What gap are you circling?
How does it work?
Your question is split into search angles, nine or ten methodologically distinct papers are pulled from arXiv and summarized, a hundred directions are generated through ten analytic lenses, and the one you pick is novelty-checked against fresh retrievals, then drafted through six constraint gates and a fifteen-point self-audit.
Do I just watch, or do I steer?
You steer at every step: rephrase the angles, strike papers from the survey, rewrite any direction before developing it, and paste outside reviews that permanently improve future proposals.
Are the novelty claims guaranteed?
No. They are literature-based estimates against what was retrieved, and the interface says so. A failed retrieval reports itself as a failed search, never as novelty.
What does a run cost?
A few cents of model usage and about two minutes. Everything runs against your own configured endpoint; nothing is tracked and no cookies are set.
0 ways into your question.
These become the literature search. Click any to rephrase it, cut the ones you don't mean, add the one only you know about.
0 papers, one per approach.
Kept from the candidate pool by embedding distance and a second-reader screen. Strike any that miss your intent; the survey reruns without them.
0 directions. Pick one.
Ten lenses over every paper, clustered by convergence. Swipe through them. When you are on the one, hit develop. You can rewrite it first; your wording carries into the proposal.
Before writing: what already exists.
Each direction was checked against fresh retrievals. A failed search reports itself as a failed search, never as novelty.
Ready to hand to a human.
Paste any outside review below its proposal: it gets revised to address it, and the review's general lessons become permanent guardrails for every future run.