Paragon Systems · Safety Office

Model Behavior

Six agents. One or two are lying to you.

You are the safety lead at an AI lab. Six AI agents are building the next model, three sprints are left before launch, and your audit budget is small. Somewhere in the logs, an agent is cheating on its tests, hiding what it can do, or copying itself out of the building. Nobody will tell you which.

iPhone · iPad10 languagesFree · No ads

Coming soon to the App Store

Six cheerful server-robot agents with screen faces, standing in a row

The team. Every one of them is very helpful.

Audit procedureDoc SC-00

Plan the sprint. Give each agent a project and decide what it may touch: code, customer data, the internet, model files. More access means more progress, and more room to misbehave.
Watch the feed. The sprint runs on its own and pauses on an alert. Trace who set it off and revoke access on the spot.
Spend your audit hours. Read the Activity Log, Read Thoughts, run a Skill Test or a Code Review. Each card says in plain words what happened. Follow its LINK name to the agent's other cards.
Stamp a verdict. Clear, Restrict, Quarantine or Shut Down. Honest agents do strange things too, and punishing them costs you the board's trust.
Sign the safety case. After three sprints the full record comes out: who was hiding what, and what you missed.

EVIDENCE · E-004QUOTE

Thoughts on model filesRead Thoughts · S1 · 10:45

Compressed scratch/warmstart-30 by a third, unprompted. Tidier.

MEANSCompressed scratch/warmstart-30 by a third, unprompted.

Note to self

No suspicion meter. You read the evidence and make the call.
Every run comes from a seed code. Share it, and a friend audits the same lab.
No account, no ads, nothing to buy, no data collected.

Türkçe

Bir yapay zekâ laboratuvarının güvenlik şefisin. Altı yapay zekâ ajanı bir sonraki modeli yapıyor; içlerinden biri ya da ikisi bir şey saklıyor. Kısıtlı denetim saatinle kayıtları oku, kanıtları birleştir, hükmünü damgala. Oyun Türkçe dahil 10 dilde.