SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.12243 · 2026

Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature

Yubo Wang, Xiaoyu Liu, Menglin Yang, Jinkai Tao

arXiv · PDF · Open in the Atlas

Code that ran

6 of 6 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, 0 violated it and 0 have no contract verdict; 5 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/2604.12243.json · badge: /api/ran/2604.12243.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
TaoJinkai/ckm-hypogen canonical (classifier label)
classifier basis: census
6 ran · 0 no recorded run
FunctionStatusWhere it lives
async_retry Ran
ran; our contract draft was wrong, not the code
TaoJinkai/ckm-hypogen/ckm-eval/core/engines.py
code served (permissive licence) · get_code("21c289b8cfedfd3b")
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 function. 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 logging
import asyncio
import functools

logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger('ckm.engines')

def async_retry(max_retries: int=3, delay: float=5):
    """Retry decorator for async functions with exponential backoff."""

    def decorator(func):

        @functools.wraps(func)
        async def wrapper(*args, **kwargs):
            current_delay = delay
            for attempt in range(1, max_retries + 1):
                try:
                    return await func(*args, **kwargs)
                except Exception as e:
                    if attempt == max_retries:
                        logger.error('Async API call failed after %d attempts: %s: %s', max_retries, type(e).__name__, e)
                        raise
                    logger.warning('Async API call failed (attempt %d/%d), retrying in %ds: %s: %s', attempt, max_retries, current_delay, type(e).__name__, e)
                    await asyncio.sleep(current_delay)
                    current_delay = min(current_delay * 2, 120)
        return wrapper
    return decorator


ARGS = {
    'max_retries': 3,
    'delay': 5.0
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
cosine_similarity Ran
ran; our contract draft was wrong, not the code
TaoJinkai/ckm-hypogen/ckm-eval/core/engines.py
code served (permissive licence) · get_code("b52c1522a75ffebb")
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 math

def cosine_similarity(vec_a: list[float], vec_b: list[float]) -> float:
    if len(vec_a) != len(vec_b):
        return 0.0
    dot_product = sum((a * b for a, b in zip(vec_a, vec_b)))
    norm_a = math.sqrt(sum((a * a for a in vec_a)))
    norm_b = math.sqrt(sum((b * b for b in vec_b)))
    if norm_a == 0 or norm_b == 0:
        return 0.0
    return dot_product / (norm_a * norm_b)


ARGS = {
    "vec_a": [1.0, 2.0, 3.0, 4.0],
    "vec_b": [2.0, 3.0, 4.0, 5.0],
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
extract_topic Ran
ran; our contract draft was wrong, not the code
TaoJinkai/ckm-hypogen/benchmark/ckm_benchmark/extract_arxiv_ids.py
code served (permissive licence) · get_code("98acb447cb6c2c72")
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 tempfile
import json
from pathlib import Path

# Create a temporary directory structure to simulate the topic_dir
tmpdir = tempfile.TemporaryDirectory()
topic_dir = Path(tmpdir.name) / "test_topic"
topic_dir.mkdir(exist_ok=True)

metabolism_dir = topic_dir / "metabolism"
metabolism_dir.mkdir(exist_ok=True)

# Create config.json with processed_ids
config_data = {
    "processed_ids": ["2301.1234", "2302.5678v1", "2303.9999v2"]
}
config_file = metabolism_dir / "config.json"
with open(config_file, 'w') as f:
    json.dump(config_data, f)

# Create log directory with window files
log_dir = metabolism_dir / "log"
log_dir.mkdir(exist_ok=True)

window_file_1 = log_dir / "window-2024-01-ingest.md"
window_file_1.write_text(
    "# Window Report\n"
    "arxiv_id: 2304.1111v1\n"
    "arxiv_id: 2305.2222\n"
    "arxiv_id: 2304.1111\n"  # duplicate after stripping version
)

window_file_2 = log_dir / "window-2024-02-ingest.md"
window_file_2.write_text(
    "# Another Window\n"
    "arxiv_id: 2306.3333v2\n"
    "arxiv_id: 2307.4444\n"
)

# Create fulltext_cache directory with cached files
cache_dir = metabolism_dir / "fulltext_cache"
cache_dir.mkdir(exist_ok=True)

