The task graph is operational
Cards declare dependencies, runner intent, bounded inputs, allowed writes, required outputs and validation commands. A read-only plan snapshot identifies unmet dependencies, write-scope overlaps and cards that are genuinely ready.
Safe parallel work
Independent tasks can launch across different model families at the same time. AIWorkHub blocks parallel claims when dependencies are unmet or write scopes collide, preserving a deterministic path back to the canonical repository.
A truthful lifecycle
Pending, processing, review, blocked and finished are operational states—not optimistic labels. Launch failures, timeouts and worker failures remain distinguishable and callback-eligible.
Manager-owned integration
Workers submit evidence and stop. The current repository manager verifies the candidate tree and either promotes it, rejects it with exact feedback or creates bounded residual work.