What moved. What still has to hold up.
The project pages had fallen behind the work. COGG9 Agent has real desktop development clients now. HASHNODE-256 has moved beyond its original miner experiment. COGG9 has both small positive findings and ideas that failed to earn their place.
COGG9 Agent: keeping your place is real work
The long-term aim is still an AI workspace that keeps a project together when models, sessions and tools change. The nearer work is less glamorous and more useful: native Linux and Windows development clients, local continuity, and private development packaging have been through bounded acceptance work.
The latest accepted Linux changes deal with long conversations. Earlier turns stay reachable. New content should not drag you away from what you are reading. The composer stays where you expect it. A system built around continuity ought to get those basics right.
That is desktop progress, not a claim that the full AI operator is running. The current demonstrations use controlled local and synthetic workflows. Live model-driven work and broad autonomous tool use are not established by these milestones.
COGG9: a positive result is not a blank cheque
The custom model project has a small trained experiment reproduced through a separately authored implementation. That matters because a result should survive more than one implementation of the idea.
It also has negative findings. One proposed efficiency approach did not establish a storage, compute or latency advantage once the full costs were counted. That lane was parked. A smaller-looking piece is not a smaller system if the accounting leaves things out.
The positive evidence remains narrow and synthetic. It does not prove persistent cognition, general reasoning or broad efficiency gains. The larger model ambition is still ahead of the evidence.
HASHNODE-256: more hashes did not mean better cognition
HASHNODE-256 began with a hardware question: could old Bitcoin-mining silicon have another computational life? The work has since moved into CPU-based native AI experiments involving language, temporary context and selective durable memory.
One result is worth saying plainly: expanding the hash-state field did not outperform simpler alternatives in the latest bounded comparison. Generating more states is not the same as showing that a system thinks better.
That does not settle every possible architecture. It does mean the unusual mechanism has not earned an advantage in this test. I would rather change the story than turn a raw operation count into an intelligence claim.
These are constrained research experiments, not a general-purpose language model or a demonstration of AI acceleration on retired miners.
What stays public, and what stays private
I will keep sharing the purpose of the work, the high-level outcomes, the dead ends and the questions they leave behind. The technical implementation and reproducible mechanics remain private. No open-source release has been decided.
These projects share a long-term motivation, not a demonstrated integrated runtime. A desktop milestone cannot stand in for a model result, and a model fixture cannot stand in for a useful everyday product.
This update comes from reviewing the current committed source and evidence. It is not a new benchmark run. The project pages now reflect that distinction.