A single proxy is easy to inflate without improving real experience

What it is

When only one proxy is guarded, the attack surface is concentrated: find the path that lifts that proxy without passing through real experience. That is single-proxy gaming. Make “I like this” the only button that dismisses a dialog, and the like rate rises while attitude does not improve. Count autoplay as watch time, and duration rises while watching did not happen. Singleness makes targeting cheap—study the definitional seam of one number; there is no need to fool a cluster of observations that constrain one another.

Why it happens

A proxy is a narrow cut into the goal. The narrower the cut, the easier it is to design irrelevant behavior that fits: one click, one default, one choice of timing. Real experience is wide; improving discovery, understanding, completion, and later evaluation together is expensive. Rational inflation takes the narrow cut. A single proxy is also predictable: the team knows review watches only this column, so design exploration collapses into “how to move this column,” and exploration itself is shaped by the definitional seam. After a successful inflate, the proxy looks more like the goal—because it rose—so there is even less appetite for a second observation, and the single point locks in.

Studying it

Red-team the live single proxy: list interface changes that would lift it without improving the goal, ship the safe ones to a small slice, and measure whether proxy and an independent criterion split. Larger splits mean a more brittle single point. In history, intervals where the proxy keeps hitting highs while task success, complaints, and voluntary return stay flat or reverse are candidates that inflation has already happened. Definition audits help too: denominator exclusions, defaults, and automatic continuation usually mean the single proxy has already been opened.

Where it stops holding

Some proxies have almost no seam, such as crash rate or payment failure; targeted inflation then requires fabricating data, which is a different problem. A single proxy can serve as reconnaissance in an exploratory period, provided it does not enter the success announcement. Swapping one single proxy for another is only a move, not a reduction in targeting. Inflation is not always malice: autoplay may be intended as “continuous experience,” and still inflates duration.

Applying it

• A proxy on the ship gate must survive at least one drill of “lift it without improving the goal”; if the drill works, it must not be the gate on its own.
• Do not make the only path for dismiss, skip, or continue be the proxy event itself.
• When a proxy hits a new high, sit an unrewarded criterion beside it; if the criterion does not move, treat it as inflation, not as an experience win.
• On finding a definitional seam (pre-checked defaults, automatic continuation, excluding difficult users), close it immediately rather than arguing motive first.

Related

• Same group: Q6.09.1 Once a metric is used for evaluation, teams optimize the metric rather than the goal it represents · Q6.09.3 Using multiple proxies in combination reduces single-point gaming · Q6.09.4 Pair reverse metrics to monitor whether the proxy relationship has already distorted
• Adjacent: Q6.05 Metric manipulability · Q6.01 Classification frameworks
• Search terms: single-proxy gaming · proxy inflation · metric loophole