Critique a design that surfaces a chain of thought to end users.

Cordage Capital is a robo-advisor. When it recommends a portfolio rebalance, a panel labeled "See how I reasoned through this" shows the model's raw reasoning text. Halyna Marchetti was brought in as a trust researcher to review the panel before its rollout widened to every client.

How to answer this, stage by stage

Nobody's grading whether you know chain-of-thought panels exist. They're grading whether you can spot the exact assumption baked into this one that nobody actually checked.

Let's learn

Cordage Capital's robo-advisor recommends portfolio rebalances. When it added a "See how I reasoned through this" panel, showing the model's raw reasoning text, it followed a new model release built to write out its own deliberation step by step.

Before the panel, clients saw one plain-language reason line under a recommendation. About 6% of them still called a human advisor asking "why," each call costing roughly twelve minutes of advisor time. After the panel shipped, those calls dropped to 2%.

The turn: a model reversing itself mid-reasoning isn't actually the problem. Models revise their thinking constantly on the way to a good answer, and that revision is often exactly what makes the final recommendation solid. The real problem is showing that revision raw, with no label telling the reader they're looking at a normal correction rather than a system that can't make up its mind.

At its worst: a client reads a trace where the model leans toward selling a stock, then writes "actually, let me reconsider given the client's stated risk tolerance" and reverses to holding it. She screenshots the contradiction and posts it publicly with the caption "this AI can't make up its mind about my retirement money." It spreads widely, and trust drops broadly even though the final recommendation was the better one.

What I would leave alone: for internal advisor use, the fully raw, unflagged trace is exactly right as-is. Advisors are trained to read a model second-guessing itself as diligence, not instability, so they need nothing softened.

The lesson: "we're showing our reasoning" is not one claim. It's at least three, faithful, versioned, and flagged, and a panel that satisfies none of them is still technically a reasoning panel. It's just not a trustworthy one.

Now here is the same thing as a story

The short version above is what you'd say defending this critique to Cordage's product leadership. Read this one for how the gap actually got found.

Halyna Marchetti reviews AI features before they roll out widely, brought in specifically because she's not on the team that built them. She reads an interface the way a proofreader reads a contract, looking for the sentence someone assumed nobody would test.

For the first two months after the reasoning panel launched to a small pilot group, it looked like a clean win. "Why" calls dropped. The product team called it a success in a leadership update, pointing at the falling call volume as proof.

Four steps, and nobody had checked what actually happens between the second box and the third.

The trigger wasn't anything that happened at Cordage. A rival robo-advisor's chain-of-thought panel went viral for a different reason: a screenshot showing its model reasoning "the client is 61, so I should recommend more conservative bonds," then flatly contradicting that logic two lines later. Financial commentators mocked it for days.

Halyna watched the story spread and thought, quietly, "we have the exact same panel."

Cordage had never actually confirmed which of these two panels it was showing.

She pulled Cordage's own reasoning logs and found the same pattern quietly present, just not yet screenshotted. Traces containing a mid-reasoning reversal, "actually, let me reconsider," had been climbing for weeks, unnoticed because the complaint category "AI contradicts itself" hadn't existed as a support tag until someone finally added it.

Here's the decision I'd take back. We shipped the raw trace with no version pin and no revision flag, because the whole feature was rushed out to answer a competitor's release. That made sense for a fast pilot. It stopped making sense the moment real client trust was riding on it.

The panel was built like a diary and shown like a script, without deciding which one it actually was.

I'd add one plain label, "reconsidered," wherever the model's own trace shows a reversal, plus a version pin on every trace so a disputed recommendation from months ago can be checked against the model that actually made it.

Four parts, and Cordage's live panel had none of them when it shipped.

Replay the same stock-hold recommendation under the new design: the client sees "reconsidered: kept this stock after weighing your stated risk tolerance more heavily," instead of a raw, unlabeled reversal. Nothing about the actual advice changes. What changes is whether it reads as a system thinking carefully or a system that can't decide.

We built the panel fast to answer a competitor's move, and treated "shows more of the model" as automatically the safer, more honest choice. It took watching a rival get mocked publicly, then finding the exact same fault line quietly growing in our own logs, to see that raw and honest were never the same promise.

AUDIT, reading a reasoning panel like evidenceNot a style critique. AUDIT is what separates a reasoning panel that's actually checkable from one that only looks like it is.

The goal corner isn't raw, and isn't cleaned. It's faithful and flagged, at the same time.

The recap, one line per letter: ask is an engineer shipping fast with no research sign-off, uncover is the unconfirmed gap between real reasoning and after-the-fact narration, demand is the missing version pin, isolate is the missing revision flag, and test is splitting real clients between raw and plain versions to measure actual decisions.

And if you want to be sure it really works, try it somewhere elseSame five letters, a university admissions office instead of a robo-advisor. A different building, and the reversal that scares people is about a person, not a portfolio.

A university's admissions office uses an assistant that drafts a reasoning summary for why an applicant was waitlisted, shown to admissions staff before a final decision. Mapped onto AUDIT: ask is whether an admissions officer or only an engineering intern reviewed the panel before staff started relying on it daily. Uncover is whether the summary reflects the model's actual weighting of test scores, essays, and extracurriculars, or a separate, cleaned explanation generated after the ranking was already set. Demand is the version pin, since admissions cycles run for months and a challenged decision needs to be traceable to the exact model that produced it. Isolate is the missing flag for cases where the model's stated reasoning shifted between an early read of the file and its final rank. Test is pulling ten waitlisted files and checking, by hand, whether the summary text actually matches the features that moved the ranking, not just whether it sounds plausible.

A robo-advisor and an admissions office share almost nothing else. They share exactly these three gaps.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "raw isn't honest by default, check faithfulness, versioning, and revision flags," and stop.
Cost: there's no engineering time to rebuild the panel this quarter. Start with the cheapest fix: label mid-reasoning reversals with one word, "reconsidered," without touching anything else.
The model gets better, for real: if the new model version reasons more clearly overall, that's still not a reason to skip a faithfulness check. A clearer-sounding trace can still be an after-the-fact story instead of the real path.

Where people run it wrong.
They treat a technical-looking panel as automatically more trustworthy than a plain-language one.
They celebrate a drop in "why" calls without checking what new complaint category might be quietly rising instead.
They ship a reasoning feature to match a competitor's announcement before anyone tests what it actually does to real decisions.

How to use it live. When someone hands you a chain-of-thought panel to critique, don't start with the wording. Ask whether it's faithful to the real computation, whether it's pinned to a version, and whether a reversal gets labeled. If any of the three is missing, say so before anything else.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0