Describe what a reasoning model does differently and when the extra cost is worth it.

Keelson builds Sinew, an AI tool that reads a contract before signature and flags the clauses that carry real risk. Stanwick Components, an industrial parts manufacturer, runs every vendor and customer contract through it. Its senior contract reviewer, Bronwyn Merrion, is the one who has to trust, or double check, whatever Sinew tells her.

How to answer this, stage by stage

Nobody is grading whether you can define a reasoning model. They are grading whether you can turn "it thinks longer" into a rule that decides, cheaply, when that's actually worth paying for.

Let's learn

Every week, before Sinew existed at all, Stanwick's legal team read every incoming contract by hand. A simple mutual NDA took Bronwyn Merrion about ten minutes. A real master agreement, the kind where indemnification, a liability cap, and an insurance clause all lean on each other, took her close to three hours, cross-referencing section numbers on a legal pad as she went.

This is the work Sinew was built to take off her desk. Not the reading. The cross-referencing.

Keelson built Sinew with two speeds. Fast Lane reads a contract once and answers once, about 20 seconds, a few cents. Reasoning Lane reads it, works through the clauses that might connect to each other, checks those steps against each other, and only then answers, about three minutes, roughly ten times the cost. For a long stretch, nobody at Keelson or Stanwick could say which contracts actually got which read.

This is the actual technical difference the question is asking about. Not "smarter." Slower, on purpose, because it checks its own work.

Stanwick sends Sinew about 70 contracts a month. Roughly 56 are simple, standard NDAs and purchase orders under a set dollar value. Roughly 14 are structurally complex, master agreements with indemnification, liability caps, and insurance requirements that reference each other.

Then came the part that mattered. Reasoning Lane on every contract wasn't really the problem, and neither were the handful of times Fast Lane guessed wrong on something genuinely simple. Those were cheap mistakes. The real problem showed up the day a contract that badly needed the careful read got the fast one instead, and nothing on the screen said so.

What it costs at its worst: a reviewer who has learned to trust a clean tag stops doing her own cross-check. The one contract structurally complex enough to hide a real problem sails through with a false green light, and by the time anyone notices, it already has a signature on it. That's worse than never having built Sinew at all, because at least without it, Bronwyn would have read every page herself.

What I would leave alone: Stanwick's simple mutual NDAs and standard purchase orders, still the majority of what comes through Sinew, stay exactly as fast as they've always been. Nothing in them has a second clause to check against, so there's no reason to make Bronwyn wait three minutes for a read that a single pass already gets right.

The lesson: a setting is a promise about a document, made once. The document doesn't send an update when it stops keeping that promise.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one for why the rule had to change at all, and what nearly went out the door before it did.

Bronwyn Merrion can read a limitation-of-liability clause and tell you, in about ten seconds, whether it's been quietly gutted by a section three pages later that redefines what counts as "liability" in the first place. Nine years reviewing contracts will do that.

Before Sinew arrived, that instinct was her whole job. She read every incoming agreement herself, indemnification against liability cap against insurance, every time, because nobody else was going to.

Sinew launched at Stanwick running Reasoning Lane on everything, no toggle yet, every contract getting the slow, careful read. For six good months, that was fine. Flags came back reliable. Bronwyn started trusting a clean tag the way she'd trust a second lawyer who never got tired.

Then Keelson, fielding cost and speed complaints from bigger customers, shipped the settings page: a Fast or Reasoning default, per contract category, per account. Stanwick's operations admin, not Bronwyn, set "master supply agreement" to Fast, because ninety percent of the ones they'd signed that year really were short and standard. Nobody told legal.

The habit thinned in three small beats. First, Bronwyn noticed master agreements cleared faster than she remembered, and figured Sinew had simply gotten quicker. Second, she stopped opening the underlying document on anything Sinew cleared, the way she used to for the first six months, because it had been right every time. Third, "no risk flagged" stopped meaning "read carefully" to her at all. It just meant cleared, full stop, and she had no way to know which read had actually happened.

Same wrong contract, two different outcomes, depending on what decided the lane: a label, or the document itself.

Kestrion Forgeworks was Stanwick's first overseas precision-parts supplier, and its agreement arrived forty pages long, nearly four times Stanwick's usual template, filed under the same old category. Fast Lane. Cleared. Nothing flagged.

It was a Friday afternoon, signature due Monday. Something nagged at Bronwyn: the file felt thick for something Sinew had cleared so quickly. Out of a habit she hadn't used in months, she opened Section 9, the indemnification promise, and started checking it against everything it might quietly touch.

She spent that evening and came back Saturday morning, close to four hours in total, cross-referencing a forty-page agreement clause by clause, the exact work Sinew was supposed to have replaced, because she no longer trusted what "cleared" actually meant and had no way to check from the screen.

