Hypothesis Generation & Testing

A literature review tells you what's been said. The next move — where research actually happens — is forming your own hypotheses and stress-testing them against the evidence. Agent Bayes is well-suited to the second half of that loop: you propose, the agent challenges.

Step 1 — Surface candidate hypotheses

If you already have hypotheses in mind, skip ahead. Otherwise, ask the agent to draw out implicit claims from what's on the map.

Looking at the current mindmap, propose three concrete hypotheses about the Late Bronze Age collapse that the assembled evidence suggests but doesn't yet directly state.

Treat the output as a starting point, not a verdict. Edit, combine, discard — these are your hypotheses, the agent is just helping you articulate them.

Step 2 — Make each hypothesis a node

Create a dedicated subtree per hypothesis. A useful structure:

H1: Climate-driven drought was the primary trigger of the collapse
├─ Supporting evidence
├─ Counter-evidence
├─ Predictions / what we'd expect to see
└─ Open questions

You can build this scaffold manually or ask the Editor to do it:

Under each of these three hypotheses, create empty child nodes for "Supporting evidence", "Counter-evidence", "Predictions", and "Open questions".

Step 3 — Populate the supporting side

For each hypothesis, select the "Supporting evidence" node and ask the agent to fill it:

Find evidence in the KB that supports this hypothesis. Be specific about which claims, papers, and data are doing the work.

This is a standard research task — the Researcher retrieves, the Editor adds citation-backed claims under the selected node.

Step 4 — Steelman the opposition

The critical step. Select "Counter-evidence" and explicitly invite the agent to argue the other side:

Find the strongest counter-evidence to this hypothesis in the KB. Don't soften — present opposing claims at their strongest, with citations.

This is where the system's bias toward preserving contradicting viewpoints earns its keep. A hypothesis that survives a steelmanned opposition is meaningfully different from one that survived because no one pushed back.

Step 5 — Predictions and falsifiers

For each hypothesis, fill in "Predictions" with claims of the form "if H is true, we should observe X". The agent can help generate these from what's already on the map:

Given this hypothesis and the evidence above, what specific empirical predictions would distinguish it from the main alternatives? List each as a falsifiable claim.

Predictions you can't check against the KB go into "Open questions" — they're what you'd need to do new work to answer.

Step 6 — Adjudicate

With supporting evidence, counter-evidence, and predictions all in one subtree, the hypothesis is now in a state where you (the human) can make a real judgment. The agent won't decide for you. What it gives you is the structured comparison.

Useful follow-up:

Compare these three hypotheses on the strength of their supporting evidence, the difficulty of explaining the counter-evidence, and the testability of their predictions. Don't pick a winner — just lay out the trade-offs.

Patterns to watch for

  • Confirmation drift. If the supporting side balloons and the counter side is thin, re-run Step 4 with sharper phrasing. The agent will go looking harder.
  • Cherry-picked citations. Confidence scores help here — low confidence + small number of sources = a claim that depends on too little.
  • Hidden assumptions. Ask the agent to make assumptions explicit: "What does this hypothesis assume that hasn't been stated?"

What's next