rabbithole

Where concept-family-explorer maps everything around a subject, rabbithole moves inward: mechanism, sub-parts, edge cases, failure modes, the history of how understanding of the concept changed, primary sources, and where experts disagree — pass after pass, until a pass turns up nothing new.

The six deepening questions

Every pass asks each of these against every claim still standing:

  1. Why is this true? — the mechanism underneath a stated fact.
  2. What happens at the boundary or limit? — edge cases, failure modes.
  3. What’s the exception? — cases the general claim doesn’t cover.
  4. Who disagrees, and on what basis? — expert disagreement, competing models.
  5. What did the primary source actually say, in its own words? — drift between original and paraphrase.
  6. What changed over time in how this was understood? — historical evolution.

Saturation, measured

Each pass’s claims are diffed against the accumulated list from every prior pass. The new-information rate — new atomic claims this pass ÷ total claims after this pass — is recorded per pass, and the run stops after two consecutive passes each score below 5%. A single low-yield pass isn’t enough to stop on; it could just be an unlucky question angle. Two in a row, using a different subset of the six questions, is the corroborating evidence that the concept is actually exhausted rather than merely under-questioned.

A pass whose new claims are mostly about a different, neighboring concept isn’t saturation — it’s a sign the “concept” was really a family, and the run hands off to concept-family-explorer instead of forcing a fake depth loop to continue.

Use it for: “deepen my understanding of X”, “exhaust X completely”, “rabbithole on X”.

Not for: a domain or family, not one concept (concept-family-explorer) · one ordinary research pass with no forced exhaustion loop (/dr) · pulling a concept’s mentions out of a corpus you already hold (llms-concept-abstractor).