidden_dim)

ARGS = {
    "in_hidden": in_hidden,
    "out_hidden": out_hidden,
}
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 (6,) — shape (6,) 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_first_option Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/judge_qwenlm.py
code served (permissive licence) · get_code("7375ac1003e75160")
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 extract_first_option(text):
    if not text:
        return ''
    match = re.search('\\(([A-Z])\\)', text)
    if match:
        return match.group(1)
    match = re.search('([A-Z])[\\.\\)\\s]', text)
    if match:
        return match.group(1)
    match = re.search('([A-Z])', text)
    if match:
        return match.group(1)
    return ''

# Test cases to construct valid arguments
ARGS = {
    'text': 'Choose option (A) from the list'
}
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_mcq_option Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/judge_qwenlm.py
code served (permissive licence) · get_code("c08ee5d994b5b602")
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 extract_mcq_option(answer):
    if not isinstance(answer, str) or not answer:
        return ''
    text = answer.strip()
    pattern = '^[ (\\[]*([A-F])(?:(?=$)|[\\.\\)\\]]|(?:[\\:\\-]\\s+))'
    match = re.match(pattern, text)
    if match:
        return match.group(1)
    return ''

# Setup test cases for extract_mcq_option
ARGS = {
    'answer': 'A) This is the correct answer'
}
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
generate_config_from_args Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/verl/model_merger/base_model_merger.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("8896f4e2bafe8abc")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from LeapLabTHU/OPDVR, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

identity_forward Ran
ran, honoured the contract we drafted
icip-cas/ShortX/shortx/utils/patch_strategies.py
code served (permissive licence) · get_code("6650c706f3db2556")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 3, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a minimal layer object
class DummyLayer(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def forward(self, hidden_states):
        return self.linear(hidden_states)

layer = DummyLayer()

# Create tiny hidden states tensor (batch_size=2, hidden_dim=4)
hidden_states = torch.randn(2, 4)

ARGS = {
    "layer": layer,
    "hidden_states": hidden_states,
    "args": (),
    "kwargs": {}
}
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 3, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
is_correct Ran
ran, violated the contract we drafted
VisionOPD/Vision-OPD/eval/cal_acc.py
code served (permissive licence) · get_code("3d659f15a1dfc162")
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

ARGS = {
    'item': {
        'judge': 'yes'
    }
}
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
l2_distance Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
icip-cas/ShortX/shortx/shortgpt/metrics.py
code served (permissive licence) · get_code("1215305d10fb1d13")
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 (6,) 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 in_hidden and out_hidden
# Shape: (batch_size=2, seq_len=3, hidden_dim=4)
in_hidden = torch.randn(2, 3, 4)
out_hidden = torch.randn(2, 3, 4)

ARGS = {
    "in_hidden": in_hidden,
    "out_hidden": out_hidden,
}
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 (6,) — shape (6,) 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
llama_attention_only_forward Ran
ran; our contract draft was wrong, not the code
icip-cas/ShortX/shortx/shortgpt/adapters/llama.py
code served (permissive licence) · get_code("7766ed6fe1ace2d2")
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 4, 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)

class MockAttention(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.hidden_size = hidden_size
    
    def forward(self, hidden_states, attention_mask=None, position_ids=None, 
                past_key_value=None, use_cache=None, cache_position=None, 
                position_embeddings=None, **kwargs):
        batch_size, seq_len, hidden_size = hidden_states.shape
        attn_output = torch.randn(batch_size, seq_len, hidden_size)
        return attn_output, None

class MockLayerNorm(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.bias = nn.Parameter(torch.zeros(hidden_size))
    
    def forward(self, x):
        return x

class MockLlamaLayer(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.input_layernorm = MockLayerNorm(hidden_size)
        self.self_attn = MockAttention(hidden_size)

hidden_size = 4
batch_size = 2
seq_len = 3

layer = MockLlamaLayer(hidden_size)
hidden_states = torch.randn(batch_size, seq_len, hidden_size)
attention_mask = torch.ones(batch_size, seq_len)
position_ids = torch.arange(seq_len).unsqueeze(0).expand(batch_size, -1)
past_key_value = None
use_cache = False
cache_position = None
position_embeddings = torch.randn(seq_len, hidden_size)

