Proof-of-purpose

Proof-of-purpose dashboard
User does 'canary' test via public apis
Adversarial prompt test gets a STRONG reaction from Anthropic.. f^_^; )

Inspiration

Arun was listening to a Sam Harris podcast a few weeks ago where he speaks with William MacAskill about AI and the future of humanity (https://www.youtube.com/watch?v=5JF0gOQvskY) and among other things, they were discussing the current intelligence explosion and near-term risks of concentrated AI power. With that problem on his mind, Arun entered the Hackathon and recruited some teammates to build a solution.

How we built it

We used Claude Code with an AFK Ralph loop to code the site while developing features as a team in parallel through GitHub issues. The full flow was: PRD (GitHub issue) -> /prd-to-issues skill to make smaller GitHub issues in vertical slices for end-to-end testing (monorepo) -> Ralph loop -> merger agent -> auto-deploy (via Railway).

Challenges we ran into

UI was more difficult to update that expected using words (UI changes usually took 2-3 loops).
Sometimes, the adversarial prompts triggered Claud Code to shut down our session (see attached images) but we reached out to them and explained our motivations.
Categorizing the constitution into just a few categories was subjective on our part (we had to cut scope for the hackathon)
We had to collapse the LLM answer into yes/no/no-answer so it was deterministic because fuzzy evaluation proved hard; we think a future judging system could address this and expand the capabilities of the protocol.

Accomplishments that we're proud of

We think we found a potential method to address a HUGE problem (maybe the biggest problem civilization faces today!) in 36 hours :)

What we learned

Working with agentic AI dramatically increased our speed, not only due to the coding speed but because we could work in parallel with the execution to talk as a team and write more detailed feature descriptions.