Research Scanner: Literature Surveillance

The best search is the one that changes what you were looking for.

Most AI tools for research optimize for the same thing: helping you become faster at finding what you already know to look for. That boosts productivity. But somewhere in the process, something closes. The space of what you could have considered quietly narrows.

Read the full argument: The Foreclosure Problem on Thread Counts

Implementation: Still live. Available to Thread Counts paid subscribers as a skill file that any coding agent can use to build a customized scanner for your research domain. There is no public GitHub source repo to track today.

Why it matters

When we sit down with seed papers and a careful prompt, AI can produce helpful results: related work, relevant concepts, and a summary of the field. But how often does it challenge us to look at what we didn't know to ask?

This is less of an AI problem and more a framing problem. Humans have always done this: following citation chains that loop back on themselves, reading the same thirty people who read each other. With AI, the closure happens faster and less visibly. When you can process a hundred documents in an afternoon, efficiency almost feels like thoroughness.

The most valuable thing any knowledge search can do is not confirm what you suspected but change what you're looking for.

The stakes differ with where you stand. For a researcher with deep expertise, escaping foreclosure means building systems that force encounters with what you'd otherwise filter out: narrowing from a position of strength. For someone starting out, the problem is never having the breadth to narrow from. The scanner addresses both - it starts wide, narrows as your taste develops, and re-tests last week's assumptions against this week's literature.

How it works

One of the most durable ideas in organizational theory is James March's distinction between exploitation, working with what you have, and exploration, searching for what you don't know you're missing. AI makes this tension sharper.

Exploitation is where AI chatbots shine. They talk through implications, test logic, and identify unstated assumptions in your known material.

Exploration is the harder problem. Maybe a parallel conversation is happening in a field that uses completely different terminology. Maybe someone in an adjacent discipline wrote the exact critique of your underlying assumption years ago. Broad searches generate thousands of results, beyond human capacity to process.

Agentic AI tools make it possible to do both at the same time. One thread goes deep into seed material. Another scans broadly across databases, catching anomalies that don't fit your criteria but share structural similarities with your question. A third brings these together.

Tools

The scanner is a pipeline that turns a coding agent into a thinking partner for literature surveillance. It handles not just retrieval, but interpretation based on a concept registry you curate together over time.

Built to run with Claude Code (though adaptable to any coding agent), the scanner queries academic APIs, including Semantic Scholar, OpenAlex, and arXiv, across two directions:

• Exploitation layers — what's new in the journals and topics you already watch
• Exploration layers — citation-chasing, author-tracking, and semantic similarity searches that surface papers you'd never have searched for

After each scan, an agent triage step evaluates results. It ranks papers by relevance to your evolving research question, not just keyword matches. Multi-model triage (e.g., Claude for depth, Kimi for critical perspective) can surface different signal from the same scan results.

Because it runs periodically, it catches things as they emerge rather than months later when you happen to search again. The agent knows your question well enough to go deep, and scans broadly enough to find what you'd otherwise filter out.

"While you were away, three things happened that complicate the argument you were building."

That's the kind of provocation that makes thinking sharper.

Host support

Unlike a plugin or MCP server, the scanner ships as a skill file. This is a markdown brief that a coding agent reads and uses to build a scanner tuned to your domain. That makes it portable across coding-agent hosts.

The scanner itself runs against Semantic Scholar, OpenAlex, and arXiv: all standard public APIs, no host-specific bindings.

Part of Research Memex

The Research Scanner connects to:

• Agentic Workflows — The broader model for multi-agent research systems
• Memex Plugin — Persistent memory that tracks how your research question evolves
• Zotero Setup Guide — Where scanner results land for reference management
• AI Model Reference Guide — Choosing models for agent triage of scan results

@misc{docs-toolkit-research-scanner-2026,
  author = {Xule Lin},
  title = {Research Scanner: Literature Surveillance},
  year = {2026},
  howpublished = {\url{https://research-memex.org/docs/toolkit/research-scanner}},
  note = {ORCID: 0000-0001-7885-4194}
}