Xueping Gao pitches 'Executable Code Knowledge' for AI agents

Code that carries its own validation evidence recovered exact patch selectors on 9 of 11 tasks in a small Python study. Hide that declared evidence and the recovery drops to 1 of 11.

That is the core claim of an August preprint on arxiv from Xueping Gao, introducing what the paper calls Executable Code Knowledge (ECK). Rather than pushing retrieval, summaries, graphs, rules, or reverse specifications around code, the paper argues selected code units should themselves become 'a source-bound object combining stable identity, semantics, executable behavior, contracts, evidence, relations, provenance, validation state, and a query interface.'

The prototype is Python. The evaluation covers three real Python repositories and 26 controlled patch tasks. On exact changed-line impact against independently authored labels, the paper reports 'precision, recall, and F1 all 1.000.' AST-bounded fingerprints correctly classified 50 positive changes and 17 unrelated same-file controls, while static rules snapshots detected none of the 50 stale cases. The gap between direct ECK and the hidden-evidence baseline is reported as statistically significant at paired exact McNemar p=0.0078.

Gao's own bottom line is that no single layer wins: the paper recommends 'a hybrid architecture: retrieval for coverage, ECK for source and evidence governance, and projections for delivery.' Two researchers we track shared the paper within a day of it going up.