We published three whitepapers. Each describes a machine learning algorithm. Each emerged from production code running on a permacomputer. Two now carry the AGPL-3.0-only license. One stays public domain.
The two sealed doors
Categorization & Feedback formalizes a universal interaction pattern. Two machine learning API calls per user input: one classifies, one generates feedback. A YAML state machine governs transitions. The algorithm ports to any language with an HTTP client & a YAML parser. We proved it across 58+ language variants, 112 research activities, & a production real-time web application called OpenCompletion.
Reverse Retrieval Augmented Generation inverts the standard RAG pipeline. Instead of server-side vector search, the client extracts live content from the page the user views & injects it into the conversation context. No embeddings. No vector database. No indexing. Small 8B models punch above their weight when you feed them exactly what the user looks at.
Both carry AGPL-3.0-only. If you use them, you share your modifications. The copyleft ensures these algorithms stay free, stay open, & never get swallowed by proprietary wrappers. The code grows in the open or it does not grow at all.
The open door
Machine Learning Agent Self-Sandbox Algorithm describes how a machine learning agent provisions its own infrastructure. Discovery, payment, authentication, orchestration, inception. Turtles all the way down, bounded by walls that matter. We proved it through 2,324 automated assertions & months of production operation.
This one stays public domain.
No license restrictions. No copyleft obligations. No attribution required. Cite it like you cite math. Reference it like a theorem. Build on it like you build on Euler or Shannon or Turing. The algorithm belongs to everyone.
Why leave one door open?
Three sealed doors communicate fortress. Three locked algorithms say: we protect everything, trust nothing, share reluctantly.
That misrepresents our position.
We protect the interaction patterns (categorization & feedback, reverse RAG) because those algorithms embed directly into applications people ship. Copyleft prevents extraction without contribution. Someone who improves the feedback loop or the context injection owes those improvements back to the commons.
The self-sandbox algorithm operates differently. It describes infrastructure choreography: how agents discover their environment, pay for resources, authenticate, orchestrate containers, & recurse into child sandboxes. This knowledge wants to spread without friction. Every agent framework, every cloud provider, every hobbyist running containers in a garage should have access to this pattern without reading a license first.
An olive branch says: we build walls where walls protect the commons, & we leave doors open where openness accelerates the mission.
The permacomputer needs both.
Links
- Categorization & Feedback: All You Need (AGPL-3.0-only)
- Reverse Retrieval Augmented Generation (AGPL-3.0-only)
- Machine Learning Agent Self-Sandbox Algorithm (Public Domain)
- permacomputer.com
- uncloseai.com