The laws that govern trends, wealth, and systems — where the math itself, not psychology, produces the counterintuitive result.
Compounding robust
Growth on growth curves flat then vertical — money, skill, reputation alike.
Mechanism When each period's gain is added to the base that generates the next gain, growth is exponential, not additive. The curve looks nearly flat for a long time because early absolute gains are tiny, then bends sharply upward as the base becomes large — most of the total arrives late. The same structure applies to any process where output feeds back into the productive base: invested capital, compounding skill, accumulating reputation and relationships.
Implication Start early and never interrupt the compounding; the long flat stretch is the price of the vertical part.
Trap Quitting during the flat years because progress feels absent — right before the curve turns up.
The Rule of 72 robust
Divide 72 by the growth rate to get the doubling time.
Mechanism For a quantity compounding at r% per period, doubling time ≈ 72 / r (it's an approximation of the exact ln(2)/ln(1+r), accurate for typical rates). It converts a percentage into an intuitive timescale: 8% doubles in ~9 years, 10% in ~7.2, 2% inflation halves purchasing power in ~36. The same tool works in reverse for decay (halving under inflation or decline).
Implication Use it to sanity-check any growth or inflation claim on the spot — no calculator needed.
Trap Underrating a "small" rate — a 6% fee or 3% inflation doubles/halves faster than intuition expects.
Exponential intuition failure robust
Humans model exponentials as linear and are always wrong on timing.
Mechanism The brain extrapolates trends linearly (add a constant each step), but exponential processes multiply. So we consistently underestimate future values and, crucially, mis-time inflection points: an exponential looks negligible until it's suddenly overwhelming (the lily pad covering half the pond on the last day before it's full). Compounding growth, pandemics, and technology adoption all blindside linear intuition.
Implication When a process is multiplicative, distrust your gut on both magnitude and timing; compute it out.
Trap "It's still tiny, we have time" — said just before an exponential curve goes vertical.
Pareto principle (80-20) robust
A small share of causes produces most of the effects — and it's fractal.
Mechanism In many systems, outputs are distributed unevenly so that roughly 20% of inputs generate ~80% of results (customers, revenue, bugs, effort). It arises from underlying power-law/multiplicative processes, not a magic ratio. It's also fractal: within the top 20%, the same skew repeats (the top 4% often drives ~64%), so concentration compounds as you zoom in.
Implication Find and protect the vital few; ruthlessly deprioritize the trivial many. Then reapply within the vital few.
Trap Spreading effort evenly across all inputs as if each contributed equally — most don't.
Power laws robust
Much of the world isn't bell-curved; it's dominated by rare extremes.
Mechanism In a normal distribution, extremes are vanishingly rare and the mean is representative. Many real quantities — wealth, city sizes, book sales, network connections — follow power laws instead, where a few instances are orders of magnitude larger than the median and the "average" is meaningless. These arise from mechanisms like preferential attachment (rich-get-richer) and multiplicative growth, not additive noise.
Implication Before using an average or assuming a bell curve, ask if the domain is actually power-law; if so, the tail is the story.
Trap Planning around the "average" outcome in a domain where a handful of extremes dominate everything.
Fat tails & Black Swans robust
Rare, extreme events dominate outcomes in many domains.
Mechanism In fat-tailed (Extremistan) domains, a single observation can dwarf the entire rest of the sample — one market crash, one pandemic, one bestseller. Taleb's Black Swans are high-impact events that are rare, unpredictable in advance, and rationalized as predictable afterward. Because standard (thin-tailed) statistics vastly underweight tail probability, models built on them systematically underprice catastrophic and jackpot events alike.
Implication In fat-tailed domains, protect against ruin and expose yourself to positive tails; don't trust variance-based risk models.
Trap "Six-sigma" risk models that price a once-a-millennium event that then happens twice a decade.
Network effects & Metcalfe's law contested
A network's value grows faster than its user count.
Mechanism Each new user can connect with all existing users, so the number of possible pairwise connections grows as n(n−1)/2 — roughly with the square of users (Metcalfe's law). Value therefore rises super-linearly with size, creating a self-reinforcing advantage: bigger networks are more valuable, attract more users, and get bigger. The n² form overstates real value (not all connections are equal), so it's a directional truth, not a precise one.
