Dark code is software that exists, runs, passes checks, and ships — but was never sufficiently understood, specified, authorized, or reconstructible by the organization that owns it.
Human software teams could produce poorly documented systems, but they usually left behind some trail of human intent: meetings, tickets, commits, code review discussion, architectural memory, and operational scar tissue. Agent-native development changes the scale. A system can now generate large amounts of plausible, functioning code faster than a human organization can comprehend, authorize, or reconstruct it after the fact.
The dangerous case is not code that obviously fails. The dangerous case is code that appears to work. It passes automated checks. It satisfies the immediate ticket. It may even improve the measured output. But the organization cannot say why the implementation is correct, what assumptions it embeds, which specification assertions it satisfies, or how to rebuild it from first principles if the artifact disappears.
ANDM position: internal development may be opaque to the outside world. The code must not be opaque to the organization itself.
ANDM makes the anti-dark-code guarantee structural. A conformant module hub requires code, spec, test-results, and commit-record structure. The spec must be strong enough that the code can be regenerated from it. The tests must validate the current code/spec batch. The commit record must identify the actor, operation, and governed artifact lineage.