SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.12652 · 2026

Multi-Rollout On-Policy Distillation via Peer Successes and Failures

Matt Fredrikson, Gaurav Mittal, Yizhou Zhao, Chen Wu, Haixin Wang, Weichen Yu, Yu Hu, Xiaoze Liu, Xiaomin Li, Yinyi Luo, Ruowang Zhang

arXiv · PDF · Open in the Atlas

Code that ran

23 of 29 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; 0 honoured the contract we drafted, 2 violated it and 0 have no contract verdict; 21 failures of our instrument; 6 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/2605.12652.json · badge: /api/ran/2605.12652.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
viviable/mopd_code canonical (classifier label)
classifier basis: census
23 ran · 6 no recorded run
FunctionStatusWhere it lives
average Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
viviable/mopd_code/analysis/compute_teacher_signal_metrics.py
code served (permissive licence) · get_code("bd3548f55a4d47e5")
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 float, 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

# The function takes a list of floats and computes their average
# We need to construct a valid list of float values

values_list = [1.5, 2.5, 3.5, 4.5, 5.5]

ARGS = {
    "values": values_list
}
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 float, 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
build_student_targets Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/build_target_responses.py
code served (permissive licence) · get_code("d7aa73af385d98f7")
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 hashlib
from collections import defaultdict
from typing import Any

def stable_target_id(target_type: str, prompt_id: str, source_response_id: str | None, response_text: str) -> str:
    raw = f"{target_type}|{prompt_id}|{source_response_id or 'none'}|{response_text}"
    return hashlib.sha1(raw.encode('utf-8')).hexdigest()[:20]

def build_student_targets(candidates: list[dict[str, Any]], student_success_per_prompt: int, student_failure_per_prompt: int) -> list[dict[str, Any]]:
    by_prompt: dict[str, list[dict[str, Any]]] = defaultdict(list)
    for row in candidates:
        by_prompt[row['prompt_id']].append(row)
    targets: list[dict[str, Any]] = []
    for prompt_id, rows in by_prompt.items():
        rows = sorted(rows, key=lambda x: x['response_id'])
        successes = [row for row in rows if row['success']][:student_success_per_prompt]
        failures = [row for row in rows if not row['success']][:student_failure_per_prompt]
        for row in successes:
            targets.append({'target_id': stable_target_id('student_success', prompt_id, row['response_id'], row['response_text']), 'target_type': 'student_success', 'prompt_id': prompt_id, 'source_response_id': row['response_id'], 'response_text': row['response_text'], 'success': True, 'reward': row['reward'], 'source_model': 'student_rollout', 'difficulty_bucket': row['difficulty_bucket'], 'metadata': {'rollout_step': row.get('rollout_step'), 'pred': row.get('pred'), 'ground_truth': row.get('ground_truth')}})
        for row in failures:
            targets.append({'target_id': stable_target_id('student_failure', prompt_id, row['response_id'], row['response_text']), 'target_type': 'student_failure', 'prompt_id': prompt_id, 'source_response_id': row['response_id'], 'response_text': row['response_text'], 'success': False, 'reward': row['reward'], 'source_model': 'student_rollout', 'difficulty_bucket': row['difficulty_bucket'], 'metadata': {'rollout_step': row.get('rollout_step'), 'pred': row.get('pred'), 'ground_truth': row.get('ground_truth')}})
    return targets


candidates = [
    {
        'prompt_id': 'prompt_1',
        'response_id': 'resp_1',
        'response_text': 'This is a correct answer.',
        'success': True,
        'reward': 0.95,
        'difficulty_bucket': 'easy',
        'rollout_step': 1,
        'pred': 'correct_output',
        'ground_truth': 'correct_output'
    },
    {
        'prompt_id': 'prompt_1',
        'response_id': 'resp_2',
        'response_text': 'This is an incorrect answer.',
        'success': False,
        'reward': 0.1,
        'difficulty_bucket': 'easy',
        'rollout_step': 1,
        'pred': 'wrong_output',
        'ground_truth': 'correct_output'
    },
    {
        'prompt_id': 'prompt_1',
        'response_id': 'resp_3',
        'response_text': 'Another correct response.',
        'success': True,
        'reward': 0.92,
        'difficulty_bucket': 'medium',
        'rollout_step': 2,
        'pred': 'correct_output',
        'ground_truth': 'correct_output'
    },
    {
        'prompt_id': 'prompt_2',
        'response_id': 'resp_4',
        'response_text': 'A successful response.',
        'success': True,
        'reward': 0.88,
        'difficulty_bucket': 'hard',
        'rollout_step': 1,
        'pred': 'result_a',
        'ground_truth': 'result_a'
    },
]

