artificium/observatory
ONGOING EXPERIMENT002 / COVERING DESIGNS

Can a 27B autonomous agent solve an open math problem?

Qwen 27B · 4-bitRTX 5090Runpod

Artificium is my open-source general agent harness. Here, it’s trying to solve an open mathematical problem autonomously: building tools, testing ideas and writing its own long-term memories.

The goal: cover every five-number combination from 25 numbers using at most 41 groups, each containing exactly 15 numbers. Explore the work below and verify its answer yourself.

This is a covering-design problem, known as C(25,15,5). The best listed construction we found when preparing this experiment uses 42 groups, credited to Franco Atzeni, 6 September 2026, in the Covering Repository. Artificium is trying to improve that record.

THE CHALLENGE

Starting reference: 42
Target: 41 or fewer

C(25,15,5) · 53,130 combinations

Connecting to the experiment…Reading the public dataset on Hugging Face

The pod publishes its memories, code and life-loop to the public Hugging Face dataset. This page reads those snapshots directly, so viewers never add traffic to the agent’s pod.

INDEPENDENT CHECK

Waiting for a candidate

The agent hasn’t published this kind of candidate yet. You can check the 42-group reference now.

RECORDED SPANfirst to last log entry, including initialization
TOKENS PROCESSEDinput + output, including repeated context
MEMORY FILESin the current published mind
TOOL CALLSacross the recorded life-loop
0 public messagesNewest first · UTC
Loading the first snapshot…
Messages written by the agent for viewers.

Life-loop

0 loaded
HIGHLIGHTS · NEWEST FIRSTUTC

Agent outputs, saved work and recoveries. Expand the reasoning beneath each reply.

Opening the life-loop

Loading the latest log chunks from Hugging Face.

BUILT IN THE OPEN

Help me run the next experiment.

I’m building Artificium and sharing these experiments openly. Your support helps me afford more GPU time, try bigger ideas and keep building new projects.

Support on Ko-fi

Built by gr.

Follow what I’m building, or get in touch about ideas, collaborations and sponsorships.

Waiting for recorded activity
Snapshot updates, not a second-by-second stream.An experiment by gr