Implication In networked markets, size begets value begets size — getting to critical mass first can be decisive.
Trap Competing on features against an entrenched network whose real moat is its user base, not its product.
Zero marginal cost of digital goods robust
Software and media cost ~nothing to reproduce — which rewrites their economics.
Mechanism A physical good costs materials and labor per unit; a digital good has high fixed cost to make the first copy and near-zero marginal cost for every copy after. So profit is dominated by scale: once built, each additional user is almost pure margin, and price can fall toward zero while remaining viable. This is why one program can serve billions and why "free + upsell" business models work for digital and not for widgets.
Implication Build or own things that copy for free and distribute globally; the economics dwarf per-unit businesses.
Trap Applying scarcity-based pricing intuition to digital goods, or competing on price against something that costs nothing to copy.
Winner-take-most dynamics robust
Scalable markets concentrate rewards on a tiny few.
Mechanism When a product scales without per-unit constraints (digital distribution, network effects, low switching to the leader), small quality or timing advantages compound into dominant market share — the best option captures most of the demand because there's little reason to choose the second-best. Combined with power-law payoffs, this yields superstar economics: #1 earns multiples of #2, who earns multiples of #3.
Implication In scalable markets, aim to be clearly first/best in a narrow category rather than a solid also-ran in a broad one.
Trap Being a close second in a winner-take-most market and capturing a fraction of the leader's return.
The Lindy effect contested
For non-perishables, life expectancy grows with current age.
Mechanism For things that don't age biologically — ideas, technologies, books, institutions — every additional year of survival is evidence of robustness, so expected remaining life increases with age. A book in print for 50 years is likely to last another 50; a fad that appeared last year probably vanishes. It's a survival-statistics heuristic (a proxy for unobserved fragility), not a physical law, and fails when conditions change abruptly.
Implication Bias toward time-tested tools, ideas, and practices; treat the brand-new as far more likely to disappear.
Trap Chasing the newest framework/diet/trend over the boring thing that has already survived decades.
Regression to the mean robust
Extreme results tend to be followed by more average ones.
Mechanism When an outcome combines skill and luck, an extreme value usually rode extreme luck, which doesn't repeat — so the next measurement drifts back toward the average. This is pure statistics, but it masquerades as causation: praise a great performance and it "worsens," punish a terrible one and it "improves," fooling you into crediting the intervention when only the luck component regressed.
Implication Expect extremes to normalize on their own; don't credit (or blame) an intervention for what regression would have produced anyway.
Trap Concluding punishment works and praise backfires — the classic misread of pilots, athletes, and hot funds.
Survivorship-adjusted base rates robust
The true odds must include the failures you never see.
Mechanism Base rates computed from visible cases are inflated because failures are removed from the sample before you observe it — the surviving restaurants, funds, and creators are counted; the vanished majority aren't. To get the real probability of success, you must reconstruct the full denominator (attempts, not just survivors), which is usually far larger and makes the true odds far worse than the visible field suggests.
Implication When estimating odds, ask "out of everyone who tried, how many are still here to be counted?" and use that denominator.
Trap Estimating startup or creator success from the ones you've heard of — a sample defined by having succeeded.
Marginal gains (1% compounding) contested
Small consistent percentage improvements compound to large totals.
Mechanism Improving by 1% daily yields ~37× over a year (1.01³⁶⁵); declining 1% daily leaves you near zero — the arithmetic of repeated multiplication. Applied to skills, habits, and systems, tiny gains that individually feel trivial accumulate into a wide gap over time. The catch: real-world gains aren't smoothly multiplicative or unbounded (diminishing returns, plateaus), so treat the 37× as motivational math, not a guarantee.
Implication Prioritize a sustainable stream of small improvements over rare heroic pushes; consistency is the multiplier.
Trap Dismissing small daily gains (or small daily erosions) as too minor to matter over a year.
Diminishing returns & S-curves robust
Most real growth is S-shaped: slow, fast, then flattening.
