A live recursive explanation engine
See exactly how an explanation is built.
Name a claim. Unfold writes one honest, ~150-word paragraph — then exposes the exact phrases inside it the explanation actually rests on. Click one, and it descends into its own paragraph. Keep going until you hit a first principle, or the honest edge of what anyone actually knows.
Nothing here is pre-written. Every paragraph below is generated live by Claude.
The idea
Most explanations hide their own foundation. Unfold shows it.
Every claim rests on other claims. A textbook asserts and moves on. A pundit asserts and hopes you don't ask. Unfold asserts too — but it marks exactly which words it's leaning on, and lets you go stand on them yourself, one layer down, for as many layers as it takes to hit solid ground.
How it works
Four steps. No hand-waving.
Ask
Name any claim, question, or assumption — yours, a headline's, or one you've just never quite trusted.
Read one honest paragraph
Claude writes a single ~150-word explanation. Warm and precise, no hedging filler — if something is genuinely unsettled, it says so.
Pull a thread
2–5 phrases inside that paragraph are load-bearing — the ideas the explanation actually depends on. They're the only clickable words on the page.
Descend
Click one. It expands inline into its own fresh explanation, one layer down — a real new model call, not a canned answer — until the chain bottoms out.
Live demo
This isn't a mockup. It's the actual engine.
Type your own claim, or drop one of ours. Every explanation below is a live call to Claude — the same engine, the same prompt, running in front of you right now.
What do you want to understand?
Type any claim, or drop one of ours below.
Where a dig ends
Two honest ways down stops.
Bedrock — a first principle
Some explanations bottom out in a definition, an axiom, or an identity that can't be meaningfully decomposed further — what “energy” means, a mathematical identity. When Unfold hits one, it stops and says so instead of manufacturing a fake extra layer.
Frontier — an open question
Some things bottom out in a brute fact science hasn't explained yet, or a genuinely open question. Unfold is built to say “nobody knows, honestly” rather than bluff a confident-sounding non-answer.
Under the hood
Built to be checked, not just believed.
Verbatim spans, not vibes
The model's own character offsets aren't trusted. The server re-locates every clickable phrase inside the explanation text, in order, and drops anything that isn't an exact, non-overlapping substring via repairSubclaims. A phrase either really is in the paragraph, or it never becomes clickable.
One call, one paragraph
Each descent is exactly one request to claude-sonnet-5 with structured outputs — the explanation, its subclaims, and its terminal state all come back in one guaranteed-shape JSON response. No retries stitching together a franken-answer.
The depth guard
Past 8 layers down, Unfold pauses and asks before continuing — each descent is a fresh model call, and the app is upfront with you about that cost before it happens.
Start now
Nothing left to take on faith. Start a dig.
Name a claim above and watch the ground appear under it, one honest layer at a time.