Haute Lumière
Commerce · II.09 · MMXXVI · daylight
TEN CONCEPT BRIEFS · Chapter II.09 — Evolution and Economic Selection
One page each. A reader who reads only these ten pages has the chapter.
BRIEF 1 — Variation, Selection, Inheritance
The idea. Anything that has all three of these will evolve, whether or not anybody intends it to. Remove any one and it will not.
- Variation — the members of the population differ.
- Selection — those differences affect who persists.
- Inheritance — what persists is passed on with reasonable fidelity.
This is substrate-neutral. It is as true of cost codes and shift patterns as of finches. It is also a diagnosis: when a sector is not improving, one of the three is missing, and it is worth knowing which.
Worked example. A regulated utility with no new entrants, a single approved operating manual and a workforce that never moves between operators has no variation and no route for inheritance. Selection pressure can be enormous and nothing will happen, because there is nothing for it to act on.
Why it matters. It converts why is nothing getting better here from a complaint into a three-way test with a different fix at each branch: widen the variation, sharpen the criterion, or build the transmission.
You already know this because you have watched an idea die when the one person who held it left, and you did not need a theory to know that the idea was never really in the organisation — it was in the person.
BRIEF 2 — Routines Are the Unit of Inheritance
The idea. Nelson and Winter's move, in An Evolutionary Theory of Economic Change (1982), was to identify what actually replicates in an economy. Not the firm — firms die. Not the plan — plans are rewritten. The routine: the repeatable, largely tacit pattern by which a thing is done.
Routines replicate imperfectly and in several directions at once. People carry them when they move. Suppliers copy them. Spin-outs inherit them wholesale. Consultants transmit them for a fee. A routine can outlive every person and every legal entity that ever held it.
Worked example. A scheduling practice invented in one plant appears six years later in four unrelated firms, three of which employ someone who once worked at that plant. The firm that invented it has been sold twice and no longer uses it. The routine is doing better than its parent.
Why it matters. If routines are what inherits, then the things that raise inheritance fidelity — documentation, secondment, apprenticeship, open method — are not soft investments. They set h² in the breeder's equation, and the response to selection is directly proportional to it.
You already know this because you can name a habit your organisation has that nobody chose, nobody wrote down, and nobody can quite get rid of.
BRIEF 3 — The Price Equation
The idea. George Price, in 1970, wrote down the exact accounting of how the mean of any trait changes in any population under any selection whatsoever.
Cov(w , z) E(w · Δz)
Δz̄ = ------------ + ------------
w̄ w̄z is the trait. w is offspring per parent. The first term is selection — the covariance between the trait and who persists. The second is transmission — how faithfully the persisters pass the trait on.
Worked example. Five firms, z = capability spend as a share of revenue. z̄ = 6.000 percent, w̄ = 1.100, E(w·z) = 7.320, so Cov(w,z) = 7.320 − 6.600 = 0.720. The selection term is 0.720 / 1.100 = 0.6545 points a generation; the transmission term is 0.094 / 1.100 = 0.0855. The mean moves 0.7400 points, of which selection is 88.5 percent.
Why it matters. It tells you where to intervene. A population whose change is 88 percent selection needs a different criterion; one whose change is mostly transmission needs better copying.
You already know this because you have seen a company's average change for two quite different reasons — because the laggards were acquired, and because everyone got slightly better — and you could tell the difference.
BRIEF 4 — An Identity Is Not a Mechanism
The idea. The Price equation is true by construction. It fits every population, in every period, including populations nothing is selecting at all. So the Price equation fits is never evidence of anything.
This is the entry point for every just-so story in evolutionary economics. If fitness is inferred from survival, then "the fit survived" and "the survivors survived" are the same sentence, and no observation can contradict it. Edith Penrose made exactly this objection in 1952 and was right about the version she was attacking.
Worked example. Take any industry, look at who grew, name a trait the growers shared, and call it selection. The account will fit perfectly. It will also fit if you run the same procedure on a population whose growth rates were drawn at random.
Why it matters. Because there is a real version and it is worth having, and the only way to keep it is to be strict about what separates it from the decorative one. That separation is Brief 10.
You already know this because you have heard a post-hoc explanation of a market outcome that would have sounded equally convincing if the outcome had gone the other way.
BRIEF 5 — The Fitness Function Is the Accounting
The idea. A population moves toward whatever covaries with persistence — and in an economy, what covaries with persistence is whatever the measurement system counts. The measurement system is the selection environment.
Worked example. The same five firms as Brief 3. Measure w as next period's revenue and Cov(w,z) = +0.720, giving +0.6545 points a generation. Measure w as this period's operating margin instead, and the weights invert — capability spend costs margin now — so Cov(w',z) = −0.620 and the term is −0.5688. Nothing about the firms changed. The swing is 1.2234 points a generation, and over ten three-year generations the two economies stand 12.23 percentage points apart.
Why it matters. It relocates the lever. The most powerful evolutionary intervention available to an organisation is not a subsidy, a mandate or a culture programme. It is a definition — what is counted, at what boundary, over what period. That is a document, and documents can be edited.
You already know this because you have watched a team's behaviour change the week a new scorecard arrived, before anyone had said a word about values.