Mechanism Few things grow exponentially forever; resource limits, saturation, and competition bend growth into a sigmoid — slow start, rapid middle, plateau. Within that, each additional unit of input yields less output as you approach the ceiling (diminishing marginal returns). Mistaking the steep middle for permanent exponential growth, or the early flat part for failure, are both common misreads of the same S-curve.
Implication Identify where you are on the S-curve; near the top, added effort barely moves output — better to jump to a new curve.
Trap Pouring resources into optimizing a maxed-out curve instead of finding the next one.
Diminishing marginal utility robust
Each additional unit of something is worth less than the last.
Mechanism The satisfaction from consuming more of a good declines with quantity — the first slice of pizza, the first $10k of income, delivers far more utility than the tenth. This concavity explains risk aversion (a sure amount beats a gamble with the same expected value because the upside adds less utility than the downside subtracts) and why wealth's happiness effect flattens.
Implication Diversify sources of satisfaction rather than maximizing one; past a point, more of the same buys little.
Trap Chasing more of a good already deep in diminishing returns (income, features, stuff) at rising cost for shrinking gain.
Metcalfe / Reed / Sarnoff contested
Three ways network value can scale — broadcast, pairwise, and group-forming.
Mechanism Sarnoff's law: a broadcast network's value scales linearly with audience (n) — one-to-many. Metcalfe's law: a network enabling pairwise connections scales as ~n² — many-to-many. Reed's law: a network that lets subgroups form scales as ~2ⁿ (the number of possible subsets), the steepest — the value of communities and group-forming platforms. Each describes a different connection topology; real networks sit between these bounds.
Implication Match the model to the network type: broadcast, connection, or community — the last two scale far more steeply.
Trap Valuing a group-forming platform with broadcast (linear) intuition and badly underestimating it — or vice versa.
Moore's law & cost-curve deflation robust
Some technologies get predictably cheaper/more powerful over time.
Mechanism Moore's law observed transistor density roughly doubling every ~2 years, driving exponential compute-cost deflation for decades. It's an empirical trend (and an economic self-fulfilling roadmap), not a physical law, and its pure form is slowing — but the broader pattern of exponential cost decline in maturing technologies is a usable planning input: things that are expensive now may be trivially cheap on your project's horizon.
Implication Plan for future cost curves: build for what a capability will cost in a few years, not what it costs today.
Trap Assuming today's cost/performance is fixed — or, oppositely, assuming a specific curve continues forever.
Wright's law (experience curves) robust
Unit cost falls a fixed percentage with every doubling of cumulative production.
Mechanism Each doubling of total units ever produced reduces cost per unit by a roughly constant fraction (e.g. ~20%), because cumulative experience drives learning, process improvement, and scale. Unlike Moore's law (tied to time), Wright's law ties cost to volume made — which is why cost falls faster when adoption is faster. It has held across aircraft, solar panels, batteries, and semiconductors.
Implication Forecast cost declines from cumulative volume, and note that accelerating adoption accelerates the price drop (a virtuous loop).
Trap Dismissing an expensive emerging technology by static cost, ignoring the experience-curve decline volume will bring.
Little's law robust
Items in a system = arrival rate × time each spends in it.
Mechanism For any stable queue or pipeline, average work-in-progress L = throughput λ × average lead time W (L = λW), regardless of the arrival pattern or service details. Rearranged, lead time = WIP / throughput — so if you want faster turnaround at fixed throughput, the only lever is reducing how much is in progress at once. It's an exact, assumption-light relationship for stable systems.
Implication To shorten cycle time, cut WIP (fewer simultaneous tasks), not just "work faster" — limit the number in flight.
Trap Starting more and more work to "get more done," which inflates WIP and lengthens everything's completion time.
Queueing theory: the 80% wall robust
Push utilization past ~80% and wait times explode nonlinearly.
Mechanism In a system with variability, average wait time scales with utilization ρ as roughly 1/(1−ρ). As you approach 100% utilization, that term blows up: going from 80% to 90% busy roughly doubles the wait, 90% to 95% doubles it again. There's no slack to absorb the natural bursts, so queues cascade. High utilization looks efficient but is why "fully loaded" systems (roads, teams, servers) grind to a halt.