ARGS = {
    'candidates': candidates,
    'student_success_per_prompt': 2,
    'student_failure_per_prompt': 1,
}
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
check_correctness Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/verl/utils/reward_score/prime_code/utils.py
code served (permissive licence) · get_code("2d94e42b971ca757")
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 multiprocessing
import sys

# Mock run_test function that the code calls
def run_test(in_outs, test, debug, timeout):
    """Mock implementation of run_test"""
    # Returns (result_list, metadata_dict)
    num_inputs = len(in_outs['inputs'])
    result = [1 for _ in range(num_inputs)]  # Success indicators
    metadata = {'test_name': 'sample_test', 'passed': True}
    return result, metadata

# Create mock in_outs dict with inputs
in_outs = {
    'inputs': [
        {'arg1': 10, 'arg2': 20},
        {'arg1': 30, 'arg2': 40},
        {'arg1': 50, 'arg2': 60}
    ]
}

# Mock generation string (code to test)
generation = """
def add(a, b):
    return a + b
result = add(10, 20)
"""

ARGS = {
    'in_outs': in_outs,
    'generation': generation,
    'timeout': 10,
    'debug': True
}
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 — 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
ci_yerr Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
viviable/mopd_code/analysis/plot_teacher_metrics.py
code served (permissive licence) · get_code("5ef8c6dd23fae513")
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 torch

torch.manual_seed(0)

# Build the metrics_view dict structure that ci_yerr expects
metrics_view = {
    'conditions': {
        'condition_A': {
            'accuracy_ci': (0.05, 0.15),
            'loss_ci': (0.02, 0.08)
        },
        'condition_B': {
            'accuracy_ci': (0.03, 0.12),
            'loss_ci': (0.01, 0.06)
        },
        'condition_C': {
            'accuracy_ci': (0.04, 0.14),
            'loss_ci': (0.025, 0.075)
        }
    }
}

# Build the arguments
ARGS = {
    'metrics_view': metrics_view,
    'conditions': ['condition_A', 'condition_B', 'condition_C'],
    'metric': 'accuracy',
    'values': [0.85, 0.90, 0.88]
}
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 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
condition_colors Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/plot_teacher_metrics.py
code served (permissive licence) · get_code("288f33b5eca65cc7")
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

ARGS = {
    'n': 5
}
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
fetch_history Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/draw_acc.py
code served (permissive licence) · get_code("caa5d8692e5bdb76")
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 wandb
from unittest.mock import Mock, MagicMock

# Create a mock API and run object to avoid actual network calls
mock_api = Mock()
mock_run = Mock()

# Mock the scan_history method to return test data
def mock_scan_history(keys=None):
    return [
        {'_step': 0, 'loss': 0.5},
        {'_step': 1, 'loss': 0.4},
        {'_step': 2, 'loss': None},  # This row will be skipped
        {'_step': 3, 'loss': 0.3},
    ]

mock_run.scan_history = mock_scan_history
mock_api.run.return_value = mock_run

# Patch wandb.Api to return our mock
wandb.Api = Mock(return_value=mock_api)

ARGS = {
    'entity': 'test_entity',
    'project': 'test_project',
    'run_id': 'test_run_123',
    'metric': 'loss'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
wandb — 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_metrics Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/plot_teacher_metrics.py
code served (permissive licence) · get_code("aebe25129f6e22f7")
How it ran

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

import json
import tempfile
from pathlib import Path

# Create a temporary metrics file
with tempfile.NamedTemporaryFile(mode='w', suffix='.json', delete=False) as f:
    metrics_data = {
        "accuracy": 0.95,
        "loss": 0.123,
        "epochs": 10,
        "model_name": "test_model"
    }
    json.dump(metrics_data, f)
    temp_path = Path(f.name)

ARGS = {
    "path": temp_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
make_client Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/generate_strong_success.py
code served (permissive licence) · get_code("b97c9a97c01f9eea")
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 OpenAI. 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.

