Reasoning Trace X-ray
A reasoning model’s hidden “thinking” usually renders as a scrolling wall of text nobody reads. This X-ray turns it into what it really is: a chain of steps, where each step computes a value from the one before it. Edit any single step — halve the doubling, flip a subtraction, change the split — and watch every step that depends on it re-run forward and the final answer fork live. It makes “the model thought about it” into an object you can grab and argue with. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the five-step reasoning trace top to bottom — it’s a small word problem worked out one move at a time (start 100, double it, give some away, split the rest).
- Change any step’s value in its number box, or click a preset — Fork step 2 or Fork step 3 — to jump straight to an edit.
- Watch the steps below your edit recompute and the Original answer → Forked answer panel update; steps above your edit never move.
- Open the X-ray panel to see exactly which step forked and how the change cascaded, then copy or share the fork.
What this clears up (the fundamentals)
- A reasoning trace is a dependency chain, not prose — each step consumes the previous step’s result, so the “thinking” has a shape you can walk. Naming that shape is the whole point of chain-of-thought.
- Errors propagate forward, never backward — edit a middle step and everything downstream changes while everything upstream is untouched. That is why one wrong intermediate move quietly poisons a confident final answer.
- Self-correction is just re-running from a step — “the model caught its mistake” means it re-derived the chain from the step it fixed. You can do the same edit by hand here.
- Extra reasoning tokens buy re-derivation, not magic — the payoff of a reasoning model over a fast one is precisely these intermediate steps being spent and checked; sometimes they change the answer, sometimes they don’t.
Where it’s used
A top-of-funnel AI-engineering teaser — the fastest way to feel why one edited reasoning step forks the whole answer, instead of taking a hidden trace on faith. It’s a miatz build-lab concept playable feeding AIE-101: play it here, then learn to build the step tree and the forward re-run yourself, and reuse the same move in AIE-401 to debug real eval failures.
FAQ
What is a “reasoning trace” or chain of thought?
The intermediate steps a reasoning-capable model works through before its final answer — normally hidden or collapsed. Here it’s laid out as an ordered chain of steps, each computing a number from the previous one, so you can see and edit the thinking.
What happens when I edit one step?
The edited step takes your value, and every step that depends on it re-runs forward from there. Steps before your edit are left exactly as they were. The final answer forks to reflect the new chain.
Why did editing a step not change the answer sometimes?
Because your value can re-derive to the same downstream numbers, or because you edited the last step, which nothing else depends on. The X-ray tells you which case you hit — a change that cancels out is itself a lesson about how the chain flows.
Is this a real model call?
No. It’s a fixed, deterministic teaching chain — the same edit always produces the same fork, with no randomness. The full playground version routes a real reasoning model via your own key (BYOK), but this concept playable computes locally so it works offline.
Is anything uploaded?
No. The trace and your edits are computed entirely in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.
Limits
A teaching model with one fixed word problem and five arithmetic steps — it shows the structure of chain-of-thought reasoning (a forward dependency chain you can fork), not the messy natural-language reasoning of a live model. Real traces branch, backtrack, and self-correct in prose; the payoff of reasoning tokens over a fast model depends on the actual problem. The move it teaches — treat each reasoning step as an editable, re-runnable object — is exactly the real practice.
Related
Part of the Demystify Playgrounds. Explore the rest from the Playgrounds home.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.