ARGS = {
    'layer': layer,
    'hidden_states': hidden_states,
    'attention_mask': attention_mask,
    'position_ids': position_ids,
    'past_key_value': past_key_value,
    'use_cache': use_cache,
    'cache_position': cache_position,
    'position_embeddings': position_embeddings,
    'args': (),
    'kwargs': {}
}
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 4, 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
llama_ffn_only_forward Ran
ran; our contract draft was wrong, not the code
icip-cas/ShortX/shortx/shortgpt/adapters/llama.py
code served (permissive licence) · get_code("2f1ea70ec01e35ff")
How it ran

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

import torch
import torch.nn as nn

torch.manual_seed(0)

class LlamaFFN(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.hidden_size = hidden_size
        self.w1 = nn.Linear(hidden_size, hidden_size * 4)
        self.w2 = nn.Linear(hidden_size * 4, hidden_size)
        self.w3 = nn.Linear(hidden_size, hidden_size * 4)
    
    def forward(self, x):
        return self.w2(nn.functional.silu(self.w1(x)) * self.w3(x))

class LlamaLayer(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.post_attention_layernorm = nn.LayerNorm(hidden_size)
        self.mlp = LlamaFFN(hidden_size)

hidden_size = 4
batch_size = 2
layer = LlamaLayer(hidden_size)
hidden_states = torch.randn(batch_size, hidden_size)

ARGS = {
    "layer": layer,
    "hidden_states": hidden_states,
    "args": (),
    "kwargs": {}
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
llama_last_attn_ffn_only_forward Ran
ran; our contract draft was wrong, not the code
icip-cas/ShortX/shortx/shortgpt/adapters/llama.py
code served (permissive licence) · get_code("51bc546d3017eaa6")
How it ran

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

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build minimal layer components
class MinimalAttention(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.hidden_size = hidden_size
    
    def forward(self, hidden_states, attention_mask=None, position_ids=None, 
                past_key_value=None, use_cache=None, cache_position=None, 
                position_embeddings=None, **kwargs):
        # Return hidden_states unchanged and None for past_key_value
        return hidden_states, None

class MinimalMLP(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.linear = nn.Linear(hidden_size, hidden_size)
    
    def forward(self, hidden_states):
        return self.linear(hidden_states)

class MinimalLlamaLayer(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.hidden_size = hidden_size
        self.input_layernorm = nn.LayerNorm(hidden_size)
        self.post_attention_layernorm = nn.LayerNorm(hidden_size)
        self.self_attn = MinimalAttention(hidden_size)
        self.mlp = MinimalMLP(hidden_size)

# Create the layer instance
hidden_size = 4
seq_length = 3
batch_size = 2

llama_layer = MinimalLlamaLayer(hidden_size)

# Create input tensors
hidden_states = torch.randn(batch_size, seq_length, hidden_size)
attention_mask = torch.ones(batch_size, seq_length, dtype=torch.long)
position_ids = torch.arange(seq_length, dtype=torch.long).unsqueeze(0).expand(batch_size, -1)
past_key_value = None
use_cache = False
cache_position = torch.arange(seq_length, dtype=torch.long)
position_embeddings = torch.randn(seq_length, hidden_size)

ARGS = {
    "self": llama_layer,
    "hidden_states": hidden_states,
    "attention_mask": attention_mask,
    "position_ids": position_ids,
    "past_key_value": past_key_value,
    "use_cache": use_cache,
    "cache_position": cache_position,
    "position_embeddings": position_embeddings,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_sample_uid Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/infer.py
code served (permissive licence) · get_code("16c462aa165ced48")
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 hashlib
import json

def make_sample_uid(item, benchmark):
    for key in ('sample_uid', 'uid', 'index', 'question_id', 'id'):
        value = item.get(key)
        if value is not None and str(value) != '':
            return f'{benchmark}:{key}:{value}'
    stable_obj = {'benchmark': benchmark, 'images': item.get('images') or [], 'query': item.get('query', '')}
    raw = json.dumps(stable_obj, ensure_ascii=False, sort_keys=True)
    return 'sha1:' + hashlib.sha1(raw.encode('utf-8')).hexdigest()


