The Thermostat Problem

There's a well-known problem in economics called Goodhart's Law: when a measure becomes a target, it stops being a good measure. The moment you start managing to the number instead of to the thing the number was supposed to describe, the number stops telling you the truth. Most organizations know this in the abstract and violate it constantly in practice, nowhere more visibly than in corporate diversity programs.

Here's the mechanism. If your house is cold, you can walk over to the thermostat and warm the reading with a lighter. The dial will say seventy degrees. The house will still be freezing, because the problem was never the number on the wall, it was the drafty windows and the door that doesn't seal. Companies do the equivalent version of this constantly: they see a diversity number they don't like, so they hire until the number looks better, without touching the conditions that produced the number in the first place. The dial moves. The house stays cold. Often the people hired into an environment that wasn't built for them leave within a year, and the number drifts right back to where it started, at real cost, and with a round of backlash attached to it that makes the next attempt harder.

Paolo Gaudiano, a complexity scientist who spent over two decades in applied agent-based modeling before turning it toward workplace diversity, draws that distinction more sharply than most. Diversity is a company's balance sheet, a snapshot that tells you where things stand but nothing about how you got there. Inclusion is the cash flow: the daily experiences that quietly push people toward the door or keep them in the room. Merit, whether pay and promotion are actually fair, is the profit and loss statement, the record of what happened over time and why. You cannot fix a balance sheet problem by editing the balance sheet. You have to fix the cash flow and the P&L, and let the balance sheet catch up on its own.

The habit of mind underneath that distinction matters more than the DEI application of it. Paolo has a phrase for the failure mode he's spent his career working against: "big data, little brain." Most data science, as he put it, is pattern-matching at scale with no domain understanding behind it. A model surfaces a correlation, hands it to an expert, and the expert is left guessing at what it means and what to do next. Agent-based modeling asks something different of you before it gives you anything back: you have to actually understand who the people in the system are and what they're responding to. The model doesn't replace that understanding. It's that understanding, made explicit enough to test before you bet a real decision on it.

The same argument underlies every World Model we build, and it generalizes well past HR. Whether the emergent outcome in question is a diversity number, a disease curve, or a product adoption rate, the failure mode is identical: treating the outcome as the thing to manage instead of treating it as evidence of a structure you haven't examined closely enough. A World Model doesn't just give you a better number to stare at. It gives you the structure underneath the number: the actual households, workplaces, and conditions people are responding to, so you can find the lever that genuinely moves the outcome instead of the one that just makes the dial read warmer for a quarter. The same logic applies just as directly to growth and adoption questions: why a product plateaus in one market and not another, why usage climbs and then stalls, which lever actually moves people versus which one just moves the dashboard.

Paolo goes further into this on The Flux, our podcast, including how he measures inclusion directly rather than waiting for it to show up as a diversity statistic. Worth a listen if the distinction is useful to you. The question underneath all of it: are you managing what the thermostat reads, or do you understand the house well enough to actually fix it?