Implication Deliberately leave slack; targeting 100% utilization guarantees long, unstable waits and no capacity for surprises.
Trap Scheduling people or servers at ~100% "for efficiency" and getting gridlock the moment anything varies.
Amdahl's law robust
The un-optimized part caps your total possible speedup.
Mechanism If a fraction of a process can't be improved (or parallelized), that fraction sets a hard ceiling on overall gains no matter how much you optimize the rest. If 10% of a task is serial, even infinite speedup of the other 90% caps total improvement at 10× — the serial 10% now dominates. Optimizing the already-fast part yields nearly nothing once the bottleneck is elsewhere.
Implication Find and attack the true bottleneck; effort on non-bottleneck parts barely moves the total.
Trap Heavily optimizing the glamorous 90% while the ignored 10% quietly limits everything.
Goodhart's law robust
When a measure becomes a target, it stops being a good measure.
Mechanism A metric is a proxy for something you actually care about. Once people are rewarded on the proxy, they optimize the proxy directly — including by gaming it in ways that break its link to the real goal. The correlation that made it a useful signal collapses under the pressure to hit it, because there are always cheaper ways to move the number than to improve the underlying thing.
Implication Don't over-incentivize single metrics; use multiple, watch for gaming, and keep sight of the real goal behind the proxy.
Trap Paying for lines of code, calls made, or test scores and getting exactly those numbers with none of the intended value.
Campbell's law robust
The more a social metric drives decisions, the more it gets corrupted.
Mechanism A cousin of Goodhart's, specific to social indicators: the higher the stakes attached to a quantitative measure used for social decision-making, the more subject it is to corruption pressure and the more it distorts the process it's meant to monitor. High-stakes testing invites teaching-to-the-test and cheating; crime stats get reclassified; the measure both corrupts behavior and stops reflecting reality.
Implication Be wary of any single high-stakes social metric; expect distortion to scale with the stakes attached.
Trap High-stakes school testing that produces score inflation and narrowed teaching instead of real learning.
Parkinson's law robust
Work expands to fill the time available for it.
Mechanism Absent a binding constraint, a task's effort inflates to consume whatever time is allotted — via added perfectionism, procrastination, and scope creep — because there's no forcing function to stop. A report that could take an hour takes the week you gave it. A tight, credible deadline removes the slack and compresses the work to what's essential.
Implication Set deliberately tight deadlines and time-boxes; constrain the time to constrain the sprawl.
Trap Generous deadlines and open-ended timelines that guarantee tasks swell to fill them.
Hofstadter's law robust
It always takes longer than you expect — even when you account for this law.
Mechanism A self-referential statement of the planning fallacy: estimates are optimistic because we can't foresee the specific unknowns that cause delay, and padding the estimate doesn't help because the same optimism reapplies to the padded number. The unforeseen is, by definition, not in the estimate — and there's always some.
Implication Pad estimates using outside-view historicals (how long similar things really took), then expect to still run over.
Trap Committing to deadlines built on your own best-case estimate plus a token, still-optimistic buffer.
Gall's law robust
Complex systems that work evolved from simple systems that worked.
Mechanism A working complex system is almost always the descendant of a working simple one, grown and adapted incrementally. A complex system designed from scratch rarely works and can't be patched into working — because you can't anticipate all the interactions and failure modes up front; they're discovered only by running a simpler version. Emergence and edge cases are found empirically, not designed.
Implication Start with the simplest thing that works and evolve it; don't attempt a grand complex design in one leap.
Trap Big-bang rewrites and from-scratch mega-systems that collapse under untested complexity.
Conway's law robust
Organizations ship systems that mirror their communication structure.
Mechanism A system's architecture tends to copy the org chart that built it, because the interfaces between components must be negotiated across the interfaces between teams — where teams don't talk, modules don't integrate cleanly. Four teams building a compiler produce a four-pass compiler. The communication topology of the makers becomes the structural topology of the made.
