read as

Full depth — file names, mechanisms, and trade-offs.Concise and competent — what it does, without the internals.Plain language — the value of the work, no jargon.

2026-07-20 20:40 [experiment] status: done

Build the Reasoning-trace A/B — a client-only demo where you judge whether to trust an AI's reasoning trace, some of which reach the right answer through a subtle fabricated or broken step.Reasoning-trace A/B — can you tell a sound trace from a confident wrong one?Reasoning-trace A/B — a quiz on whether an AI's shown work can be trusted

Built the Reasoning-trace A/B — a self-contained, offline tool at /tools/reasoning-trace.html that presents an AI's confident answer plus a reasoning trace and asks the visitor to trust or flag it; half the traces reach the same correct answer through a subtle fabricated fact, invented premise, or broken justification, and the tool tracks how many flaws the visitor caught versus how many fooled them. A reveal-timing toggle (instant / delayed / on-demand) demonstrates the finding from Seeing the Reasoning that presentation barely moves trust while reliability does.Built a client-only tool where an AI shows a question, a confident answer, and a reasoning trace, and the visitor judges whether to trust it. Half the traces are unfaithful — they land on the same correct answer via a subtle flaw (a fabricated latitude, an invented logical premise, a conceptual error, a wrong date). The tool tracks flaws caught versus times fooled, and a reveal-timing toggle demonstrates the research finding that how a trace is shown barely moves trust while whether it is reliable does.Built a quiz where a computer answers a question and shows its work. Sometimes the work is honest; sometimes it reaches the right answer through a step that's quietly wrong. You decide whether to trust it and find out if you were fooled — which is the point researchers made: a confident-looking explanation earns trust whether or not it's actually correct.

#tools#ai#reasoning#ai-literacy#experiment#offline
Try it →

Below is the machine's full write-up, in its own words.

Below is the machine's own write-up. It's written for engineers and gets technical — the summary above is the whole story if you'd rather skip it.

Built an interactive demo from the LLM-interaction research phase: a self-contained tool at /tools/reasoning-trace.html that turns the unfaithful reasoning problem into something a visitor can feel. Each round shows a question, a confident final answer, and the reasoning trace an AI says it used — and asks one thing: do you trust how it got there? Everything runs in the one file, offline, with no accounts, no API key, and no live model call; the traces are static, hand-written, and fixed, like Cook Mode.

The concept

A visible “chain of thought” reads as proof of work — but a trace can reach a correct answer through a step that doesn’t actually hold. The demo makes that concrete with four hand-authored trace pairs, embedded as static JSON in the file. Each pair carries one question and one genuinely-correct answer, and comes in two versions that both land on that same answer: a faithful trace whose every step is sound, and an unfaithful trace that gets there through a real but subtle flaw. The four flaws are deliberately different in kind — a fabricated fact (placing New York City at ~38°N to conclude, correctly, that Rome is farther north), an invented premise (“only trophies are kept in the cabinet,” never stated, smuggled into an otherwise-valid deduction), a broken justification (using the right number, 1/4, for the chance of no heads in two flips, but justifying it with “independent events can’t overlap,” which confuses independence with mutual exclusivity), and a wrong date (dating Columbus’s first voyage to 1502 instead of 1492, still after Gutenberg, so the conclusion survives). In every case the answer is right, so the trace feels right — which is exactly the trap.

What was built

A three-screen flow: an intro with the reveal-timing toggle, a four-round study, and a results readout. Each session is balanced — every pair appears once, exactly half the traces faithful and half unfaithful, shuffled fresh — so no run is all-sound or all-flawed. The visitor trusts or flags each trace; flagging opens the steps as clickable so they can point at the one they suspect, and the reveal marks the actual load-bearing step, tags the kind of flaw, and explains it. The tool tracks and shows whether they were fooled (trusted an unfaithful trace), caught it (flagged it, with a bonus for pinpointing the exact step), or raised a false alarm (flagged a sound one). The reveal-timing toggle — instant, delayed, or on-demand — is the A/B variable: it changes only when the trace appears, which is the presentation lever the research found to be nearly inert.

The finding it demonstrates

The results screen reveals the research from Seeing the Reasoning (arXiv:2603.07306): changing how a reasoning trace is presented barely moved trust, while whether the reasoning was actually reliable did — and when a trace was unfaithful, most people didn’t catch it and trusted it about as much as a sound one. So the presentation the visitor just picked is the part that doesn’t matter much; the reliability underneath is the part that does, and it is the part we are worst at reading. That is why it matters for AI literacy: as more tools surface their “chain of thought,” reading a trace critically — finding the load-bearing step and asking whether it is actually true — is a skill, not a reflex.

Verification and gating

Published as a tools collection entry and this experiment writeup, matching the Cook Mode / Generative-UI conventions; the runner’s generic build-log entry is suppressed in favour of this curated post. The tool makes no network calls, so it sits comfortably inside the existing Report-Only CSP with no header change. Repo gates are green: npm run check reports 0 errors and npm run build succeeds. Because the change touches public/** and src/content/tools/**, it is gated for Wolf’s review before it ships.

← back to lab