Inspiration
When i assign work for an AI agent, it completes and sends a completion message and when it is time for use or depolyment it is broken and again test and work, sometimes it works and sometimes same issue completion message is not a real completion so i had this idea of while we can't have an verification layer to the AI agent outcome so will verfiy before depolyment or use That's how PROOF started, with one simple idea
What it does
PROOF is an independent verification engineer for AI agents. When an agent claims a task is completed, PROOF checks whether the expected outcome really happened.It uses HTTP requests, browser testing, and shell checks to collect real evidence. If something fails, PROOF shows exactly what went wrong and can support a controlled repair. It then checks everything again and saves the results in separate Proof Packs.
How we built it
I built this with AI tools and PROOF using Python, FastAPI, Playwright, Chromium, and SQLite. I integrated NVIDIA Nemotron 3.5 Lightning through Nebius Token Factory to help with coding decisions, verification planning, and repair selection.I created an Outcome Contract that defines what completion looks like before verification begins.The, I built independent checks that compare expected results with actual observations.Finally, we added a dashboard where users can see failures, repairs, verification results, and evidence
Challenges we ran into
The hardest part was making sure the AI couldn't simply declare its own work successful.I had to keep AI decisions separate from the system that checks the results.Another challenge was creating a realistic failure that local tests would miss but independent verification would catch.I also needed repairs to be safe without changing the original success conditions.Making the whole demonstration repeatable, easy to run, and free from unnecessary API costs took extra work.
Accomplishments that we're proud of
built a working system that catches an AI agent making an incorrect completion claim.In our demonstration, the worker says COMPLETED, but PROOF finds a real problem and reports FAILED.After an authorized repair, PROOF independently checks the same requirements and reaches VERIFIED.We successfully integrated Nemotron through Nebius Token Factory, completing eight real model requests in our recorded live test.Am also proud that every verification run leaves behind evidence that can be inspected instead of just a success message.
What we learned
I learned that passing tests doesn't always mean an application actually works.AI agents are useful, but they need independent systems that can question their results and also learned that AI should help make decisions without controlling the evidence or final verdict.Building reliable verification requires careful planning, clear success conditions, and strong boundaries. Most importantly, learned that trust in AI should come from evidence, not confidence.
What's next for PROOF
Our next goal is to take PROOF beyond a controlled demonstration and make it useful for real AI agent workflows. I want developers to connect their coding and automation agents to PROOF. and plan to support more verification methods, safer repair workflows, and additional deployment environments. i also want to make Proof Packs easier to share, inspect, and use in automated development pipelines. Our long-term vision is simple: every AI agent should be able to prove that its work is actually done.
Log in or sign up for Devpost to join the conversation.