Measuring Performance

The fork everyone gets wrong

Measurement is where performance management actually lives or dies, and it's where the field is most confused. Faced with bad numbers, organizations pick one of two doors, and both are wrong.

Door one: measure more. Add metrics, build the scorecard, cascade KPIs down every level. The result, in Meyer's words, is firms "swamped with measures" — fifty to sixty top-level ones is routine — while the people who run them rate them "poor" and "do not experience [the constant changes] as improvements."¹ More measurement doesn't produce more insight; past a point it produces noise, gaming, and burnout.

Door two: stop measuring. This is the fashionable move — the one Harvard Business Review now documents as a wholesale retreat from ratings.² But watch where it lands: having thrown out the number, reformers fall back on the manager's unaided judgment. And we know exactly what that is. The largest studies of people rating people find about 60% of the variance in a rating is about the rater, not the rated — the idiosyncratic rater effect — and it gets worse with more elaborate scales, not better.³ It's the more general truth Kahneman, Sibony, and Sunstein call noise: "wherever there is judgment, there is noise — and more of it than you think," including occasion noise, where the same rater judges the same case differently on different days.⁴ "Manager's gut" isn't the cure for a broken thermometer; it's a thermometer with no markings.

Performix takes neither door. The answer to bad measurement is better measurement — fewer things, measured as latent constructs, with honest uncertainty, justified by the decision they inform, and pointed at the performer's own control rather than the org's. That is what "psychometric-first" means, and it's the part of the product that nothing else in this literature does.

Principle 1 — Measure what's worth measuring

Most measurement programs never ask the prior question: is this worth measuring at all? The discipline that does is value of information — Hubbard's central move in How to Measure Anything: a measurement is worth its cost only if it would change a decision, and you can compute, before you measure, how much reducing that uncertainty is worth (expected value of information).⁵ Marr arrives at the same place from the management side with Key Performance Questions — design every metric to answer a specific question you actually have, not to fill a dashboard cell; the goal is "management by insights, not management by numbers."⁶

This is why Performix doesn't open with a battery of surveys. It starts at the CAMS minimum, then narrows: each next question is chosen because the prior answers make it the one most likely to move the diagnosis, and deeper instrumentation is gated by value of information — we go deeper only where knowing more would actually change what you'd do.⁷ Measure the constraint, not the catalog.

Principle 2 — Measure the latent thing, with stated uncertainty

Meyer names the deepest problem precisely: "the performance we want to measure and the performance we can measure are not the same."¹ What you care about (can this team do the work; is it aligned; is it supported) is latent — you can't read it off directly. What you have are observable indicators. The mistake is treating an indicator as if it were the thing.

The mature response is psychometric. You estimate the latent construct from multiple indicators using item response theory, and — this is the part conventional appraisal omits entirely — you report it with its uncertainty, not as a false-precision single number.⁸ Aguinis's textbook defines measurement as the "systematic description of an employee's strengths and weaknesses";⁹ Performix's definition is estimation of a latent construct with a stated confidence interval. The difference between description and estimation-with-uncertainty is the difference between an opinion and a measurement.

Two design laws follow, both load-bearing:

• Every score is a 0–100 index that decomposes. A headline number you can't drill into is an assertion; a headline number that opens to its sub-measures and the data underneath is a measurement. CAMS scores are always decomposable to their sub-constructs and evidence.¹⁰
• Always carry uncertainty and a reference. A number with no interval and no benchmark can't be acted on responsibly — you don't know if it's signal or noise, good or bad. Intervals (Wilson/credible) and a reference point travel with every score.¹⁰

Principle 3 — Watch for measures going stale

Meyer's most under-appreciated finding is that measures run down: almost all of them tend, over time, to "lose variance and hence the capacity to discriminate between good and bad performance."¹ A metric that once separated strong teams from weak ones quietly stops doing so — everyone optimizes to it, the spread collapses, and you're now tracking a number that no longer tells you anything while believing you're well-instrumented.

This is a psychometric property — loss of item information — stated in business language, and it has a direct product consequence: a measure's discriminating power must itself be monitored, and items that have run down get retired or refreshed. (Upstream, this is exactly what an adaptive item lifecycle is for.) A measurement system that never audits its own discrimination is slowly going blind.

Principle 4 — Respect the pathologies

Measurement that drives consequences induces predictable dysfunctions, and the management-control literature has named them — use the names, and design around them:

• Myopia — financial and short-cycle measures push short-term behavior at the expense of the long-term thing you actually wanted.¹¹
• Gaming / surrogation — people optimize the measure instead of the goal. Harris and Tayler call the deeper version surrogation: "the tendency to mentally replace strategy with metrics" — and it's "a common subconscious bias: whenever metrics are present, people tend to surrogate." It's what hijacked Wells Fargo. Their remedies map onto our design: involve the people executing a strategy in forming it, loosen the link between metrics and incentives (pay-for-performance makes surrogation worse), and use multiple metrics so no single one stands in for the whole.¹¹¹²
• The controllability principle — don't hold people to outcomes they don't control; scoring teams on uncontrollable factors is both unfair and uninformative.¹¹

Performix's posture against these isn't exhortation, it's structure: measure the binding constraint (not a wall of proxies to game), carry uncertainty and a reference (so noise can't masquerade as a verdict), and surface measurement as protected feedback rather than a managerial scorecard — which removes most of the incentive to game in the first place.