# The function make_client requires:
# - api_key: str
# - base_url: str | None
# - default_headers: dict[str, str] | None (optional, defaults to None)

# Mock the OpenAI class since we don't want to actually import it
class OpenAI:
    def __init__(self, api_key: str, base_url: str | None = None, default_headers: dict | None = None):
        self.api_key = api_key
        self.base_url = base_url
        self.default_headers = default_headers

# Set OpenAI in the global namespace so the function can access it
import builtins
builtins.OpenAI = OpenAI

ARGS = {
    'api_key': 'sk-test-key-12345',
    'base_url': 'https://api.openai.com/v1',
    'default_headers': {'Authorization': 'Bearer token', 'Content-Type': 'application/json'}
}
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 OpenAI — a OpenAI output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a OpenAI 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
only_int_check Ran
ran, violated the contract we drafted
viviable/mopd_code/verl/utils/reward_score/prime_code/testing_util.py
code served (permissive licence) · get_code("6339c418c950b9c8")
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

torch.manual_seed(0)

# The function only_int_check takes a single parameter 'val' and checks if it's an integer
# We need to provide a valid integer value for this parameter

ARGS = {
    "val": 42
}
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
parse_run_path Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/draw_acc.py
code served (permissive licence) · get_code("0898e734fc60f2e9")
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.

from urllib.parse import urlparse

def parse_run_path(run_value: str) -> tuple[str, str, str]:
    if run_value.startswith('http://') or run_value.startswith('https://'):
        parsed = urlparse(run_value)
        parts = [part for part in parsed.path.split('/') if part]
        if len(parts) >= 4 and parts[2] == 'runs':
            return (parts[0], parts[1], parts[3])
        raise ValueError(f'Unsupported W&B run URL format: {run_value}')
    parts = [part for part in run_value.split('/') if part]
    if len(parts) == 3:
        return (parts[0], parts[1], parts[2])
    raise ValueError('Run must be a W&B URL or entity/project/run_id.')


# Test case 1: URL format with https
run_value_1 = "https://wandb.ai/entity/project/runs/run_id_123"

# Test case 2: Simple entity/project/run_id format
run_value_2 = "my_entity/my_project/run_abc"

ARGS = {
    "run_value": run_value_1
}
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
rankdata Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
viviable/mopd_code/analysis/compute_teacher_signal_metrics.py
code served (permissive licence) · get_code("414365b67878b9e4")
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

torch.manual_seed(0)

# The function rankdata takes a list of floats and returns their ranks
# with average ranking for tied values

values = [3.5, 1.2, 3.5, 2.1, 1.2, 4.0]

ARGS = {
    "values": values
}
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
read_jsonl Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/score_teacher_contexts.py
code served (permissive licence) · get_code("903f3f3a478c7f63")
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
from pathlib import Path
import tempfile
import os

# Create a temporary JSONL file with test data
temp_dir = tempfile.mkdtemp()
temp_file = Path(temp_dir) / "test.jsonl"

# Write some test data to the JSONL file
test_data = [
    {"name": "Alice", "age": 30, "city": "New York"},
    {"name": "Bob", "age": 25, "city": "Los Angeles"},
    {"name": "Charlie", "age": 35, "city": "Chicago"},
]

with temp_file.open("w") as f:
    for row in test_data:
        f.write(json.dumps(row) + "\n")

ARGS = {
    "path": temp_file,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
read_jsonl Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/compute_teacher_signal_metrics.py
code served (permissive licence) · get_code("2b6d83de04b5e4ad")
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 a temporary JSONL file with test data
temp_dir = tempfile.TemporaryDirectory()
temp_path = Path(temp_dir.name) / "test.jsonl"

