You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$20,000 left to reach the funding goal.
AI agents compress their memory and a summary can look right today while dropping who can revoke the agent. Contradish catches this before the memory is saved
Goal is to find out how often AI agents’ own memory summaries look correct today but quietly break their principal’s ability to revoke or correct them later and to make it easier to prevent that failure.
I will achieve this goal by running my preregistered 40-case study on frontier models then testing the Contradish guard which blocks unsafe memory rewrites before they’re saved in a real agent framework. I will publish the results and all data/code openly.
I'll know it worked if the study produces a clear answer which means a public and preregistered measure of how often frontier models’ memory summaries are right today but unsafe later. A low rate counts as success too as long as the result is solid. The guard should work on a real framework and catches the unsafe rewrites from the study and let the safe ones through and other people use it meaning an agent developer trying the guard or researchers citing or building on the results.
Model API costs and 3-4 months of my time on the formal extensions and framework integration and public write-up.
I (Michele Joseph) will lead this project. I studied computer science at Cornell and am on a gap year from Harvard Law School to focus on compression-aware AI research and Contradish.
The most likely cause would be if the problem turns out to be rare but I’d still publish the result because it’s useful to know. Another possible outcome is nobody adopts it but the study and the open spec would still put the problem on record for others to build on.
$0, the project has been self-funded so far but I’ve applied to BlueDot Impact and Lightcone Commons