Notes on multi-model research, engineering decisions, and the philosophy of AI-assisted scholarship.
Every LLM has blind spots — training cutoffs, citation habits, regional biases. Running several in parallel and diffing their outputs turns individual weaknesses into collective strength.
When DeepSeek says a concept originated in 1927 and another model says 1953, who wins? Our adjudicator doesn't pick sides — it weighs evidence, flags disagreement, and writes the nuanced truth.
General reviews tell you what scholars think. Evolution tracking tells you how they got there — origins, inflection points, and the debates that shaped today's consensus.
API pricing changes monthly and no reseller can keep up. BYOK means you always pay cost, choose your own model mix, and never get locked into someone else's margin.
Server-sent events, per-model buffers, reconnect replay, and what happens when a model dies mid-stream — a tour of the plumbing behind Scholarly's real-time columns.
A research tool that quietly fabricates results is worse than useless. Here's why every mock response in Scholarly carries a visible demo badge — and always will.
Long-form articles are on the way — for now these are the ideas we're thinking about.