SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.02589 · 2026

PeerRank: Autonomous LLM Evaluation Through Web-Grounded, Bias-Controlled Peer Review

Nurit Cohen-Inger, Yanki Margalit, Erni Avram, Ran Taig, Oded Margalit, Caura Ai

arXiv · PDF · Open in the Atlas

Code that ran

11 of 11 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, 0 violated it and 0 have no contract verdict; 11 failures of our instrument; 0 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/2602.02589.json · badge: /api/ran/2602.02589.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
caura-ai/caura-PeerRank canonical (classifier label)
classifier basis: census
11 ran · 0 no recorded run
FunctionStatusWhere it lives
get_color Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/figure_utils.py
code served (permissive licence) · get_code("76b9fd5321ad4c7c")
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.

# Setup code for get_color function

ARGS = {
    'model': 'gpt-5.6-sol'
}
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
get_short_name Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/config.py
code served (permissive licence) · get_code("044fec136989f0d2")
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 torch.nn as nn

ARGS = {
    "model": "gpt-5.6-sol",
    "max_len": 12,
}
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
get_text_color_for_background Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/figure_utils.py
code served (permissive licence) · get_code("41cdaee442f2d812")
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.

# Setup code for get_text_color_for_background

ARGS = {
    'hex_color': '#FF5733'
}
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
get_timing_data Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/generate_figures_PeerRank.py
code served (permissive licence) · get_code("651f5dab5cafde82")
How it ran

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

import torch
import numpy as np
from scipy import stats

# Construct the data dict that get_timing_data expects
data = {
    'phase2': {
        'timing_stats': {
            'model_a': {'avg': 0.123},
            'model_b': {'avg': 0.456},
        }
    },
    'phase3': {
        'timing_stats': {
            'model_c': {'avg': 0.789},
            'model_d': {'avg': 0.321},
        }
    }
}

ARGS = {
    'data': data
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, 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
load_data Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/generate_figures_PeerRank.py
code served (permissive licence) · get_code("8606cbbc71dbca7d")
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
import tempfile
import os

# Create a temporary directory with test data files
temp_dir = tempfile.mkdtemp()
data_dir = Path(temp_dir)

# Create phase1 questions file
phase1_data = [
    {"id": 1, "question": "What is AI?"},
    {"id": 2, "question": "How does ML work?"}
]
with open(data_dir / 'phase1_questions_v1.json', 'w', encoding='utf-8') as f:
    json.dump(phase1_data, f)

# Create phase2 answers file
phase2_data = [
    {"id": 1, "answer": "AI is artificial intelligence"},
    {"id": 2, "answer": "ML learns from data"}
]
with open(data_dir / 'phase2_answers_v1.json', 'w', encoding='utf-8') as f:
    json.dump(phase2_data, f)

# Create phase3 rankings file
phase3_data = [
    {"rank": 1, "score": 0.95},
    {"rank": 2, "score": 0.87}
]
with open(data_dir / 'phase3_rankings_v1.json', 'w', encoding='utf-8') as f:
    json.dump(phase3_data, f)

# Create phase4 report file
with open(data_dir / 'phase4_report_v1.md', 'w', encoding='utf-8') as f:
    f.write("# Final Report\n\nThis is the summary report.")

ARGS = {
    'revision': 'v1',
    'data_dir': data_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
remap_scores_to_models Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank_phase3.py
code served (permissive licence) · get_code("b0e93efc0814802c")
How it ran

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

import torch
import numpy as np
import difflib
import re

# Test case 1: Exact match
scores = {
    "gpt-4-turbo": 0.85,
    "claude-3-opus": 0.92,
}

label_to_model = {
    "gpt-4-turbo": "openai/gpt-4-turbo",
    "claude-3-opus": "anthropic/claude-3-opus",
    "Response A": "openai/gpt-4-turbo",
    "Response B": "anthropic/claude-3-opus",
}

# Test case 2: With substring matches and typos
scores_complex = {
    "A": 0.88,
    "B": 0.91,
    "gpt-4": 0.87,
    "geminii-3.7-flash": 0.89,  # typo for gemini-3.7-flash
    "unknown-model": 0.75,
}

label_to_model_complex = {
    "Response A": "openai/gpt-4-turbo",
    "Response B": "anthropic/claude-3-opus",
    "gpt-4-turbo": "openai/gpt-4-turbo",
    "gpt-4-sol": "openai/gpt-4-sol",
    "gemini-3.7-flash": "google/gemini-3.7-flash",
    "claude-opus-3": "anthropic/claude-opus-3",
}

