Split Decision Beta

The decision

One global daily phrase was serving two incompatible jobs.

The single daily phrase was the identity of the solo game, where there's exactly one puzzle a day and everybody gets the same one. It was also where team puzzles came from. That worked right up until somebody joined a second team, because the moment they've seen the phrase once, the information asymmetry the whole co-op mechanic runs on is gone: they already know the answer their teammate is working toward. So I decoupled team puzzles from the solo daily.

On metrics, the more useful part is what pressure-testing changed. My North Star was Recurring Co-op Duo-Days, which counted every team-day where both members played their assigned word. On paper it's the right metric, because it only moves when co-op actually delivers.

The problem is that computing it means recognizing the same two people tomorrow, and identity in the app is per-device with no recovery. Somebody clears their data or picks up a second phone or reinstalls, and they come back as a new player, which quietly splits a returning duo into two identities. The bias runs downward, it runs hardest on the engaged pairs the metric exists to reward, and it's invisible, because you can't tell "this duo churned" from "this person got a new phone." A number that's most wrong exactly when you're most successful is worse than no number, so I declared it uncomputable until identity recovery ships.

What I put in its place for launch were two things I could measure inside a single game: how often a player spends a Signal on their teammate's word, and how often a Rescue gets claimed when someone is stuck. Both of those need both people present and paying attention in the same session, which is exactly why they survive the identity churn that broke Duo-Days.

Two smaller ones. I derived the teams-per-user cap from an actual comprehension test instead of asserting it, because I wasn't willing to ship a possibly-confusing product on arithmetic nobody had run. And when I caught myself describing a change as "only two lines of code," I split the claim in two, because the genuinely risky part, which was cross-surface dedupe, had been getting laundered by the cheap headline.