Research
Status labels are honest. Active work is not presented as established scientific fact.
Can a peer mesh with gossip membership, sandboxed execution, and content-addressed artifacts provide useful distributed compute without a trusted scheduler?
Does an independent, execution-backed verification plane reduce unjustified confidence in agent-authored changes compared to narrative self-reports?
Can an evidence-driven performance loop specialized to AMD Zen produce more trustworthy optimization decisions than generic benchmark folklore?
Can weekly USPTO XML dumps be stream-parsed into locally owned SQLite/JSONL artifacts suitable for search and downstream ML prep without loading archives into memory?