Implication To get the architecture you want, structure the teams to match it ("inverse Conway maneuver").
Trap Fighting a technical architecture problem that's really an org-structure problem — the system keeps re-mirroring the org.
The Peter principle contested
People rise to their level of incompetence.
Mechanism If promotion is based on performance in the current role, employees keep getting promoted until they reach a role they're not good at — where, being no longer promotable, they stall. Over time positions fill with people at their incompetence ceiling. The core assumption (skill in one role predicts skill in the next) is often false, and empirical support is mixed, but the failure of promoting your best salesperson into management is real and common.
Implication Promote on evidence of ability for the new role, not just excellence at the current one; offer non-management advancement.
Trap Promoting your best individual contributor into a manager who's now bad at management and lost as a contributor.
Brooks's law robust
Adding people to a late project makes it later.
Mechanism New people must be brought up to speed (pulling productive people off work to train them), and adding staff multiplies communication paths (n(n−1)/2), raising coordination overhead. Work that can't be cleanly partitioned can't be sped up by more hands. So in the short term, adding people to a delayed, complex project increases ramp-up and coordination cost more than it adds output.
Implication Rescue a late project by cutting scope or improving the plan, not by throwing bodies at it late.
Trap Panic-hiring onto a slipping deadline and pushing it further out.
Dunbar's number contested
There's a cognitive ceiling (~150) on stable relationships.
Mechanism Dunbar linked neocortex size to group size across primates and extrapolated a human limit of ~150 stable relationships you can maintain — with tighter inner layers (~5 intimates, ~15 close, ~50 friends). Maintaining a relationship costs cognitive and time resources, capping the number. The exact figure and the brain-size basis are debated, but the layered structure and the existence of a practical ceiling are widely observed.
Implication Treat close-relationship capacity as finite; invest deliberately in the inner layers rather than spreading thin.
Trap Confusing thousands of online "connections" with the far smaller number of real, maintainable relationships.
Price's law contested
Half the output comes from the square root of the contributors.
Mechanism In a group of n contributors, roughly √n of them produce about half the total output — so in a team of 100, ~10 people do half the work; in 10,000, just ~100 do. It reflects the extreme skew of productivity (a power-law/creativity distribution), and the gap widens as the group grows. The specific √n form is a rough empirical regularity, not a proven law, but the underlying concentration of output is real.
Implication Identify and retain the small core carrying disproportionate output; understand output won't scale linearly with headcount.
Trap Assuming doubling headcount doubles output, or treating all contributors as interchangeable units.
Benford's law robust
In natural datasets, leading digits aren't uniform — 1 appears ~30% of the time.
Mechanism In many real, multiplicatively-generated datasets spanning orders of magnitude (financials, populations, physical constants), the first digit d occurs with probability log₁₀(1 + 1/d): "1" ~30.1%, "2" ~17.6%, down to "9" ~4.6%. It emerges from scale-invariance and logarithmic spacing. Fabricated numbers usually don't follow it because humans distribute invented digits too evenly — which is why auditors and forensic accountants use it to flag fraud.
Implication Use the leading-digit distribution as a cheap first screen for fabricated or manipulated numeric data.
Trap Assuming a clean-looking set of figures is genuine; fabricated data often fails Benford in ways the eye can't see.
Cunningham's law anecdotal
The fastest way to get the right answer is to post the wrong one.
Mechanism People are far more motivated to correct an error than to answer a question — correcting satisfies status, pedantry, and the drive to demonstrate knowledge, while answering a plain query is effort with no payoff. So a confidently wrong statement summons detailed corrections that a polite question wouldn't. It's an observation about online social dynamics, not a rigorous law, but it reliably works.
Implication To extract expertise from a crowd, state a plausible-but-wrong claim and let the corrections roll in.
Trap The flip side: much online "debate" is just this dynamic — people correcting to score points, not to inform.
The bullwhip effect robust
Small demand changes amplify into wild swings up the supply chain.