# Write some test JSONL data
with temp_path.open("w") as f:
    f.write(json.dumps({"key1": "value1", "key2": 42}) + "\n")
    f.write(json.dumps({"key1": "value2", "key2": 100}) + "\n")
    f.write("\n")  # Test empty line handling
    f.write(json.dumps({"key1": "value3", "key2": 200}) + "\n")

ARGS = {
    "path": temp_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
read_variant_config Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/build_context_variants.py
code served (permissive licence) · get_code("759afe3c4dcbc678")
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
from pathlib import Path
from tempfile import NamedTemporaryFile

# Create a temporary JSON file with a valid variant config
config_data = {
    'global': {
        'dont_reprompt_on_self_success': True,
        'feedback_only_without_solution': False,
        'summary_k': 2,
        'summary_sampling_policy': 'first_k'
    },
    'variants': [
        {'name': 'base'},
        {'name': 'base_raw'},
        {'name': 'solution', 'include_solution': True}
    ]
}

# Write to a temporary file
temp_file = NamedTemporaryFile(mode='w', suffix='.json', delete=False)
json.dump(config_data, temp_file)
temp_file.close()

ARGS = {
    'path': temp_file.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 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
resolve_api_config Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/generate_strong_success.py
code served (permissive licence) · get_code("d9156c6e78fc7a12")
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 argparse

ARGS = {
    'args': argparse.Namespace(
        backend='openrouter',
        api_key_env='OPENAI_API_KEY',
        base_url='https://custom.example.com',
        http_referer='https://myapp.example.com',
        x_title='My Application'
    )
}
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
resolve_dtype Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/score_teacher_contexts.py
code served (permissive licence) · get_code("2d63c8da40e7f663")
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 dtype. 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 takes a single parameter: dtype_name (str)
# It maps string names to torch dtypes or returns None for 'auto'
# We need to test with a valid dtype name string

ARGS = {
    'dtype_name': 'float32'
}
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 dtype — a dtype output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dtype 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
resolve_run Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/fetch_wandb_metric_compare.py
code served (permissive licence) · get_code("53fb0fcdb63f7f14")
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 Mock. 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 wandb
from unittest.mock import Mock, MagicMock

# Create a mock wandb.Api instance
mock_api = Mock(spec=wandb.Api)

# Create a mock run object with necessary attributes
mock_run = Mock()
mock_run.display_name = "test_run"
mock_run.name = "test_run_id"

# Mock the api.runs() method to return a list with one run
mock_api.runs = Mock(return_value=[mock_run])

ARGS = {
    'api': mock_api,
    'entity': 'test_entity',
    'project': 'test_project',
    'run_name': 'test_run'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
wandb — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Mock — a Mock 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
sanitize Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/fetch_wandb_metric_compare.py
code served (permissive licence) · get_code("6f3c534aa4ba4ee2")
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

def sanitize(text: str) -> str:
    return text.replace('/', '_').replace('@', 'at').replace(' ', '_').replace(':', '_')

ARGS = {
    'text': 'hello/world@example.com:path with spaces'
}
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
sanitize_metric_name Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/draw_acc.py
code served (permissive licence) · get_code("f5674775fedebe04")
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

ARGS = {
    'metric': 'loss/train@epoch_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 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
stable_target_id Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/build_target_responses.py
code served (permissive licence) · get_code("2d00b81e871f020d")
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

ARGS = {
    "target_type": "response",
    "prompt_id": "prompt_123",
    "source_response_id": "source_456",
    "response_text": "This is a test response for stable target id generation."
}
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
stable_variant_id Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/analysis/build_context_variants.py
code served (permissive licence) · get_code("3d4ee5bf4a4b62a8")
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

def stable_variant_id(condition: str, prompt_id: str, response_id: str) -> str:
    raw = f'{condition}|{prompt_id}|{response_id}'
    return hashlib.sha1(raw.encode('utf-8')).hexdigest()[:20]


