AI says "done" but humans still verify — a proposal to track real completion state
A recurring problem shows up in long-running AI agent work: a new agent needs the same project explanation repeated, the AI reports "done" but a human still has to verify and clean up, a handoff passes along information but not responsibility, and nobody can tell what was actually checked, what remains unresolved, or where to restart if something goes wrong. This isn't just a context-window limitation — running out of context explains why an agent forgets things, but it doesn't explain why a plausible-looking "done" report leaves nothing behind that the next human or AI can safely verify, roll back, or continue from.
That gap is called False . is a proposed framework to make AI inspectable rather than just plausible-looking.
At its core, before an output is accepted as "done," it records what was actually checked, what is still unresolved, who has authority over the next step, and where to restart from if something is wrong. A lightweight version can start with just one ordinary AI output, without blind , multiple agents, or a formal benchmark.
Key points
- Points out that in long-running AI work, humans still verify and clean up even after the AI reports "done"
- Frames this as a distinct problem — False — separate from context-window limits
- Proposes : record what was checked, what's unresolved, and where to restart before accepting an output as done
- A lightweight version can start with a single AI output, no or formal benchmark required