# Test case 1: item with sample_uid
item_with_uid = {
    'sample_uid': '12345',
    'query': 'test query',
    'images': ['img1.jpg', 'img2.jpg']
}

benchmark_name = 'test_benchmark'

ARGS = {
    'item': item_with_uid,
    'benchmark': benchmark_name
}
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
pick_image_path Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/infer.py
code served (permissive licence) · get_code("ebfbc91b7ed393e0")
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)

# The function pick_image_path takes an item parameter
# It calls item.get('images') which expects a dict with a 'get' method
# It returns the first element of the result (or empty string if None)

# Create a valid item dict
item_with_images = {
    'images': ['path/to/image1.jpg', 'path/to/image2.jpg']
}

ARGS = {
    'item': item_with_images
}
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
prepare_zoombench Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/prepare_data.py
code served (permissive licence) · get_code("d9c5a589cd427f50")
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.

from pathlib import Path
import tempfile
from unittest.mock import patch, MagicMock
import pyarrow as pa
import pyarrow.parquet as pq

# Create a temporary directory for the test
temp_dir = tempfile.mkdtemp()
out_dir = Path(temp_dir)

# Create mock data structure that matches what the function expects
mock_data = {
    'id': [1, 2],
    'query': ['What is this?', 'Describe the image'],
    'response': ['This is a test', 'A test image'],
    'image': [
        {'bytes': b'\xff\xd8\xff\xe0\x00\x10JFIF\x00\x01\x01\x00\x00\x01\x00\x01\x00\x00'},  # minimal JPEG header
        {'bytes': b'\xff\xd8\xff\xe0\x00\x10JFIF\x00\x01\x01\x00\x00\x01\x00\x01\x00\x00'}
    ],
    'crop_image': [
        {'bytes': b'\xff\xd8\xff\xe0\x00\x10JFIF\x00\x01\x01\x00\x00\x01\x00\x01\x00\x00'},
        None
    ]
}

# Create a PyArrow table
table = pa.table(mock_data)

# Write the table to a parquet file
local_dir = out_dir / 'ZoomBench_data' / 'data'
local_dir.mkdir(parents=True, exist_ok=True)
parquet_path = local_dir / 'test.parquet'
pq.write_table(table, str(parquet_path))

# Mock the snapshot_download function to avoid network calls
def mock_snapshot_download(repo_id, repo_type=None, local_dir=None):
    return local_dir

ARGS = {
    'out_dir': out_dir
}

# Patch snapshot_download to use our mock
import huggingface_hub
original_snapshot_download = huggingface_hub.snapshot_download
huggingface_hub.snapshot_download = mock_snapshot_download
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
huggingface_hub, pyarrow — 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
process_multi_modal_inputs_for_minicpmo Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/verl/utils/dataset/vision_utils.py
code served (permissive licence) · get_code("ee4a16e546e12473")
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 tiny tensors for the function parameters
batch_size = 2
seq_length = 5
vocab_size = 100

# input_ids: token IDs for the batch
input_ids = torch.randint(0, vocab_size, (batch_size, seq_length))

# attention_mask: binary mask with left padding (argmax will find first 1)
attention_mask = torch.tensor([
    [0, 0, 1, 1, 1],
    [0, 1, 1, 1, 1]
], dtype=torch.long)

# position_ids: positional embeddings
position_ids = torch.arange(seq_length).unsqueeze(0).expand(batch_size, -1)

# cu_seqlens: cumulative sequence lengths for flash attention
cu_seqlens = torch.tensor([0, seq_length, 2 * seq_length], dtype=torch.long)