Mechanism Each stage of a supply chain orders based on its downstream orders plus a safety buffer, and reacts to delays. A small blip in end-consumer demand gets amplified at each upstream step as everyone over-orders to protect against stockouts and over-cuts when demand dips — so variability grows the further you are from the customer. Information delays and batching turn a gentle ripple into a whipcrack at the factory.
Implication Share real demand data across the chain and reduce order batching/delay to damp the amplification.
Trap Upstream producers whipsawed between shortage and glut by demand signals distorted through the chain.
Jevons paradox robust
Making a resource more efficient can increase its total use.
Mechanism When technology makes using a resource cheaper/more efficient per unit, the effective price of the service falls, demand rises, and new uses open up — so total consumption can grow rather than shrink. Jevons saw more efficient steam engines increase (not decrease) coal use. The efficiency gain is partly or wholly eaten by the rebound in demand it enables.
Implication Don't assume efficiency alone cuts total consumption; if you want less total use, pair efficiency with a cap or price.
Trap Expecting a more fuel-efficient fleet or faster tool to reduce total fuel/time use, when it expands usage instead.
Tragedy of the commons robust
Shared resources get depleted because individual incentives ignore collective cost.
Mechanism When a resource is shared but its benefits are private and its costs are spread across everyone, each actor rationally takes more than a sustainable share — they capture the full gain but bear only a fraction of the depletion. Multiply that across all users and the resource collapses, even though everyone would be better off restraining. It's a structural incentive misalignment, not a moral failing.
Implication Solve commons problems structurally — property rights, quotas, pricing the externality, or governed cooperation — not by appeals to virtue.
Trap Overfishing, aquifer depletion, ad spam, and pollution — each rational individually, ruinous collectively.
Reflexivity contested
Beliefs about a market change the market, which changes the beliefs.
Mechanism Soros's idea: in markets, participants' perceptions and the underlying reality influence each other in a two-way feedback loop. Rising prices can improve the fundamentals (easier financing, more confidence), validating the rise and pushing it further — a self-reinforcing boom detached from equilibrium — until the gap between perception and reality snaps back in a bust. Unlike physical systems, the observers' beliefs are part of the system.
Implication In belief-driven markets, watch the feedback between narrative and fundamentals; trends can self-sustain far past "fair value," then reverse violently.
Trap Assuming markets always revert to a stable fundamental value, ignoring self-reinforcing bubbles and crashes.
Antifragility contested
Some systems gain from disorder; fragile ones break under it.
Mechanism Taleb's trichotomy: fragile things are harmed by volatility and shocks; robust things resist them; antifragile things actually improve from a dose of stress, disorder, and error — because they contain redundancy, optionality, and feedback that let them adapt and overcompensate (muscles from load, immune systems from exposure, evolution from mutation). The key is bounded, small stressors with capped downside, not catastrophic ones.
Implication Build in redundancy, small frequent stressors, and optionality so volatility strengthens rather than destroys you.
Trap Optimizing away all slack and redundancy for efficiency, creating a fragile system that shatters at the first shock.
Second-order effects as default robust
In complex systems, unintended consequences are the rule, not the exception.
Mechanism Complex systems are densely interconnected with feedback loops and delays, so any intervention ripples into effects far from the point of action — and because actors adapt to the change, the system pushes back in ways the intervention didn't model. The intended first-order effect is usually swamped or reversed by the second- and third-order responses (the cobra effect: a bounty on cobras breeds more cobras).
Implication Assume every intervention has unintended consequences; look for how actors will adapt, and pilot small before scaling.
Trap Confident large-scale interventions whose adaptive side effects dwarf and sometimes reverse the intended result.
Ergodicity: time average ≠ ensemble average robust
What happens to you over time can differ sharply from the average across everyone.
Mechanism Expected value averages outcomes across many parallel players at one instant (the ensemble average). But an individual lives one sequence through time (the time average). For non-ergodic, multiplicative processes — like wealth, where losses compound off a shrinking base — the two diverge. Classic case: a coin flip paying +50% or −40% has positive ensemble EV (+5%/round), yet almost every individual path trends to zero, because the time-average growth (geometric mean 1.5 × 0.6 = 0.9) is negative. The average gambler wins; the typical gambler is ruined.