What she found: Section 9 promised Kestrion would cover Stanwick's costs if a defective part caused a recall. Section 14, read alone, looked like an ordinary limit, nobody pays more than the value of parts delivered in the last 90 days. Nobody had connected the two. Read together, the recall promise was worth at most about forty thousand dollars. A real recall on precision parts could run into millions. Section 22 required Kestrion to carry insurance, which sounded like a backstop, until you read that the policy only covered Kestrion's own mistakes, not a promise Kestrion made in a contract.

The old decision sat in a much smaller meeting, eighteen months earlier, when Keelson had five customers and was deciding how to let people control Reasoning Lane's cost as the company grew. Building an automatic classifier that read each document's own structure felt like a bigger engineering bet than the team was ready to make. A settings page felt clean, easy to demo, easy to explain in a sales call. Lysandra Novaczek, newer to the product team then, backed it. Every customer's contracts really were simple back then. The bet was reasonable, for exactly as long as that stayed true.

Run the same afternoon again, months later, after Lysandra's redesign shipped. Kestrion Forgeworks' contract lands in Sinew. Before Bronwyn even opens her inbox, the structural gate has already counted three connected risk clause types and routed it to Reasoning Lane. Three minutes after upload, the flag is sitting there: indemnification effectively capped by Section 14, verify against real exposure. Bronwyn checks it against the raw text herself, about ten minutes, confirms it, and flags it to procurement that Monday morning, days ahead of signature instead of hours before it. Her legal pad stays in the drawer.

What I'd tell myself, sitting in that first settings meeting: we built customers a knob because a knob was easy to explain in a demo. We should have built the document a voice, because the document is the only thing that actually knows what's in it.

SPARK, the five calls behind the reasoning switch

Not a script for sounding technical. SPARK is what forces you to name the one cheap check that decides, before Sinew spends a dollar thinking, whether this contract actually needs it.

The whole page in one picture. A quick read for what doesn't connect. A slow, checked read for what does.

The anchor, close up. Not a setting a customer picks. A count, taken fresh off the document, every single time.

The flag has to show its work, which sections it chained through, or a reviewer is just trusting a new kind of label.

The recap, one line per letter: situation is a reviewer whose "cleared" tag stopped meaning one consistent thing, payoff is trust that's actually earned by a structural check instead of a category label, anchor is the cross-reference gate itself, tuned against real labeled contracts rather than a launch-day guess, risk is either extreme costing Stanwick on a different day, and keep out draws the line at what has nothing to chain through, never at "when in doubt, run it anyway."

And if you want to be sure it really works, try it somewhere else

Same five letters, a hospital's overnight reading room instead of a legal team, and this time the missed connection isn't a liability cap. It's a nodule that grew between two scans.

Solmark builds Nightread, an AI tool that pre-reads radiology reports overnight for Kellenbeck Health, a hospital network, and flags reads that need a second look before a radiologist signs off in the morning. Thoraya Duskwright runs the overnight reading queue, roughly 90 reports a shift.

Same shaped gate, a report instead of a contract. Count what actually connects, then decide the lane.

Mapped onto SPARK: the situation is Thoraya's overnight queue, most reports routine, a few genuinely dense, a finding's size mentioned in one section, a prior scan's measurement mentioned in another. The payoff is every radiologist trusting a "no second look needed" tag only when it's earned. The anchor is the same shaped gate, count how many findings a report correlates against another section or a prior scan, route those to Reasoning Lane so the model actually checks whether the numbers agree, otherwise Fast Lane. The risk is the same two-sided cost: over-trigger and radiologists lose a night to routine studies waiting on a slow read; under-trigger and a nodule that quietly grew gets the all-clear. Keep out draws the same line: a routine single-finding chest film, or a check for a missing required field, never needs the slow pass.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the anchor, say the rule, and what it protects against.
Cost: no budget for a new classifier this quarter. Have the hospital's existing report software run the same structural count as a plain rule, and gate the existing reasoning call behind it, no new model spend.
The model got better, for real: say Nightread's single-pass accuracy doubles overnight. The gate still matters, because a better single-pass model still only reads a report once. It can still miss a correlation across two sections it never re-reads.

Where people run it wrong.
They route every long report to reasoning mode because more pages feels riskier, instead of counting what actually cross-references.
They tune the cut line once at launch and never recheck it as report templates change hospital to hospital.
They let a busy technologist override the gate for speed, which quietly recreates the exact 2am judgment call the rule was built to remove.

How to use it live. Before answering a "what does the expensive mode do, and when's it worth it" question cold, ask yourself one thing: what's the cheap signal about this specific request, checkable before you spend a dollar, that tells you whether it actually needs multiple steps. Naming that signal first is usually exactly what the question is listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0