# multi_modal_inputs: dictionary with image-related data
# image_bound: list of tensors with image token boundaries per sample
image_bound_0 = torch.tensor([[1, 3], [4, 6]], dtype=torch.long)  # 2 images in first sample
image_bound_1 = torch.tensor([[2, 4]], dtype=torch.long)  # 1 image in second sample

# pixel_values: list of lists of image tensors
pixel_values_0 = [torch.randn(3, 8, 8), torch.randn(3, 8, 8)]  # 2 images for first sample
pixel_values_1 = [torch.randn(3, 8, 8)]  # 1 image for second sample

# tgt_sizes: list of target sizes per sample
tgt_sizes_0 = torch.tensor([8, 8], dtype=torch.long)  # size info for first image
tgt_sizes_1 = torch.tensor([8, 8], dtype=torch.long)  # size info for second image

multi_modal_inputs = {
    'image_bound': [image_bound_0, image_bound_1],
    'pixel_values': [pixel_values_0, pixel_values_1],
    'tgt_sizes': [tgt_sizes_0, tgt_sizes_1]
}

ARGS = {
    'input_ids': input_ids,
    'attention_mask': attention_mask,
    'position_ids': position_ids,
    'cu_seqlens': cu_seqlens,
    'multi_modal_inputs': multi_modal_inputs
}
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
qwen_attention_only_forward Ran
ran; our contract draft was wrong, not the code
icip-cas/ShortX/shortx/shortgpt/adapters/qwen.py
code served (permissive licence) · get_code("cf3900721c85d810")
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 3, dtype family float, shape (2, 3, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockLayerNorm(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.bias = nn.Parameter(torch.zeros(hidden_size))
    
    def forward(self, x):
        return x * self.weight + self.bias

class MockSelfAttention(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.hidden_size = hidden_size
    
    def forward(self, hidden_states, attention_mask=None, position_ids=None, 
                past_key_value=None, use_cache=None, cache_position=None, 
                position_embeddings=None, **kwargs):
        batch_size, seq_len, hidden_size = hidden_states.shape
        attn_output = torch.randn(batch_size, seq_len, hidden_size)
        return attn_output, None

class MockQwenLayer(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.input_layernorm = MockLayerNorm(hidden_size)
        self.self_attn = MockSelfAttention(hidden_size)

hidden_size = 4
batch_size = 2
seq_len = 3

layer = MockQwenLayer(hidden_size)
hidden_states = torch.randn(batch_size, seq_len, hidden_size)
attention_mask = torch.ones(batch_size, 1, seq_len, seq_len)
position_ids = torch.arange(seq_len).unsqueeze(0).expand(batch_size, -1)
past_key_value = None
use_cache = False
cache_position = torch.arange(seq_len)
position_embeddings = torch.randn(seq_len, hidden_size)

ARGS = {
    "layer": layer,
    "hidden_states": hidden_states,
    "attention_mask": attention_mask,
    "position_ids": position_ids,
    "past_key_value": past_key_value,
    "use_cache": use_cache,
    "cache_position": cache_position,
    "position_embeddings": position_embeddings,
    "args": (),
    "kwargs": {}
}
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 3, dtype family float, shape (2, 3, 4) — shape (2, 3, 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
qwen_ffn_only_forward Ran
ran; our contract draft was wrong, not the code
icip-cas/ShortX/shortx/shortgpt/adapters/qwen.py
code served (permissive licence) · get_code("361a3caeaaaf446b")
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 3, dtype family float, shape (2, 3, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class SimpleLayerNorm(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.bias = nn.Parameter(torch.zeros(hidden_size))
    
    def forward(self, x):
        mean = x.mean(-1, keepdim=True)
        std = x.std(-1, keepdim=True)
        return (x - mean) / (std + 1e-6) * self.weight + self.bias

class SimpleMLP(nn.Module):
    def __init__(self, hidden_size, intermediate_size):
        super().__init__()
        self.gate_proj = nn.Linear(hidden_size, intermediate_size)
        self.up_proj = nn.Linear(hidden_size, intermediate_size)
        self.down_proj = nn.Linear(intermediate_size, hidden_size)
    
