Who owns the row set at a seam?

The layer closer to the user, because it can see the cases. A model team serving a surface team runs the surface team's rows against the model, the way a vendor's FDE runs the customer's rows.

The pass rate on those rows at handoff. From then on, falling under it is a recall rather than a conversation. Vohra's requirements for quality, latency, and cost are three floors.

Because my 89% was true on the held-out set and false on traffic, and nobody had written down when to look again. The re-run is on production rows, with the no-answer count as the alarm, and it would have caught the 61% at week three instead of week six.

AI PM Handoffs Between Stack Layers: Rows, Floor, Date

How should AI PMs on different layers of the stack hand off work?

Answered by Falk Gottlob · Reviewed Nov 10, 2026 · 1 min read

THE SHORT ANSWER

With three things written down at each seam, because the shared vocabulary will not carry them. Spotify's three principal PMs (Henrik Lindström on research, Divita Vohra on the recommendation foundation model, Humberto Corona on the listener experience) wrote on November 9, 2026 that quality, evals, and shipping shift meaning with the layer each owns, and their best result this year came from agreeing the measure before the architecture; the post puts no number on any seam. A seam without a number is where a classifier of mine scored 89% held-out and 61% in production six weeks later, and where a feature of mine passed 200 real cases and failed on one slice that was a fifth of traffic. So at every handoff: a row set of real cases with agreed answers, owned by the layer closer to the user; a floor, the pass rate at handoff, under which is a recall; and a date for the re-run on production rows, with the count of no-answers as the alarm. The user's 'that is not what I meant' is the row set's intake, and its rate is the repair rate, read weekly by surface.

Write the number at the seam. The word eval crossed the table at Spotify three times and meant three things; the row set, the floor, and the date are what the word drops.

The three sections, the three seams, and my two seam failures are in Spotify's Three AI PMs Share a Vocabulary. Put a Row Set, a Floor, and a Date on Every Seam.. The per-judge card with the date on it is The Threshold Card: Tune the Judge Before You Trust the Score, and Date the Re-Tune, and the row set as the statement of work is The FDE Writes the Eval: Almost Right Has a Number.

How should AI PMs on different layers of the stack hand off work?

SOURCES

THE LONG VERSION

RELATED ANSWERS