The distribution is not the curve you were handed
Most differentiation systems fail at the first step, before any money moves: they assume the shape of performance instead of measuring it. The forced ranking imposes a normal curve — a fixed share of stars, a fat middle, a mandated bottom — onto a population whose real output almost never distributes that way. Knowledge-work performance is typically right-skewed and long-tailed: a small number of people produce a disproportionate share of value, the middle is genuinely tight, and the bottom is thinner than the model demands. Uniform raises make the opposite error. They assume no shape at all, treating a flat percentage as fair because it is equal. Both are misallocations because both substitute an assumption for an observation. The uniform raise overpays the bottom decile and starves the top; the forced curve manufactures losers to fill a quota and taxes the middle to feed it.
Differentiation is a signal, and signals get gamed
The mechanism that makes differentiation work is also the mechanism that makes it dangerous. A meaningful pay gap between a high and an average performer is information — it tells the organization, and the individual, what the firm actually values and will fund. But the moment that signal carries large stakes, the inputs to the signal come under pressure. Managers learn to advocate, ratings inflate, and the measured distribution drifts toward whatever the system rewards rather than whatever the business needs. This is the central trade-off of the People Lens: the sharper the differentiation, the stronger the incentive to corrupt the measurement that drives it. A flat system is uninformative but hard to game; a steep system is informative but invites capture. The job is not to pick a slope. It is to build a measurement credible enough to survive the stakes you attach to it.
The failure mode is calibration theater
The most common way this breaks is not too little differentiation but counterfeit differentiation. Calibration meetings convene, distributions get nudged into compliance, and the ranking that emerges reflects managerial negotiating power, recency, and visibility far more than contribution. The outputs look rigorous — every number defends a budget — but the inputs were never validated against anything real. The damage compounds because the system is self-confirming: this year's rating shapes next year's assignments, which shape next year's measured output. A defensible process does the unglamorous work the theater skips:
- Separate the assessment of contribution from the allocation of dollars, and conduct them in that order, so the budget does not silently author the ratings.
- Calibrate against observable artifacts — shipped work, decisions owned, outcomes attributable — rather than against the shape of the curve.
- Test ratings for adverse-impact and manager-leniency patterns before they harden into pay, and treat a clean distribution as a hypothesis to interrogate, not a result to celebrate.
The decision: spend your differentiation budget where the variance is real
Differentiation is not free, and its cost is not the money — it is the trust you draw down every time someone perceives the gap as arbitrary. That makes it a budget to spend deliberately, not a policy to apply uniformly. Concentrate it where genuine performance variance is largest and most observable, and where the work is loosely coupled enough that one person's output is legible on its own. Compress it where output is interdependent, where measurement is noisy, or where the difference between the 50th and 70th percentile is real but unmeasurable at the precision your dollars imply. The defensible answer is therefore deliberately uneven: steep where you can see, flat where you cannot, and honest about which is which. A leader who differentiates everywhere with false precision and one who refuses to differentiate anywhere are making the same mistake — letting the convenience of a uniform rule stand in for the harder duty of matching the decision to the distribution actually in front of them.