Principle 5 — Measure for self-control, not control

The oldest and best statement of the whole stance is Drucker's. Measurement, he wrote, exists "to make self-control possible"; to use it "to control people from outside and above, that is, to dominate them, is to abuse measurements — and it is the common violation of this principle that largely explains why measurement is the weakest area in management."¹³

This is the line Performix is built on. The same measurement can be an instrument of domination (a rating that goes in your file, a number that decides your raise) or an instrument of self-control (a team seeing its own binding constraint and what would move it). Protected feedback, decomposable indices, honest uncertainty, and a diagnostic rather than evaluative frame are all in service of the second. Measurement aimed at the performer's own action is the only kind that reliably improves performance — which is, after all, the point.

How we measure our own success: variance explained

There's a measurement we hold ourselves to, and it's the honest one: what percent of the variance in a team's performance can be explained — at the start of an engagement versus at the end? That delta is the value we deliver, and raising it is our constant focus.

This is the mission made quantitative. "Rule out chance" is increasing explained variance — the explained part is, by definition, the part that isn't luck. "What stands up to scrutiny" is whatever actually moves that number out of sample; story doesn't, structure does. And unlike a satisfaction score or an activity count, variance-explained can go down — so it can be wrong, which is exactly why it means something when it goes up. The binding constraint, in this language, is simply the largest tractable chunk of currently-unexplained variance — which is why the diagnosis and the metric are the same discipline.

It also answers the question every buyer eventually asks — don't you fix it and then we don't need you? No, and not as a sales dodge: performance is non-stationary. Fix one constraint and the next becomes binding; people turn over; competitors reset the bar; strategy changes the definition of "good." And explained variance decays on its own — Meyer's "running down" again (Principle 3): a model you stop re-measuring goes stale. So there is always a new constraint, a new wrinkle, a new problem, and a standing need to keep ruling out chance as the target moves.

Where the engine lives

One boundary, stated plainly so the claim is honest: Performix is psychometric-first, but it does not reimplement psychometrics. The adaptive measurement engine — IRT scoring, the item lifecycle, computerized adaptive testing, the value-of-information machinery — lives in the research substrate upstream, and AI is a consumer of that substrate, not the engine of it.¹⁴ Performix's job is to ask the right next question, surface the estimate with its uncertainty and provenance, and make it decomposable and actionable. The measurement is real, and rigorous, precisely because it isn't generated on demand by a language model — it's read off a substrate built from real psychometrics. That is the whole differentiator, and it is why the measurement chapter is the one the rest of the field can't simply copy.

---

Footnotes

1. Meyer, Rethinking Performance Measurement: Beyond the Balanced Scorecard — "swamped with measures"; the want-vs-can gap; the "running down" of measures (loss of variance/discrimination). ↩ ↩² ↩³

2. Cappelli & Tavis, HBR's 10 Must Reads on Performance Management — the retreat from annual ratings toward manager judgment. ↩

3. Buckingham & Goodall, Nine Lies About Work, Lie #6 — ~60% of rating variance is the rater; complex scales worsen it. ↩

4. Kahneman, Sibony & Sunstein, Noise: A Flaw in Human Judgment (2021) — "wherever there is judgment, there is noise"; system noise and occasion noise. ↩

5. Hubbard, How to Measure Anything — Ch.7 "Quantifying the Value of Information"; measurement's purpose is "measurements that support decisions"; measurement as significant uncertainty-reduction "with little effort." ↩

6. Marr, Strategic Performance Management — "management by insights, not numbers"; Key Performance Questions. ↩

7. The Performix measurement canon — start at the CAMS minimum, adaptively narrow, gate depth by value of information (EVPI/EVSI). ↩

8. van der Linden, Handbook of IRT; Embretson & Reise, IRT for Psychologists; the computerized-adaptive-testing canon. The engine lives upstream of Performix. ↩

9. Aguinis, Performance Management (3e) — measurement as "systematic description of … strengths and weaknesses." ↩

10. The index doctrine: every composite ships as a 0–100 index, always decomposable to sub-measures and data, and always carrying an uncertainty interval and a reference point. ↩ ↩²

11. Merchant & Van der Stede, Management Control Systems — myopia; the controllability principle; combinations-of-measures remedies. ↩ ↩² ↩³

12. Harris & Tayler, "Don't Let Metrics Undermine Your Business" (HBR Sep–Oct 2019, in HBR's 10 Must Reads on Performance Management) — surrogation = "confuse what's being measured with the metric being used"; "whenever metrics are present, people tend to surrogate" (Kahneman & Frederick's three conditions); the Wells Fargo case; remedies = co-formulate strategy, loosen the metric↔incentive link, use multiple metrics. ↩

13. Drucker, People and Performance — "to make self-control possible … to use them to control people from outside and above … is to abuse measurements … measurement is the weakest area." ↩

14. Psychometric-first, AI-as-consumer: the adaptive engine and substrate live upstream; Performix surfaces them. ↩