Pair reverse metrics to monitor whether the proxy relationship has already distorted

What it is

A reverse metric (counter-metric) watches specifically whether the relationship between proxy and goal has already flipped. If session length rises while incomplete-task rate and voluntary-close rate rise with it, length as a proxy for “involvement” has distorted. A reverse metric is not a second success standard to be lifted, and not a guardrail that protects other goals. It asks: is this proxy still indicating the original thing. Once distortion has happened, a rise in the headline proxy should be read as an alarm, not a win.

Why it happens

The typical shape of distortion is the headline proxy starting to co-move with an observation that, if the proxy were still honest, should not worsen with it. If involvement truly increased, incompletes should fall or at least not rise; if satisfaction truly increased, complaints and unsubscribes should fall. When the headline is lifted by a shortcut, those reverse observations move first. Writing them as formal counts gives distortion a signal that does not depend on later argument. A reverse metric should be mechanistically opposite the headline and still close to the goal: not just any number that happens to be falling. It must not enter the same reward, or the team will flatten the reverse too and the watch fails. The value of distortion monitoring is timeliness: announcing success with the headline after the relationship has already broken is the most expensive misread.

Studying it

For every pressured proxy, pre-specify one or two reverse metrics and a rule “headline up and reverse worse equals distortion,” then back-test on historical shortcut events to see whether the reverse alarmed before an independent criterion. Compare product lines with and without reverse watches on miss rate of the goal criterion during headline highs. An injection also works: safely introduce a known inflate and test whether the reverse fires by the rule; if it does not, the reverse was chosen wrong or the threshold is too wide.

Where it stops holding

Reverse metrics and guardrails often sit on the same table with different jobs: a guardrail protects other goals from being stepped on by the north star; a reverse protects the meaning of the proxy itself. Writing the reverse as a target to optimize creates a new single point. Too many reverses become another overloaded table. For some headlines, worsening already co-moves with the reverse (length and failure both rise in an outage); an outage flag is needed so an incident is not read as inflation. A reverse also does not replace periodic relationship tests against an independent criterion; it is a daily distortion sentinel only.

Applying it

• Pair every rewarded proxy with one reverse metric, written as “headline up and reverse worse = distortion; success may not be announced.”
• Give the reverse no upward target and keep it out of personal evaluation.
• Drill with a known inflate; the reverse must fire, or change the observation.
• After a distortion trigger, demote the headline to watch-only, close the shortcut or change the proxy, and only then talk about performance.

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 · Q6.09.3 Using multiple proxies in combination reduces single-point gaming
• Adjacent: Q6.05 Metric manipulability · Q6.03 North star metrics
• Search terms: reverse metric · counter-metric · proxy distortion