Notes on holding, not exiting.

Last spring, a strategic partner offered Aperture Labs a distribution deal that came with a condition: triple headcount in eighteen months to hit a volume target the partner needed for their own board deck. The revenue math worked. We said no.

The condition wasn't really about Aperture's customers — eleven regional insurers who need field-level accuracy, not faster onboarding of insurers who don't exist yet. Growing headcount to serve a volume target no customer had asked for would have meant hiring against a plan rather than against a need, which is how good product teams quietly become bad ones.

We're not against growth. Aperture added three insurers this year and will likely add more next year. We're against growth that's sized to someone else's timeline instead of the business's own. That distinction is the entire reason a permanent-capital structure exists: nobody outside the company gets a vote on how fast it has to move.

When we acquired Fieldstock in 2019, the first email we got from a well-meaning advisor suggested "unifying the visual identity" across our commerce companies to build recognition. We still get a version of that email every year or two.

We've never done it, and we don't plan to. Harlow & Finch's customers don't care that it shares an owner with Fieldstock — they'd care a great deal if the specialty-coffee retailer they trust suddenly looked like it belonged to a conglomerate. A shared visual system is built for the parent company's convenience, not the customer's.

The instinct to unify comes from treating a registry like a product portfolio, where consistency reduces a buyer's confusion between SKUs. A holding company isn't selling SKUs. It's holding relationships that took years to earn, in categories that have nothing to do with one another. Sameness would cost us the one thing each acquisition was worth buying for.

"Applied AI" gets used to describe almost anything with a model in it, which makes the term nearly useless. Inside KVIAZ it means something specific: a single, named, repeatable task that a customer currently pays a person to do, done by a model to a measurable accuracy bar, with a clear cost of being wrong.

Aperture Labs' bar is 99.6% field-level accuracy on claims documents, audited monthly against a held-out sample. Fathom's bar is caption latency and word-error rate against FCC accessibility requirements. Neither company describes its product as "AI-powered" to customers — they describe the job it replaces and the accuracy it hits.

This filter rules out a lot of pitches that land in our inbox: general copilots, chat interfaces bolted onto existing software, anything whose value proposition is "faster" without a stated bar for "correct." We'd rather hold two narrow companies that hit their numbers than five broad ones that can't state what their number is.

Almanac Quarterly's publisher called Elin directly, not through a banker, three weeks before the review journal's parent company planned to shut it down. There was no process, no data room, and no time for the kind of diligence we'd normally run.

We closed the acquisition in nineteen days. The terms were simple: KVIAZ would fund the next four issues immediately and commit, in writing, to keeping the journal quarterly and in print — the two changes a more conventional buyer had already proposed making, in the form of a move to digital-only and an annual anthology instead of four issues.

Almanac's subscriber base is small and specific: about 4,100 people who want fiction in translation reviewed at length, on paper, four times a year. A faster or cheaper version of that product isn't a better version of it. It's a different, worse product wearing the same name — which is exactly the trade we bought the magazine to avoid making.