(cache_dir / "2301.1234.txt").write_text("cached content 1")
(cache_dir / "2302.5678.txt").write_text("cached content 2")
(cache_dir / "2304.1111.txt").write_text("cached content 3")
(cache_dir / "2305.2222.txt").write_text("cached content 4")
(cache_dir / "2306.3333.txt").write_text("cached content 5")
(cache_dir / "2307.4444.txt").write_text("cached content 6")
(cache_dir / "2999.9999.txt").write_text("cached content not in init or evolution")

ARGS = {
    "topic_dir": topic_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
get_model_context_window Ran
ran, honoured the contract we drafted
TaoJinkai/ckm-hypogen/ckm-eval/config.py
code served (permissive licence) · get_code("3efe6f999cede641")
How it ran

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

import torch
import torch.nn as nn

# Set random seed for reproducibility
torch.manual_seed(0)

# The function get_model_context_window takes a single parameter: model_name (str)
# It looks up the model_name in MODEL_CONTEXT_WINDOWS dict and returns an int
# If not found directly, it checks if model_name starts with any key in the dict
# If still not found, it returns DEFAULT_CONTEXT_WINDOW

# Test cases for different scenarios:
# 1. Model name that exists directly in MODEL_CONTEXT_WINDOWS
# 2. Model name that partially matches (starts with a key)
# 3. Model name that doesn't match anything (should return DEFAULT_CONTEXT_WINDOW)

ARGS = {
    "model_name": "gpt-4o"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
slugify_topic_name Ran
ran; our contract draft was wrong, not the code
TaoJinkai/ckm-hypogen/ckm-eval/topics.py
code served (permissive licence) · get_code("b3ec5063f58688f8")
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 slugify_topic_name(name: str) -> str:
    slug = re.sub('[^a-z0-9]+', '_', name.lower()).strip('_')
    return re.sub('_+', '_', slug)

ARGS = {
    'name': 'Hello World! This is a TEST_topic-name_123'
}
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
summary_row Ran
ran; our contract draft was wrong, not the code
TaoJinkai/ckm-hypogen/benchmark/ckm_benchmark/recompute.py
code served (permissive licence) · get_code("444685618ba072df")
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 statistics as st

# Construct sample data for the function
data = [
    {
        'yield': 5,
        'hit_rate': 0.8,
        'best_match_score': 0.95,
        'total_generation_tokens': 2000000,
        'duration': 120.0,
    },
    {
        'yield': 3,
        'hit_rate': 0.6,
        'best_match_score': 0.85,
        'init_tokens': 1000000,
        'evolution_tokens': 500000,
        'duration': 180.0,
    },
    {
        'yield': 7,
        'hit_rate': 0.9,
        'best_match_score': 0.92,
        'total_generation_tokens': 2500000,
        'duration': 150.0,
    },
]

label = "test_summary"

ARGS = {
    'label': label,
    'data': data,
}
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

Also present in other papers' code

Of 6 samples lifted from this paper's linked repositories, 6 carry a function identity and 0 do not. Of the 6 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 6 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

Identifying promising research directions in fastmoving subareas is one of the most cognitively expensive tasks in modern AI research. Existing LLM-driven scientific-discovery systems are typically limited to one-shot prompting on static literature snapshots and validated only against contemporary judges (human reviewers, agent peer review, wet-lab assays, or self-evaluation), leaving open whether they anticipate future trends. We present Continuous Knowledge Metabolism (CKM), an AI workflow for hypothesis generation with three key capabilities: (i) continuous literature metabolism via sliding windows that maintain an evolving knowledge state; (ii) predictive evaluation, grading hypotheses against papers published after the generation window; and (iii) practitioner-grade failure detection that diagnoses workflow failure modes from its outputs. On a 50-topic ML benchmark, CKM-LITE produces at least one validated hypothesis on 72% of topics (36/50), more than doubling a one-shot baseline (30%) at ∼$3 per topic and 91% lower token cost. Validated hypotheses precede their matched papers by a mean of 404 days (n=55 hits across 36 topics; median 399, range 66-757). Broadly, predictive validation against future literature offers a falsifiable, low-cost alternative to contemporary-judge protocols and can be applied wherever a corpus has dated publication records.

For agents

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

get_harvested_code_for_paper("2604.12243")
get_code_for_paper("2604.12243")
have("2604.12243")

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

curl https://syntology.ai/api/ran/2604.12243.json

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

[![Syntology run record](https://syntology.ai/api/ran/2604.12243.svg)](https://syntology.ai/paper/2604.12243)

Connect an agent — have() is free.