Cosmos — find the research nobody has written down yet

What it is. Cosmos is an MCP tool that takes a research prompt and returns hypotheses sitting in one specific quadrant: logically implied by the literature, yet still unpublished. Not "novel ideas" — low-hanging fruit the field has implicitly set up but nobody has picked.

How you use it. Ask it a question — "sparse autoencoders and superposition" — from Claude, the API, or any MCP client. Out comes a ranked roadmap: each card is a testable hypothesis, the real arXiv papers that imply it, and the nearest existing paper proving it's unwritten.

Why it matters for PhD students. Picking a thesis problem is the hardest, highest-stakes guess of a research career — you can burn a year on something already done, or too speculative to ground. Cosmos surfaces the defensible gap: ideas the literature logically points to, filtered to the ones no paper has claimed. A starting menu of fundable, derisked directions — with citations attached.

The two axes — judgment, not vibes

• Gemini = free exploration. It reads the corpus and generates boldly — recombining real methods with real open gaps into fresh hypotheses. The creative engine, deliberately unconstrained.
• Tavily = the literature + the novelty check. It fetches recent arXiv as clean text, and is hit a second time to ask "has anyone already written this?" — killing the already-published.
• Prometheux = symbolic judgment. Vadalog reasoning decides whether a hypothesis is genuinely implied (grounded ∧ fills_gap) — a derivation, not a guess. It didn't hallucinate the idea; it derived it, and shows the chain.

So Gemini dreams, and Tavily + Prometheux are the cold harnesses that keep only what's implied and unwritten. Generation and judgment, cleanly separated.

How it works (the pipeline)

prompt
  → PLANNER     (Gemini)   decompose into focused sub-queries
  → SCOUT       (Tavily)   parallel arXiv search → clean text snippets
  → ANALYST     (Gemini)   extract claim / method / gap per paper
  → SYNTHESIZER (Gemini)   compose hypotheses = real method × real gap
  → CRITIC  (3-stage filter, cheapest first):
        2a. unwritten?  embedding nearest-paper (ClickHouse)
        2b. unwritten?  Tavily search the hypothesis itself
        3.  implied?    Prometheux symbolic reasoning
  → ROADMAP   (implied AND unwritten, with citations)

The product is the intersection: implied (Prometheux) AND unwritten (Tavily + embeddings).

The build

Genuinely a nice one — four sponsor tools wired into a pipeline that flows end to end, a clean two-axis idea that got sharper under pressure, and the satisfying moment when a too-strict filter went from 0 → 17 real hypotheses on live arXiv. Fun problem, fun stack, good hackathon.

Built With

• claude
• clickhouse
• gemini
• prometheux
• python