BRIEF 6 — Multi-Level Selection
The idea. Selection acts at every level where there is variation, differential persistence and inheritance — on individuals within groups, and on groups within populations, at the same time and often in opposite directions. The Price equation partitions exactly into the two.
Worked example. Six firms, two clusters of three. z is spare capacity lent to a neighbour; lending costs the lender c = 0.030 per point and benefits every firm in the cluster by b = 0.060 per point of the cluster mean.
BETWEEN clusters +0.2348 pp per generation
WITHIN clusters -0.0696 pp per generation
total +0.1652 pp per generationInside both clusters the lending firms lose — the within-cluster covariance is −0.0800 in each. Between clusters the lending clusters win, 3.37 times harder. The trait rises while losing everywhere it is measured locally.
Why it matters. It explains the behaviour every executive has seen and no single-level model handles: practices that are individually costly, collectively valuable, and therefore permanently vulnerable to a scorecard drawn at the wrong boundary.
You already know this because you have seen a team win its own targets by doing something that cost the company more than it gained.
BRIEF 7 — Wilson and Sober's Five Criteria
The idea. Group selection is not a mood. It operates when, and only when, five conditions hold, as set out in Unto Others (1998) and defended in Wilson and Wilson (2007):
- The population is structured into groups that persist and that form and disperse in repeated cycles.
- Groups differ in the trait — between-group variance is strictly positive.
- The trait affects group fitness.
- Group composition is heritable: offspring groups resemble parent groups.
- Between-group selection exceeds within-group selection.
"Selfishness beats altruism within groups. Altruistic groups beat selfish groups. Everything else is commentary."
Worked example. Criterion five, made exact for the model in Brief 6. The trait rises when (b−c)·V_between > c·V_within, that is when
b / c > 1 + V_within / V_between = 1 + 2.6667 / 9.0000 = 1.2963Here b/c = 2.0000, so the critical b* is 0.03889 and the headroom is 1.54 times. Set b = 0.020 and b/c falls to 0.6667: the same six firms now yield −0.1789 points a generation and the trait is selected out.
Why it matters. Criterion two has a commercial consequence nobody expects: as groups are made alike, V_between falls, the required b/c goes to infinity, and group selection becomes impossible. Homogenisation is not neutral.
You already know this because you have seen a merger remove the last organisation that was doing it differently, and with it the only evidence that another way worked.
BRIEF 8 — Replicator Dynamics and the Time Constant
The idea. When two variants compete and the better one holds a selection advantage s, its share follows ẋ = x(1−x)s — a logistic curve. The closed form gives the time to get from one share to another:
t = ln[ (x₁/(1−x₁)) · ((1−x₀)/x₀) ] / sWorked example. From 2 percent of a market to 50 percent, the odds ratio is 49.00 and its logarithm 3.8918.
s = 0.03 / yr → 129.7 years s = 0.15 / yr → 25.9 years
s = 0.08 / yr → 48.6 years s = 0.30 / yr → 13.0 yearsTo 90 percent the odds ratio is 441.0, and even a 15 percent advantage needs 40.6 years.
Why it matters. It puts a clock on "the market will sort it out." A better routine with a modest advantage, left to selection alone, takes longer than a career. If something spread faster than this, it was not selection — somebody copied it deliberately, and deliberate copying is the fast path.
You already know this because you can name a plainly better practice in your own industry that has been plainly better for twenty years and is still not standard.
BRIEF 9 — The Breeder's Equation, Applied to Routines
The idea. R = h² · S. The response to selection per generation equals the transmission fidelity times the selection differential — how far above the population mean the selected members sit.
Worked example. S = 4.00 points; h² = 0.35, assumed, not measured. Then R = 1.40 points a generation. With a three-year generation and a ten-year horizon there are 3.33 generations, delivering 4.67 points against a 10.00-point requirement: a shortfall of 5.33 points, and 46.7 percent of the mandate.
To close it, R must reach 3.00 a generation, which means S = 8.57 points (select harder), or h² = 0.750 (copy better), or a generation of 1.40 years (cycle faster). Halve h² to 0.175 and everything downstream halves with it.
Why it matters. It makes the promise falsifiable before the money is spent. Most transformation programmes fail this arithmetic and none of them run it.
You already know this because you have sat in a room where a five-year target was set by working backwards from the desired end point rather than forwards from the achievable rate.
BRIEF 10 — The Test That Separates the Two Versions
The idea. A selection claim earns the word science when it can be wrong about a number, in advance. Three conditions:
- The trait is measured before the period, not reconstructed from the winners.
- Fitness is operationalised independently of the outcome it explains — a defect rate, a cost per unit, a repeat-purchase rate — so fit and survived are two measurements, not one.
- The model states a number and an interval before the period runs.
Worked example. Predicted Δz̄ = 0.6545 points, ±0.25, so the claim dies unless the realised change falls in [0.4045, 0.9045]. Realised 0.71 survives. Realised 0.40 falsifies. Realised 1.10 falsifies.
And the null that must be beaten: ten generations of 1.40-point steps with a reversing sign have an expected displacement of zero and a standard deviation of 1.40 × √10 = 4.43 points. An industry mean that moved four points in thirty years is inside one sigma of no selection at all.
Why it matters. It is the difference between an instrument and a vocabulary.
You already know this because you already apply this standard to a forecast — you ask what number, by when, and how wrong it can be before you stop believing it.