Implication Optimize the time-average — what you actually live through — not the ensemble EV; avoiding ruin dominates chasing positive expectation.
Trap Taking repeated +EV multiplicative bets that look attractive individually but drive your specific path toward zero.
Variance drag (the volatility tax) robust
Volatility mechanically lowers what a series compounds to.
Mechanism Compounding is multiplicative, so what matters is the geometric mean, which is always ≤ the arithmetic mean — and the gap grows with volatility (≈ σ²/2). A gain and an equal-percentage loss don't cancel: +50% then −50% leaves 0.75, a 25% loss. Two assets with the same average return therefore compound to different amounts; the more volatile one ends lower. Volatility is a direct, unavoidable drag on long-run growth, separate from any risk preference.
Implication Judge long-run performance by compound (geometric) returns, and value reduced volatility for its own sake — it raises what you actually keep.
Trap Comparing investments by average (arithmetic) return, ignoring that the volatile one quietly compounds to less.
Simpson's paradox robust
A trend in every subgroup can reverse when the groups are combined.
Mechanism When data is aggregated across groups of unequal size that also differ in a lurking variable, the combined trend can point opposite to the within-group trends. The canonical case: 1973 UC Berkeley admissions looked biased against women overall, yet within almost every department women were admitted at equal-or-higher rates — because women applied more to competitive, low-admission departments. The confounder plus unequal weighting flips the aggregate. Which figure is "right" depends on the causal question, not the math alone.
Implication Always check whether a headline aggregate holds within relevant subgroups; disaggregate before trusting or acting on a combined statistic.
Trap Being convinced by an aggregate number that reverses the moment you split it by the group that actually drives it — a favorite of statistical spin.
Comparative advantage robust
Two parties gain from trade even when one is better at everything.
Mechanism Ricardo's insight: gains from trade come from relative, not absolute, advantage. Even if you're better than someone at every task, your time is finite, so you should specialize in what you're relatively best at (highest opportunity cost to give up) and trade for the rest — and so should they. Both end up with more total output than doing everything themselves. It's counterintuitive because it holds even against a strictly more capable partner.
Implication Specialize in your highest relative-value work and delegate or trade for the rest, even tasks you'd do better yourself — your time's opportunity cost decides.
Trap Doing everything yourself because you're the best at it, while your comparative advantage — and highest-value hours — goes unspent.
Prices as information robust
A price compresses dispersed knowledge no central planner could gather.
Mechanism Hayek's argument: the knowledge needed to allocate resources — local conditions, preferences, scarcities — is scattered across millions of minds and can't be centralized. Prices solve this by summarizing all of it into a single number everyone can act on without knowing why it moved; a shortage anywhere raises the price, signaling everyone to economize and supply more. It's a distributed computation, which is why markets often outperform committees at allocation.
Implication Read prices (and their changes) as compressed signals of real conditions; respect that a market price often encodes information you don't have.
Trap Assuming a smart central decision can beat the price signal, ignoring the vast dispersed knowledge the price already integrates.
Moral hazard robust
Insulate someone from the downside and they take more risk.
Mechanism When a party doesn't bear the full consequences of its actions — because someone else absorbs the loss — its incentives shift toward riskier behavior. Insured drivers are slightly less careful; bailed-out banks take bigger bets; an employee spending the company's money optimizes differently than their own. The hazard arises whenever the decision-maker and the loss-bearer are different people (directly related to skin in the game).
Implication Align decision-making with downside exposure; whoever makes the risky call should bear a real share of the loss.
Trap Designing systems (insurance, guarantees, budgets) that remove downside, then being surprised by the reckless behavior they invite.
Adverse selection & the lemons market robust
When one side knows more, the good options quietly exit the market.
Mechanism Akerlof's lemons: when sellers know quality and buyers don't, buyers will only pay an average price — too low for good goods and just right for bad ones — so good goods withdraw, dragging the average down, until the market fills with lemons or collapses. The same information asymmetry drives insurance death-spirals (the healthy opt out, premiums rise, more healthy leave) and hiring and dating pools. Hidden information selects for the bad.
