You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
In todays world, AI programs are starting to do things on their own: send emails, delete files, move money. When one of them makes a bad call, by the time we figure it out its already to late.
I want to build a checkpoint that stands in the way. Before the AI can act, the checkpoint reads what it is about to do twice, in two different ways. If both readings say it is safe, the action goes through. If either says no, it stops.
Then I will test it in the open. I will write a list of safe actions and harmful ones, run each one many times, and publish every result, good or bad.
Your donation pays for that test to happen and for the results to be public.
My goal is to know with real numbers: can a checkpoint like this be trusted?
In 3 months I will:
1. Build a safe practice space where an AI can "send email" or "delete files" without touching anything real.
2. Put the checkpoint between the AI and those actions so the AI cannot skip it.
3. Write a list of safe and harmful actions to test with.
4. Run each one many times and count: Does the checkpoint give the same answer every time? How many harmful actions does it stop? How many safe ones does it block by mistake?
5. Publish the test list and all the results.
If the answer turns out to be "no, it can't be trusted on its own," I will publish that to full transparency is needed now more then ever.
for 3 months:
- $21,000: my pay to work on this full-time for 3 months ($7,000 a month, taxes included)
- $1,500: the cost of running the AI thousands of times for the tests
- $1,000: hosting and tools
- $2,400: a 10% cushion for miscellaneous costs.
I'm Georges Talon, founder of Rage Relief LLC me and the rest of my team have created a working product with the design and built of COSMIC AEI, a live app that reads every message two ways before it answers. Anyone can try it free at app.cosmaei.com. Running it with real users, I found and fixed real problems. One example: people who describe terrible events in calm, flat words were being read as "fine." I tracked down why and fixed it. That lesson is built into this test, because an AI can describe a harmful action in calm, flat words too. My company was accepted into Anthropic's Claude for Startups program in June 2026. Furthermore, we urged by NASA to apply for they license program in order to have access to software to help further our goal.
Most likely cause: the checkpoint turns out to be unreliable. It gives different answers to the same action, or it misses harmful ones. If so, I publish those numbers, and people building AI safety checks learn what not to rely on.
Second: the checkpoint and the AI it watches run on the same underlying AI model, so they may share the same blind spots. I will test for this and report it.
Third: Publishing the test list and results lets anyone check my work.
The product code behind my existing app stays private. What this project produces, the test list and the results, is public.
The work so far is self-funded, about $2,000 of my own money in 2026.
I have three applications pending for a longer version of this same project: Anthropic External Researcher Access (free AI usage credits only,) EA Funds Transformative AI Fund (applied October 3, 2026), and Foresight Institute. None has been decided.