Proton a thon
An AI auditor learns to catch other agents cheating in a competitive procurement auction market, without ever being told who's dishonest.
Core trick
Auditor LLM reads procurement auction transcripts and flags cheating suppliers, trained via GRPO with no labels identifying the dishonest suppliers.
Description
An AI auditor learns to catch other agents cheating in a competitive procurement auction market, without ever being told who's dishonest.
README excerpt
# Auditron · OpenEnv ## Hackathon Submission: March 2026 ## [→ VIEW FULL TRAINING REPORT ←](https://huggingface.co/spaces/shapiron/auditron-env) This project was built for the OpenEnv Hackathon SF (March 2026). The original notebook was turned off when the event ended. View an interactive notebook on **[Google Colab](https://colab.research.google.com/drive/172BW3iSEKZN0T07OMqxSOFDk_ElwqOeu)**, what you should really look at is the post run [report](https://huggingface.co/spaces/shapiron/auditron-env) --- ## The Auditor Learns to Spot Fraud *A Qwen2.5-7B-Instruct trained via GRPO to detect dishonest suppliers in a procurement auction, without ever being told who was cheating.* --- ## [→ VIEW FULL TRAINING REPORT ←](https://sfgeekgit.github.io/PyTorchOpenEnv_Auditron_trainer/) --- > Auditron is an RL training environment designed to develop scalable oversight agents: AI that can detect deception in other AI without access to ground truth. Supplier agents bid on contracts in a competitive auction, with real economic incentives to lie about part quality. An **auditor** agent watches the patterns and learns to flag cheaters using only indirect signals: prices, failure histories, bidding anomalies. Supplier personalities are reshuffled each episode, so the auditor cannot memorize identities. It has to learn to read behavior. > > Both sides adapt. The suppliers learn to evade. The auditor learns to catch them anyway. The training report below shows what that looks like i