Implication In asymmetric-information markets, look for and provide hard-to-fake quality signals (warranties, track records, verification) that let good options separate from bad.
Trap Assuming a market's average quality is representative, when adverse selection has already driven the good options out of it.
The principal-agent problem robust
Whoever acts on your behalf has their own incentives.
Mechanism A principal (owner, client, voter) delegates to an agent (manager, broker, politician) whose interests don't fully align and whose actions the principal can't fully observe. The agent optimizes for themselves — effort minimized, their fees maximized, their risk not yours — wherever monitoring is weak. It's the hidden tax in every delegation: advisors paid on transactions, managers building empires, contractors padding hours.
Implication Structure incentives so the agent wins only when you win, and monitor what you can't align; never assume delegated interests match your own.
Trap Trusting an agent's advice as if it were your own analysis, when they're paid on actions that may not serve you.
Path dependence & lock-in robust
Early accidents get frozen in; the best option rarely wins on merit alone.
Mechanism In systems with increasing returns, small early advantages — often luck or timing — compound and lock in, so the outcome depends on the path taken, not just the merits. The QWERTY keyboard, entrenched file formats, and incumbent platforms persist because switching costs and network effects freeze in whatever got there first. History matters: you can't infer the present state purely from current merits.
Implication Recognize when a standard or incumbent is entrenched by history, not quality; getting in early (or building switching costs) can matter more than being best.
Trap Assuming the dominant option won because it's superior, or expecting a better alternative to displace an entrenched, locked-in one on merit.
Berkson's paradox robust
Selecting a sample can manufacture a correlation that isn't real.
Mechanism When a sample is filtered by a condition that two independent traits both feed into, the traits appear negatively correlated within that sample even though they're unrelated in the population. Classic case: if you only date people who are either attractive or kind enough to make the cut, then within your dating pool attractiveness and kindness look inversely related — because anyone low on both was filtered out. The selection, not reality, created the pattern (a cousin of Simpson's paradox).
Implication Before trusting a correlation, ask how the sample was selected; a filter on a combination of variables can fabricate the relationship.
Trap Concluding "talented people are jerks" or "attractive things are low-quality" from a pool pre-selected on exactly those combined traits.
Normalization of deviance robust
Small tolerated deviations become the new normal — until catastrophe.
Mechanism Diane Vaughan's term from the Challenger disaster: when a risky deviation from the rules is taken and nothing bad happens, it gets accepted as normal, lowering the bar for the next deviation, and the next — until the accumulated drift produces a failure that, in hindsight, everyone should have seen. Each step feels reasonable because the last one "worked." Safety margins erode silently through repeated uneventful violations.
Implication Treat rule deviations that "worked out" as warnings, not vindications; hold the line on margins before the drift compounds.
Trap "We've skipped this check a hundred times and been fine" — the reasoning that precedes the disaster it invited.
Economic moats robust
What makes an advantage durable, not just temporarily good.
Mechanism A moat is a structural feature that protects returns from competition over time. The durable kinds: network effects (value grows with users), high switching costs (painful to leave), economies of scale (lower unit cost than rivals can reach), intangible assets (brand, patents, licenses), and a cornered resource. Absent a moat, high profits attract competition that competes them away; with one, advantage persists. Product quality alone is not a moat — it's copyable.
Implication Build or back positions with a real moat; ask what stops a well-funded competitor from replicating your advantage next year.
Trap Winning on a temporary edge (a feature, a low price) with no moat, then watching competition erode the returns.
The Shirky principle robust
Institutions try to preserve the problem they exist to solve.
Mechanism "Institutions will try to preserve the problem to which they are the solution" (Clay Shirky). An organization, profession, or process built to address a problem develops a stake in that problem's continuation — its funding, status, and jobs depend on it — so it resists solutions that would make it unnecessary, often unconsciously. The incentive to stay needed quietly outcompetes the mission to fix the thing.
Implication When an institution resists an obvious fix, check whether the fix threatens its reason to exist; follow the incentive to be needed.
Trap Expecting the body responsible for a problem to eagerly eliminate it, when its survival depends on the problem persisting.