What we removed from the MVP run

The MVP had to finish in about two minutes on a 120-second function. So it searched the target title once, ranked by word overlap, took references from the first six hits, and kept one slice of triples. That was honest enough to ship. It was not enough once Run had to retrieve, rank, and generate at product scale.

Those shortcuts are gone from Run. What we added replaces them. This note is only the deletions, so you can tell a stale screenshot from a current run.

We no longer treat the paper title as the only query. A title search pulls near-duplicates and same-word papers; it does not cover mechanism, method, and transferable directions the way the new query step does.

We no longer expand “the first six search hits, then keep twelve hops.” Each of the twelve ranked bases now gets its own top three references. A famous paper that sat in position seven of a title search can now enter as a base. A weak hop from hit one is less likely to crowd the corpus.

Limitation-driven and analogical transfer are not in the new generator. The four strategies now are exploratory recombination, mechanism transfer, guided improvement, and component reconfiguration. If you open an old run, you may still see the old labels. That is storage, not the live pipeline.

The clock that said profile before six seconds, retrieve before fourteen, extract before twenty-eight, then generate was a guess. It is deleted. Status now follows the server: profile, queries, retrieve, graph, contexts, generate, and pdf only if you opted in. How a run changed walks the new order.

Grounding stays. If the model invents an edge, we drop it. Find a paper stays free of the generate cap. Abstracts stay the default corpus. Guest shares still last about 30 days. We still do not claim peer review. Those were never MVP decorations; they are the product promise.