# Test case 3: Empty and edge cases
scores_empty = {}
label_to_model_empty = {}

scores_no_match = {
    "completely-unknown": 0.50,
}

label_to_model_limited = {
    "gpt-4": "openai/gpt-4",
}

ARGS = {
    "scores": scores,
    "label_to_model": label_to_model,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
remap_scores_to_models Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank_ui.py
code served (permissive licence) · get_code("67312eaf38e60f5c")
How it ran

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

import torch
import numpy as np

torch.manual_seed(0)

# Construct test arguments for remap_scores_to_models
scores = {
    "bert": {"accuracy": 0.95, "f1": 0.92},
    "gpt2": {"accuracy": 0.88, "f1": 0.85},
    "roberta_base": {"accuracy": 0.97, "f1": 0.94},
}

label_to_model = {
    "bert": "bert-base-uncased",
    "gpt2": "gpt2",
    "roberta_base": "roberta-base",
}

ARGS = {
    "scores": scores,
    "label_to_model": label_to_model,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sanitize_prompt Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/providers.py
code served (permissive licence) · get_code("2cdfd127e45b5201")
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 torch.nn as nn

torch.manual_seed(0)

# The function sanitize_prompt takes a single string parameter
# and performs character replacements on it.
# No complex objects are needed - just a string.

ARGS = {
    "text": "Hello, this is a test with special characters: ' ' " " – — … \xa0"
}
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
serpapi_search Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/providers.py
code served (permissive licence) · get_code("e697af0f40c4bee3")
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 coroutine. 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 os
import asyncio

# The function is async and takes:
# - query: str (a search query string)
# - max_results: int (number of results, default 5)

# Since this is an async function that makes network calls,
# we'll set up arguments that would allow it to be called.
# We cannot actually execute it without a real SERPAPI_KEY and network access,
# but we can construct valid argument values.

ARGS = {
    'query': 'artificial intelligence',
    'max_results': 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
aiohttp — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type coroutine — a coroutine output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a coroutine 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
set_active_categories Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/config.py
code served (permissive licence) · get_code("c91159eb1f2e6405")
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.

# Examine the function to understand what it needs
# The function takes two optional list[str] parameters: include and exclude
# It filters ALL_CATEGORIES based on these parameters

# Set up test cases for the function
ARGS = {
    'include': None,
    'exclude': None,
}

# Alternative test case 1: with include parameter
# ARGS = {
#     'include': ['reasoning', 'creative'],
#     'exclude': None,
# }

# Alternative test case 2: with exclude parameter
# ARGS = {
#     'include': None,
#     'exclude': ['current events'],
# }

# Alternative test case 3: with both include and exclude
# ARGS = {
#     'include': ['factual', 'reasoning'],
#     'exclude': ['logic'],
# }
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
tavily_search Ran
ran; our contract draft was wrong, not the code
caura-ai/caura-PeerRank/peerrank/providers.py
code served (permissive licence) · get_code("b5132ba6b51d4482")
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 coroutine. 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 asyncio
import os

# Set up environment variable for Tavily API key
# Since this is a test setup without actual network access, we can set a dummy key
os.environ['TAVILY_API_KEY'] = 'test_api_key_12345'

# Define arguments for tavily_search function
ARGS = {
    'query': 'what is artificial intelligence',
    'max_results': 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
aiohttp — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type coroutine — a coroutine output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a coroutine 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

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 11 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

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

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

Repositories linked to this paper

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

Abstract

Evaluating large language models typically relies on human-authored benchmarks, reference answers, and human or single-model judgments-approaches that scale poorly, become quickly outdated, and mismatch open-world deployments that depend on web retrieval and synthesis. We introduce PeerRank, a fully autonomous end-to-end evaluation framework in which models generate evaluation tasks, answer them with category-scoped live web grounding, judge peer responses and aggregate dense peer assessments into relative performance estimates, without human supervision or gold references. PeerRank treats evaluation as a multi-agent process where each model participates symmetrically as task designer, respondent, and evaluator, while removing biased judgments. In a large-scale study over 12 commercially available models and 420 autonomously generated questions, PeerRank produces stable, discriminative rankings and reveals measurable identity and presentation biases. Rankings are robust, and mean peer scores agree with Elo. We further validate PeerRank on TruthfulQA and GSM8K, where peer scores correlate with objective accuracy. Together, these results suggest that bias-aware peer evaluation with selective web-grounded answering can scale open-world LLM assessment beyond static and human curated benchmarks. Code and dataset are available at: https://github.com/caura-ai/caura-PeerRank

For agents

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

get_harvested_code_for_paper("2602.02589")
get_code_for_paper("2602.02589")
have("2602.02589")

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

curl https://syntology.ai/api/ran/2602.02589.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.02589.svg)](https://syntology.ai/paper/2602.02589)

Connect an agent — have() is free.