From checklist to instinct
In b1 you met the six-point fair-metric checklist and in b2 you wrote objectives and key results. Tonight the two meet: how do you know a key result is one you can fairly grade a team on, rather than one they will quietly game? The answer is a law almost as reliable as gravity - Goodhart's Law - and a habit of scoring every metric before you deploy it. You will end the session having driven a real lab that scores metrics and simulates the gaming for you.
When a measure becomes a target 6 min live
Goodhart's Law: "When a measure becomes a target, it ceases to be a good measure." A metric works as a quiet signal because it correlates with something you care about. Turn it into a target - especially one tied to reward - and people optimise the metric directly, snapping the correlation. The measure and the goal drift apart. This drift has a name: surrogation.
LiveThe classics - and why paired metrics resist gaming3 min▶
- The cobra effect - a bounty on dead cobras bred cobra farms. The metric (cobras handed in) rose; the goal (fewer cobras) reversed.
- Lines of code - reward it and engineers write more, worse code. AI code generation makes it meaningless entirely.
- Tickets closed - reward it and agents close-and-reopen and rush customers, while satisfaction quietly falls.
- The defence: measure as a system. Pair a quantity metric with a quality one (deploys with change-failure rate; tickets closed with CSAT). You cannot juice one without the other catching you. DORA works precisely because its four metrics hold each other honest.
Design for the gaming before you deploy. Assume any metric you announce will be optimised the laziest way that moves the number. Ask: "what is the cheapest way to make this go up without doing the real work?" If that cheap path exists, either pair the metric or do not grade on it.
Score every metric before you set it 4 min live
The checklist from b1 becomes a fast score. Rate the metric on outcome-vs-activity, controllability, game-resistance, and holism. A high score means it is fair to grade on; a low score means watch it as context, or pair it, or drop it. The lab below does this scoring in front of you.
The Fair-Metric Lab ★ live scoring + Goodhart simulator
This is a real lab, not a slideshow. Pick a Cartwheel metric and it scores its fairness from the metric's actual attributes. Then flip "tie this metric to individual pay" and watch the deterministic Goodhart simulation: for a gameable metric, the measured number races to target while the true outcome it stood for collapses. For a fair metric, the two stay together even under pressure.
Turn a gameable metric into a fair one ★ 8 min · do your own
Take a low-scoring metric from the lab and repair it.
Pick a metric that scored low (lines of code, hours logged, tickets closed) and name the cheapest way to game it.
Find the outcome it was really a proxy for (valuable software, real contribution, satisfied customers).
Replace or pair it so gaming the number would require actually doing the work - then re-score.
Pairing beats policing. You will never out-police a gameable metric; people are creative under reward pressure. The durable fix is structural: pair quantity with quality so the only way to move the number is to do the real thing. That is why "deploys per week" alone is dangerous but "deploys per week with change-failure under 15%" is safe.
Try it yourself - this week ◐ 30-45 min total
- Score three of your team's key results in the lab's four dimensions. Which are fair to grade on?
- For each low scorer, write the cheapest way someone could game it without doing the real work.
- Pair or replace one gameable metric so the number cannot move without the outcome moving.
- Find one metric currently tied to pay or bonus that would fail the lab. Raise it with your leadership.
- Optional: read the Splunk or CNA explainer on Goodhart's Law and note one example from your own org.
Sources covered (cited)
Taught from the measurement-integrity canon. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · What does Goodhart's Law predict?
A metric works as a signal because it correlates with what you care about. Make it a rewarded target and people optimise the proxy directly, snapping the correlation - surrogation.
2 · What is the durable defence against a gameable metric?
You cannot out-police creativity under reward pressure. Pairing quantity with quality (deploys + change-failure, tickets + CSAT) makes gaming require doing the real work.
3 · In the lab, a gameable metric tied to pay shows...
That widening gap between measured and true is surrogation made visible, and the core argument for keeping OKR scores decoupled from compensation.