    def forward(self, x):
        gate = torch.sigmoid(self.gate_proj(x))
        up = self.up_proj(x)
        hidden = gate * up
        return self.down_proj(hidden)

class QwenLayer(nn.Module):
    def __init__(self, hidden_size=4, intermediate_size=8):
        super().__init__()
        self.post_attention_layernorm = SimpleLayerNorm(hidden_size)
        self.mlp = SimpleMLP(hidden_size, intermediate_size)
    
    def forward(self, hidden_states):
        return hidden_states

hidden_size = 4
layer = QwenLayer(hidden_size=hidden_size, intermediate_size=8)
hidden_states = torch.randn(2, 3, hidden_size)

ARGS = {
    'layer': layer,
    'hidden_states': hidden_states,
    'args': (),
    'kwargs': {}
}
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 3, dtype family float, shape (2, 3, 4) — shape (2, 3, 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
qwen_last_attn_ffn_only_forward Ran
ran; our contract draft was wrong, not the code
icip-cas/ShortX/shortx/shortgpt/adapters/qwen.py
code served (permissive licence) · get_code("5649b66daf070ead")
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 3, dtype family float, shape (2, 5, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build minimal nn.Module for self_attn
class MockAttention(nn.Module):
    def forward(self, hidden_states, attention_mask, position_ids, past_key_value, use_cache, cache_position, position_embeddings, **kwargs):
        # Return hidden_states unchanged and None for second return value
        return hidden_states, None

# Build minimal nn.Module for input_layernorm
class MockLayerNorm(nn.Module):
    def forward(self, x):
        return x

# Build minimal nn.Module for post_attention_layernorm
class MockPostAttentionLayerNorm(nn.Module):
    def forward(self, x):
        return x

# Build minimal nn.Module for mlp
class MockMLP(nn.Module):
    def forward(self, x):
        return x

# Build the parent module that contains self_attn, input_layernorm, post_attention_layernorm, mlp
class MockDecoderLayer(nn.Module):
    def __init__(self):
        super().__init__()
        self.input_layernorm = MockLayerNorm()
        self.self_attn = MockAttention()
        self.post_attention_layernorm = MockPostAttentionLayerNorm()
        self.mlp = MockMLP()

# Create an instance to use as 'self'
mock_layer = MockDecoderLayer()

# Construct arguments
batch_size = 2
seq_len = 5
hidden_dim = 4
num_heads = 2

hidden_states = torch.randn(batch_size, seq_len, hidden_dim)
attention_mask = torch.ones(batch_size, 1, seq_len, seq_len)
position_ids = torch.arange(seq_len).unsqueeze(0).expand(batch_size, -1)
past_key_value = None
use_cache = False
cache_position = torch.arange(seq_len)
position_embeddings = torch.randn(seq_len, hidden_dim)

ARGS = {
    "self": mock_layer,
    "hidden_states": hidden_states,
    "attention_mask": attention_mask,
    "position_ids": position_ids,
    "past_key_value": past_key_value,
    "use_cache": use_cache,
    "cache_position": cache_position,
    "position_embeddings": position_embeddings,
}
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 3, dtype family float, shape (2, 5, 4) — shape (2, 5, 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
resolve_benchmark_json Ran
ran; our contract draft was wrong, not the code
VisionOPD/Vision-OPD/eval/prepare_data.py
code served (permissive licence) · get_code("225a4721e03c0bbb")
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

ARGS = {
    'benchmark': 'zoombench'
}
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
should_retry_existing_record Ran
ran, violated the contract we drafted
VisionOPD/Vision-OPD/eval/infer.py
code served (permissive licence) · get_code("36cd5e00420a29f4")
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

