The clock stops when somebody moves

Two numbers are on the screen and they disagree. The readout says retention is down since the event went live. Someone's sheet says it held. Nobody in the room can say where either figure came from or which players it counted. Twenty minutes go to that. The event is still running while this happens. By the time somebody volunteers to go and check, the decision you called the meeting to make has moved itself to next week.

Both numbers arrived fast. Neither arrived usable. Elapsed time is what an analysis layer sells, and the clock it gets judged on is not the one most teams measure. Teams measure from the query to the answer. The clock that decides anything runs from the thing happening in your game to a person acting on it, and everything after the answer appears is still on that clock.

AppsFlyer's State of Gaming for Marketers 2026, drawn from 9.6 thousand gaming apps and 24.8 billion installs, found that 46% of marketers' AI assistant queries were reporting and performance breakdowns. Not decisions. Requests for the figure. That is the marketing side of the building. We see the same pattern one floor down, in live ops. The calendar got heavier over the same stretch. AppMagic's live-ops data, summarized by GameDev Reports, has games in its sample averaging 73 live-ops events a month at the start of 2025 and 89 by the end. AppMagic does not publish that sample, so read it as a direction and not a population. A fifth more surface to watch, and no change in how long it takes to settle an argument about one number.

So the obvious build is the fast one. Get the readout out earlier. That build is half a system, and the missing half is the expensive one. A wrong number that arrives three days early is worse than a right one that arrives late. The late number gets argued with. The early one gets acted on.

Prediction is the cheap part. Anders Drachen and six co-authors showed in 2016 that a heuristic reading a player's first session and first week predicts short-term retention about as well as standard classification algorithms do. A usable early signal is cheap. The right to be believed is not.

In the live-ops platform we built for a top-20 mobile games studio, every figure carries per-field provenance. Open a number and see the inputs that produced it, and whether it was sourced, estimated, modeled or derived. The twenty-minute argument becomes a ten-second check, and the check happens in the room instead of in somebody's afternoon.

Forecasting is harder to police, because reality arrives late. Moloco studied 55 mobile game launches covering $1.5 billion in in-app purchase revenue. Mobile Marketing Reads, reporting on that study, pulled out the figure that should worry anyone selling a lifetime-value model. 87% of eventual top spenders in whale-driven games had not made a first purchase in month one, and by month 10 the top 5% of first-year payers were generating 81% of monthly in-app purchase revenue. Fit a forecast to month-one payers and it will miss the players who pay for the game. It will miss them smoothly, with a confidence interval that looks fine.

Nobody can wait for that money to arrive and grade the model against it. The only grading available today is history that already shipped, which is why that system's forecasts get backtested against 10 shipped events and 35 runs of real history before anybody reads one out in a meeting.

A check that quietly skipped looks exactly like a check that passed. A blank where a warning belongs is how a team learns to stop reading warnings. So every economy gate reports its own coverage, what it ran and what it lacked the inputs to run, and "we could not check this" becomes something the reader can see.

Patrick ran the June's Journey team at Wooga, 115 people across product, engineering, design, analytics and live-ops. At that size a number passes through several hands before it reaches the room, and every hand is a place a definition can shift without anybody deciding to shift it. Install date or first-session date. Reinstalls counted or not. A system does not fix that by being one more hand. It fixes it by putting the definition where the number gets used.

We will not ship a surface that reports a figure it cannot source, and we will not ship a forecast with no backtest behind it, on any date. Holding that line costs schedule. The client wants the dashboard by the quarter, and a dashboard handed over on the promised day that nobody can interrogate is a liability delivered on schedule. It is a faster route back to the same twenty-minute argument, with more authority sitting behind the wrong side of it.

An analysis layer earns its standing the way a good analyst does, by being checkable on demand by somebody who walked in wanting to disagree. A system that cannot show its work has only made the wrong answer arrive on time. The clock stops when somebody moves.