Why conscRAG shows four findings and hides the rest

The generator can pair four contexts with four strategies and two directions each. That is useful when you want a pile of candidates. A person sitting with one target paper does not need a pile. They need four directions that are different from each other and tied to real neighbors.

So the stored run keeps `ideas` (the four) and `candidates` (the rest). A share link reloads both. If you only ever open the default view, you are not missing a secret fifth strategy. You are choosing not to scroll.

Expand after you have clicked evidence on the default four. If those four already point at the same two papers, the extra candidates will too — the corpus is the bottleneck, not the ranker. If they point at different rings of Figure 1, expand and filter by strategy. Recombination and reconfiguration are the ones that wander; improvement stays closer to the target.

Old posts that said “four strategies, four ideas” as if that were the generator are outdated. Read how findings connect to retrieved works for the citation rule, and what we added for the new generator. The four you see first are a reading order, not a claim that the model only thought four thoughts.

It cannot tell you the idea is unpublished. It cannot replace a related-work section. It cannot see workshop PDFs that never reached OpenAlex. Peer review is still the check; we wrote that before this update and it is still true: why peer review matters.

If you want a number, use the corpus count and the triple count on the result line, not the finding index. Finding 1 is not better science than Finding 4. It is one strategy’s best grounded card from this run.