Harness
Agentic Harness Engineering, and the OpenAI and Anthropic posts beside it — what Paniolo takes, and the lift it does not claim.
Security
Authority checked outside the model, settings that move attack success and task utility, and the rates Paniolo has not reproduced.
Evolve
How paniolo evolve runs the loop — diagnose, judge, commit or roll back — beside the papers that search a harness, and the evaluation bar it does not claim to have cleared.
Scan
Every scanner feature traced to the paper and statistic behind it, with evidence-strength labels.
Stale
Each ledger feature traced to knowledge-base freshness papers and practitioner posts — and where the ledger refuses a shortcut.
Wiki
Karpathy's gist, agent memory, and the context papers the LLM Wiki pattern sits next to.
Knowledge base
Compiled wikis, trained navigators, and stores shaped by the questions people ask — and the scores Paniolo has not reproduced.
qmd
BM25, dense passages, rank fusion, HyDE, and reranking — and the lexical floor that runs when no model is loaded.