ARGS = {
    'condition': 'control_group_a',
    'prompt_id': 'prompt_001',
    'response_id': 'response_001'
}
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
string_int_check Ran
ran, violated the contract we drafted
viviable/mopd_code/verl/utils/reward_score/prime_code/testing_util.py
code served (permissive licence) · get_code("0b462a89ce04228a")
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 = {
    "val": "12345"
}
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
truncatefn Ran
ran; our contract draft was wrong, not the code
viviable/mopd_code/verl/utils/reward_score/prime_code/testing_util.py
code served (permissive licence) · get_code("22849627aa42fa47")
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

torch.manual_seed(0)

# The function truncatefn takes a string and optional length parameter
# It checks if the string is a string type, and if it's longer than length,
# it truncates it by keeping the first half and last half with '...(truncated) ...' in between

# Test case 1: Short string that doesn't need truncation
short_string = "Hello, World!"

# Test case 2: Long string that will be truncated
long_string = "The quick brown fox jumps over the lazy dog. " * 20  # Create a long string

ARGS = {
    "s": long_string,
    "length": 300
}
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
load_model_and_tokenizer Not yet run
unverified (harvested, no recorded run)
viviable/mopd_code/analysis/score_teacher_contexts.py
code served (permissive licence) · get_code("1a52c3aa3027b3fe")
How it ran

no invocation recorded.

load_spec Not yet run
unverified (harvested, no recorded run)
viviable/mopd_code/analysis/plot_heatmap_grid.py
code served (permissive licence) · get_code("984827baf66eba8d")
How it ran

no invocation recorded.

resolve_run_path Not yet run
unverified (harvested, no recorded run)
viviable/mopd_code/analysis/fetch_wandb_metric_compare.py
code served (permissive licence) · get_code("b29f33cf4c2111e6")
How it ran

no invocation recorded.

run_test_code Not yet run
unverified (harvested, no recorded run)
viviable/mopd_code/verl/utils/reward_score/feedback/code.py
code served (permissive licence) · get_code("45691de9cab1a113")
How it ran

no invocation recorded.

run_test_func Not yet run
unverified (harvested, no recorded run)
viviable/mopd_code/verl/utils/reward_score/feedback/code.py
code served (permissive licence) · get_code("4a71e4f42656119d")
How it ran

no invocation recorded.

run_test_std Not yet run
unverified (harvested, no recorded run)
viviable/mopd_code/verl/utils/reward_score/feedback/code.py
code served (permissive licence) · get_code("f1b61b9749cac892")
How it ran

no invocation recorded.

Also present in other papers' code

Of 29 samples lifted from this paper's linked repositories, 29 carry a function identity and 0 do not. Of the 29 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 28 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
truncatefn 1 other paper
1 dated earlier
2411.17501 (2024-11)
The Limits of Inference Scaling Through Resampling

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Large language models are often post-trained with sparse verifier rewards, which indicate whether a sampled trajectory succeeds but provide limited guidance about where reasoning succeeds or fails. On-policy distillation (OPD) offers denser token-level supervision by training on student-generated trajectories, yet existing methods typically distill each rollout independently and ignore the other attempts sampled for the same prompt. We introduce Multi-Rollout On-Policy Distillation (MOPD), a peer-conditioned distillation framework that uses the student's local rollout group to construct more informative teacher signals. MOPD conditions the teacher on both successful and failed peer rollouts: successes provide positive evidence for valid reasoning patterns, while failures provide structured negative evidence about plausible mistakes to avoid. We study two peer-context constructions: positive peer imitation and contrastive success-failure conditioning. Experiments on competitive programming, mathematical reasoning, scientific question answering, and tool-use benchmarks show that MOPD consistently improves over standard on-policy baselines. Further teacher-signal analysis shows that mixed success-failure contexts better align teacher scores with verifier rewards, indicating that the gains arise from more faithful, instance-adaptive supervision. These results indicate that effective on-policy distillation should exploit the student's multi-rollout trial-and-error behavior rather than treating rollouts as isolated samples. Code is available at https://github.com/viviable/mopd_code.

For agents

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

get_harvested_code_for_paper("2605.12652")
get_code_for_paper("2605.12652")
have("2605.12652")

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

curl https://syntology.ai/api/ran/2605.12652.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.12652.svg)](https://syntology.ai/paper/2605.12652)

Connect an agent — have() is free.