# Test case 1: Normal dict with valid model_answer
item_valid = {
    'model_answer': 'This is a valid response'
}

# Test case 2: Dict with empty string model_answer
item_empty = {
    'model_answer': '   '
}

# Test case 3: Dict with API_ERROR
item_api_error = {
    'model_answer': '[API_ERROR] Something went wrong'
}

# Test case 4: Dict with FUTURE_ERROR
item_future_error = {
    'model_answer': '[FUTURE_ERROR] Future error occurred'
}

# Test case 5: Dict with non-string model_answer
item_non_string = {
    'model_answer': 12345
}

# Test case 6: Dict without model_answer key
item_missing = {
    'other_field': 'value'
}

# Primary test case - a dict with model_answer key
ARGS = {
    'item': {
        'model_answer': 'This is a test response'
    }
}
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
variance_change Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
icip-cas/ShortX/shortx/shortgpt/metrics.py
code served (permissive licence) · get_code("a1a3b6eb1f6a9fa4")
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 (6,) 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 with shape (batch=2, seq_len=3, hidden_dim=4)
# in_hidden and out_hidden are used to compute variance across features
in_hidden = torch.randn(2, 3, 4)
out_hidden = torch.randn(2, 3, 4)

ARGS = {
    "in_hidden": in_hidden,
    "out_hidden": out_hidden,
}
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 (6,) — shape (6,) 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
calc_hrbench Not yet run
unverified (harvested, no recorded run)
VisionOPD/Vision-OPD/eval/cal_acc.py
code served (permissive licence) · get_code("106f8da519c5a301")
How it ran

no invocation recorded.

pope_extract Not yet run
unverified (harvested, no recorded run)
VisionOPD/Vision-OPD/eval/judge_qwenlm.py
code served (permissive licence) · get_code("5a8d7a90cea59181")
How it ran

no invocation recorded.

prepare_vstar Not yet run
unverified (harvested, no recorded run)
VisionOPD/Vision-OPD/eval/prepare_data.py
code served (permissive licence) · get_code("d982d36f7304ff08")
How it ran

no invocation recorded.

Also present in other papers' code

Of 24 samples lifted from this paper's linked repositories, 24 carry a function identity and 0 do not. Of the 24 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 1 the same month, 1 later) and 23 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

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

FunctionAlso present inEarliest-datedOthers named
generate_config_from_args 57 other papers
44 dated earlier, 4 the same month, 6 later, 3 undated
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2509.25787 (2025-09)
Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking
2510.08630 (2025-10)
ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection
and 54 more not named here

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

Structured pruning is a hardware-friendly way to compress LLMs, but it is mostly validated on multiple-choice recognition tasks, while the same compressed checkpoints can collapse on the free-form generation that deployment actually requires. Two observations trace this gap. First, greedy pass@1 nearly vanishes after compression, yet pass@k recovers substantially under repeated sampling: useful generations are demoted, not erased. Second, in the recoverable regime we study, failure manifests mainly as suffix repetition. Recovery should therefore train on the compressed model's own on-policy states with dense token-level supervision, which On-Policy Distillation (OPD) provides by reusing the pre-compression model as a frozen teacher. However, long on-policy rollouts spend early recovery budget on low-information repetitive suffixes, delaying loss descent. To mitigate this waste, we propose ShortOPD, a short-to-long OPD schedule that detects teacher-confirmed repetitive suffixes, treats the surviving prefix as each rollout's effective length, and allocates future rollout budgets to the effective lengths the policy can currently use. Across math, code, and open-ended generation, ShortOPD raises the compressed model's score to about 9× its unrecovered value and 1.6-4.4× standard recovery recipes (SFT w/o KD, KD, and SeqKD), and it matches a fixed 8192-token rollout horizon within two points using a quarter of the training time (8.5 vs. 35.9 hours) and 71% fewer rollout tokens. We hope this recipe helps move structured pruning beyond marginal gains on perplexity and multiple-choice benchmarks, a step closer to deployment-ready generation quality.

For agents

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

get_harvested_code_for_paper("2607.13124")
get_code_for_paper("2607.13124")
have("2607.13124")

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

curl https://syntology.ai/api/ran/2607.13124.json

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

[![Syntology run record](https://syntology.ai/api/ran/2607.13124.svg)](https://syntology.ai/paper/2607.13124)

Connect an agent — have() is free.