How do you assess a vendor's model dependency and what happens if their provider changes terms?

Norrsjö Fisheries is a coastal fishing co-op sorting and tagging catch for quota reporting. Solveig Mattila runs fleet operations there. Reeflex sells AI that photographs a catch bin and identifies the species inside, built on top of a foundation vision model from Vesparion.

How to answer this, stage by stage

Nobody is scoring whether you know what a foundation model is. They're scoring whether you can say what breaks when the company behind your vendor's vendor changes its mind.

Let's learn

Here is what happens when a vendor's own upstream cost quietly becomes a customer's problem.

Before Reeflex, dockside sorters at Norrsjö identified species by eye for every catch bin, about three minutes each, catching most quota-reporting errors but occasionally mixing up two similar-looking species under a rushed evening unload. Reeflex photographed every bin and tagged species in under a second, agreeing with the sorters' own calls closely enough that within a few months, sorters stopped double-checking any bin Reeflex marked confident.

Nobody at Norrsjö had really looked at the middle box until the quota showed up.

Here's the turn: the extra trust in Reeflex was never the real risk. The risk was that Reeflex's own AI runs on Vesparion, a separate foundation model provider, and Norrsjö's contract with Reeflex was signed as an unlimited-scan flat fee tied silently to Vesparion's pricing staying where it was. That made sense when Vesparion's per-image price had been flat for two years. It stopped making sense the day Vesparion tripled it, citing rising compute costs.

At its worst, a silent dependency like this doesn't just cost a few misclassified bins. It risks the co-op's quota compliance report itself, since regulators expect every bin classified with the same rigor, model-assisted or not.

What I would leave alone: Reeflex's basic bin-photo capture and storage doesn't need this scrutiny. That part runs on Norrsjö's own hardware and isn't touched by anything happening upstream at Vesparion.

The lesson: a vendor's price staying the same tells you nothing about whether their own costs are staying the same. The real question was never "is Reeflex reliable." It was "what happens to us the day the company behind Reeflex isn't."

Now here is the same thing as a story

The short version above is what you'd say to Norrsjö's board. Read this one for how the quota actually reached the dock, one label at a time.

Solveig Mattila had run fleet operations at Norrsjö for nine years, sharp enough to call a species from the shape of a tail fin at ten feet in bad light. When Reeflex went live, its tags agreed with her own calls closely enough that within four months her sorters had stopped opening the confident ones by hand, running every bin, easy and hard alike, straight through the AI check on the label printer at each station.

Sorters didn't scan a little less across the board. They flipped to scanning only the hard calls, and left the easy ones to memory.

Vesparion posted a price change one Tuesday: a near tripling of its per-image inference cost, citing rising compute demand across its whole customer base. Reeflex, absorbing the hit across its entire book of customers, quietly capped Norrsjö's account at eight thousand scans a month, down from an unlimited plan nobody had ever needed to think about. No renegotiation call came first. The quota simply appeared in the dashboard one morning.

Solveig's sorters, watching their monthly allotment shrink fast, made the only choice that felt sensible under pressure: save the scans for the bins where they genuinely wanted a second opinion, the messy mixed bycatch hauls, and skip the AI check on hauls that looked obviously like one species. It felt like the responsible move. It was also exactly backward from what the quota needed.

Within three weeks, an internal audit of tagged bins found the misclassification rate had risen from a steady three percent to nine, entirely on species Reeflex had once handled well. Two of those misclassified bins had already gone out under the wrong species code on a quota report before anyone noticed the pattern.

This is the quiet rule sorters built for themselves. Nobody wrote it down, and nobody meant for it to raise the error rate.

Vesparion's move was a number on an invoice. What reached the dock was a hard line with no warning at all.

FLIPS, in one screenNot a longer contract clause. FLIPS is what finds the exact habit that snapped when a dependency two layers up changed its terms.

This is what "model dependency" actually means, drawn out instead of defined.

The recap, one line per letter: find is Solveig, nine years reading a catch by eye, locate is sorters trusting every confident tag without a second look, identify is the flip from scanning everything to rationing toward the hardest calls, pinpoint is the flat-fee contract silently exposed to Vesparion's own pricing, and show is a notice-window clause turning a silent quota into a negotiated one.

The gap between the second and fourth marks is the entire cost of the missing clause.

And if you want to be sure it really works, try it somewhere elseSame five letters, a ticketing platform's fraud-detection vendor instead of a fishing co-op. A different flip family, a workaround instead of rationing, breaks the second story.

Gatefold Live runs Sentrigate, an AI fraud-risk score for ticket purchases, built on a foundation model from Combray Labs. Mapped onto FLIPS: find is Deshawn Purnell, Gatefold's platform reliability lead, locate is his team trusting Sentrigate's live risk score without a manual second check, identify is a different flip family this time, a workaround flip: when Combray Labs deprecated the model version Sentrigate ran on, the migration made risk scores noisier for weeks, and Deshawn's team, with no official way to compare old scores to new ones, started quietly exporting flagged transactions to a personal spreadsheet to cross-check by hand.

The old decision here isn't a pricing clause, it's an absent one: Gatefold's integration with Sentrigate never stored a local copy of past risk scores or which model version produced them. That made sense when Sentrigate's underlying model had been stable for two years. It stopped making sense the week Combray Labs forced a migration and nobody at Gatefold could tell whether a given score reflected the old model or the new one.

Gatefold built the same kind of private fix Norrsjö's sorters did. Neither one was told to.

Swap the trigger and it still runs.
Speed: an interviewer caps you at a minute. Say "ask which foundation model they run on, and what happens to you if that provider's terms change," and stop.
Cost: there's no leverage to demand a notice-window clause from a vendor this small. Say so honestly, and instead track your own error rate by case difficulty from day one, so a silent ration shows up fast even without a contractual guarantee.
The provider gets better, for real: if Vesparion's next version genuinely improves accuracy at the same price, that's still a version change worth a notice clause, since a better model is still a different one underneath your vendor.

Where people run it wrong.
They evaluate the vendor's own reliability and stop there, never asking what the vendor itself depends on.
They assume a stable price means a stable service, missing that a cost hike can arrive as a quota instead of an invoice line.
They watch one overall error rate and miss that rationing toward hard cases raises it without the model changing at all.

How to use it live. If an interviewer pushes for the mechanism, say it plainly: ask what happens two layers up, not just one, because the company you signed with is rarely the only company whose decisions reach your desk.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0