False Corroboration Across and Within Sub-Agents

In a run of the worked example, sub-agent 3 (pricing and margin structure) reports "API inference margins are compressing roughly 30% year over year," citing an analyst report R and a blog post B that quotes R's figure. Sub-agent 4 (buyer motivations) reports "enterprise buyers cite cost as the primary driver," citing R's survey appendix and a news article N summarizing R. Both sub-agents marked their sources as consistent, so both analysis steps stayed on the cheap tier with high confidence. The drafted report says "multiple sources confirm" in both sections.

1. Identify the two distinct failure modes hidden in this trace and explain why the ask-the-web-style dedupe a sub-agent runs on its own fetched batch handles one of them but structurally cannot handle the other. Include what went wrong with the per-sub-agent escalation decision.
2. Design the aggregation-stage similarity check so it catches the cross-sub-agent case without over-merging: what should it compare, what should it collapse, and what should it explicitly not collapse? What happens to the two findings' confidence?
3. Explain why the subject insists this happens before outline planning, and propose how provenance information could also feed back into Stage 2's escalation signal. Would sub-agent 3, once reduced to a single genuine source, escalate to the reasoning tier or not?