Rewards must strongly relate to the target behaviour

What it is

An incentive system must reward the target behaviour itself, not a proxy metric or whatever happens to be measurable. When reward-behaviour contingency is weak, participants optimise the rewarded behaviour rather than the target behaviour—points for logins breed login-farming, points for posts breed filler posts. The incentive system precisely manufactures what it measures, and what it measures drifts from the real goal. The first design principle of incentive systems: reward exactly the behaviour you want, and only it.

Why it happens

Weak coupling fails through surrogate-metric substitution. Target behaviours (students learning, patients recovering, high code quality) are hard to measure directly, so incentive systems fall back on proxies (time spent, check-in counts, lines of code)—and participants optimise measured behaviour instantly and creatively: every gap between proxy and real goal gets filled with behaviour (grinding hours without engaging, checking in for check-in's sake, writing redundant code for line counts). Goodhart's law states it: "when a measure becomes a target, it ceases to be a good measure." Strong coupling anchors rewards on verifiable outcomes of the target behaviour itself (comprehension assessments, functional recovery indicators, review-passed merges)—the smaller the gap, the smaller the substitution space. Coupling strength also shapes crowding-out's form: rewards strongly contingent on behavioural outcomes act as the behaviour's confirmation; rewards coupled to irrelevant proxies reframe the behaviour as "serving the metric," damaging autonomy more deeply.

Where it stops holding

Perfect strong coupling is often unattainable—real outcomes carry time lags (education's effects appear years later), external dependencies (recovery depends on illness), or individual attribution problems (team production). Design becomes balancing under imperfect measurement: multi-metric combinations (time + assessment + application behaviour) resist manipulation better than single metrics but raise system complexity and participants' comprehension cost. Coupling's ethical edge sharpens under pressure: strong coupling plus high-value rewards induces cheating (metric fraud, see the leaderboard cards), while weak coupling plus low-value rewards is safe but weakly motivating—coupling strength, reward value, and anti-gaming mechanisms must be designed as one parameter set. "What to reward" is also a question of power: whoever defines which behaviour deserves reward installs their values into the system, ethically weighty in education and healthcare (rewarding obedience or rewarding questioning).

Applying it

• Design incentives backwards from verifiable outcomes of the target behaviour: first define "how we confirm the goal is met," then point rewards at that confirmation signal—never start from "what is easy to measure."
• Use multi-metric combinations with counterweights: each proxy gets a balancing metric (study time with comprehension tests, post volume with acceptance rate) so gaming one metric cannot raise the composite score.
• Verification: audit each rewarded metric's growth against its counterweight metric after launch—a rewarded metric rising while its counterweight stays flat or falls is direct evidence of surrogate optimisation (metric gaming), and the gap between metric and goal needs narrowing.

Related

• Same group: W11.02.1 Points can erode existing intrinsic motivation · W11.02.2 Behaviour falls back when rewards stop
• Nearby: W11.03 Leaderboards' side effects · U1.01 Metric design and misleading indicators · W11.02 Crowding out intrinsic motivation
• Search terms: goodhart's law · reward contingency · surrogate metrics · incentive alignment