Using multiple proxies in combination reduces single-point gaming

What it is

Requiring several proxies with different seams to move together raises the cost of targeted inflation. That is multiple-proxy triangulation. Duration can be lifted by autoplay, but if task completion and next-day voluntary opens must rise with it, autoplay is not enough. Triangulation is not averaging the numbers into a new single point. It keeps each seam so that one shortcut can hardly fool them all at once.

Why it happens

Each proxy’s definitional seam points at different irrelevant behavior. One seam eats a pre-checked box, another eats survey timing, a third eats automatic continuation. To walk through three seams at once, the interface has to do several unrelated things, whose cost approaches actually improving the goal, so the relative return to gaming falls. Triangulation works only if the proxies are not collinear and the seams do not overlap: three numbers derived from the same clickstream event look like three, and the attack surface is still one. Triangulation also changes exploration: design has to find changes that move different observations together, and those changes are more likely to pass through the goal. A weighted average folds triangulation back into a single point, because the gamer only has to optimize that sum.

Studying it

Estimate correlation and shared event dependence among candidate proxies; a set that is highly correlated or shares one event cannot count as triangulation. Red-team the weakest member of the set and watch whether the others rise with it. If they stay put, triangulation is working; if they all rise in step, the set is still a single point. Compare the share of shortcut-type levers among shipped changes under a “single-proxy gate” versus a “conjunction of proxies” gate.

Where it stops holding

Too many proxies return overload: teams cannot act, or privately watch only the easiest one to inflate, and triangulation exists in name. Two or three with clearly different seams is usually enough. Triangulation cannot stop high-quality gaming that improves every proxy at once, such as wholesale fabrication or wholesale pressure to give good ratings; that needs audit and reverse observation. Treating triangulation as multi-indicator construct validation asks more of it than it can carry; it lowers single-point gaming, and does not prove the construct is fully covered.

Applying it

• Write the ship gate as a conjunction: at least two proxies with different seams move together, and they must not be averaged in advance.
• When choosing proxies, draw the seams first; overlapping seams do not enter the set.
• Red-team only the weakest member; the others should stay put; if they follow, rebuild the set.
• Do not crush the triangulation back into a single composite for evaluation.

Related

• Same group: Q6.09.1 Once a metric is used for evaluation, teams optimize the metric rather than the goal it represents · Q6.09.2 A single proxy is easy to inflate without improving real experience · 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: multiple-proxy triangulation · metric conjunction · gaming resistance