Gap Finding

Once you've assembled what's known, the next question is what isn't. Gap finding is the workflow for surfacing under-explored areas, weak claims, and open questions that the literature hasn't yet answered. It's often the most valuable output of a review — it's what turns a survey into a contribution.

What counts as a gap

Three distinct things, useful to keep separate:

  1. Coverage gaps — topics adjacent to your subject that the KB doesn't cover, even though the literature exists. Often a curation problem: add the missing papers.
  2. Evidence gaps — claims on the map that rest on too few or too weak sources. The work was done, but not enough of it.
  3. Conceptual gaps — questions the literature hasn't yet asked, or assumptions it hasn't examined. These are the openings for new research.

Each kind needs a different response.

Step 1 — Audit coverage

Ask the agent to compare what's on the map against what the KB contains:

Compare the current mindmap to the KB. Are there documents in the KB whose contributions aren't reflected on the map? List them with a one-line summary of what they add.

Often there are — large workflows leave some sources behind. Decide per item whether to fold the missed contribution in or leave it out deliberately.

Step 2 — Audit evidence strength

For each major branch, ask the agent to evaluate the load each citation is carrying:

Within the selected subtree, identify claims that rest on only one citation or on sources with low confidence scores. List them with their citations and a note on what would strengthen each.

Outputs will land in two buckets:

  • Add sources — the claim is true but undersupported. Action: find more evidence (often in your KB; sometimes new papers needed).
  • Soften the claim — the evidence is genuinely thin. Action: rewrite the claim to match what the evidence actually shows.

Don't accept the agent's verdict uncritically here. A claim with one citation can still be defensible if the source is authoritative.

Step 3 — Surface conceptual gaps

This is the high-value step. A useful instruction:

Looking across the current mindmap, identify questions that the literature on this topic hasn't yet directly addressed, or assumptions that authors share but don't examine. Be specific — name the assumption, name who shares it, name what would be required to test it.

Expect 3–8 candidates. Treat them as drafts — your job is to evaluate which are genuinely open, which are open but uninteresting, and which the agent missed (since "the literature hasn't addressed X" is a hard claim to make from retrieval alone).

For each gap you want to take seriously, create a node and develop it:

Open question: Are pre-collapse warning indicators visible at non-palatial sites?
├─ Why this is unaddressed
├─ What current work assumes
├─ What evidence would resolve it
└─ Feasibility

Step 4 — Cross-check the gaps

For each candidate gap, sanity-check that it really is one:

For each "open question" node, search the KB to verify that this question is genuinely under-addressed. If you find papers that do address it, list them and revise the gap accordingly.

Some "gaps" the agent surfaces are actually addressed in papers it didn't surface in the original review. This step closes that loop.

Step 5 — Triage

Not every gap is worth pursuing. Use whatever criteria fit your work:

  • Tractability — is there a feasible way to address it?
  • Impact — would resolving it change the picture meaningfully?
  • Fit — does it match your skills, time, and resources?

Tag the survivors with a label like research-agenda so they're easy to find later.

Tips

  • Gap finding works best after a real review pass. Running it on a thin mindmap mostly surfaces "we don't have data on X" trivialities.
  • Be skeptical of "no one has studied X" claims. Retrieval can miss things. Step 4 is non-negotiable.
  • The most useful gaps are usually structural, not topical. "No one has compared A and B" or "everyone assumes Y" tends to be higher-leverage than "no study on this specific phenomenon".

What's next