Skip to content

Aggregate-Marginal Divergence

Prime #
623
Origin domain
Statistics & Experimental Design
Subdomain
aggregate vs marginal dynamics → Statistics & Experimental Design

Core Idea

Aggregate-marginal divergence is the structural pattern in which a widely tracked aggregate or average metric trends in one direction while the corresponding marginal or per-unit metric trends in the opposite direction. The aggregate dashboard moves favourably while the next unit — the next customer, case, student, region, or year — moves unfavourably, and the trajectory of the next unit is the trajectory of the system. Decisions made from the aggregate dashboard therefore systematically misread the underlying trajectory: the system appears to be improving while the marginal contribution is decaying, or appears to be declining while the marginal contribution is improving. The structural commitment is that the aggregate is integration over a heterogeneous mix, and the trend in that integral can diverge from the trend at the leading edge.

The cognitive failure mode is reading the aggregate as a proxy for the margin. The structural fact is that aggregate and marginal can sustainably diverge — not as a transient or an accounting artefact, but because the population mix is shifting under the average, because past contributions can compound while present ones erode, and because some subgroups are inertial enough to mask the marginal change for a long time. Naming the pattern forces the diagnostic question — what is the direction of the marginal unit? — alongside the dashboard's aggregate trend.

The divergence is structural rather than incidental, which is what makes it durable. The aggregate is a lagging summary dominated by an inertial stock of past contributions; the margin is a leading indicator carried by the latest flow. When the two point in opposite directions, the aggregate can keep moving favourably for as long as the past stock outweighs the new flow, which can be years — and during that whole window the decision-relevant quantity is invisible on the headline metric.

How would you explain it like I'm…

Happy Average, Sad Newcomer

Imagine your whole grade's average test score keeps going up, so it looks like things are great. But the newest kids joining keep doing worse and worse. The big average looks happy while the next new person is sliding downhill. Aggregate-Marginal Divergence is when the big total looks good but the next one added is actually getting worse (or the other way around).

Average Hides The Edge

Sometimes a big average or total moves one way while the very next unit, the next customer, student, or year, moves the opposite way. The dashboard says things are improving, but the newest addition is actually getting worse, and the newest addition is where the system is really heading. People who only watch the big average misread the real direction. This can last a long time, because the average is mostly made of a big pile of past results that change slowly. The new arrivals can be sliding for years before the average notices, so you have to ask: which way is the next one going?

Total Up, Margin Down

Aggregate-Marginal Divergence is the pattern where a widely tracked aggregate or average metric trends one direction while the matching marginal or per-unit metric trends the opposite direction. The dashboard moves favourably while the next unit — the next customer, case, student, region, or year — moves unfavourably, and it's the next unit that reveals the system's true trajectory. So decisions made from the aggregate systematically misread reality: the system looks like it's improving while the marginal contribution decays, or looks like it's declining while the marginal contribution improves. The structural commitment is that the aggregate is an integration over a heterogeneous mix, and the trend in that integral can diverge from the trend at the leading edge. This divergence is sustainable, not a glitch: the population mix shifts under the average, past contributions compound while present ones erode, and some subgroups are inertial enough to mask the marginal change for years. Naming it forces the diagnostic question — which way is the marginal unit going? — alongside the headline trend.

 

Aggregate-Marginal Divergence is the structural pattern in which a widely tracked aggregate or average metric trends in one direction while the corresponding marginal or per-unit metric trends in the opposite direction. The aggregate dashboard moves favourably while the next unit — the next customer, case, student, region, or year — moves unfavourably, and the trajectory of the next unit is the trajectory of the system. Decisions made from the aggregate dashboard therefore systematically misread the underlying trajectory: the system appears to be improving while the marginal contribution is decaying, or appears to be declining while the marginal contribution is improving. The structural commitment is that the aggregate is integration over a heterogeneous mix, and the trend in that integral can diverge from the trend at the leading edge. The cognitive failure mode is reading the aggregate as a proxy for the margin. The structural fact is that aggregate and marginal can sustainably diverge — not as a transient or an accounting artefact, but because the population mix is shifting under the average, because past contributions can compound while present ones erode, and because some subgroups are inertial enough to mask the marginal change for a long time. Naming the pattern forces the diagnostic question — what is the direction of the marginal unit? — alongside the dashboard's aggregate trend. The divergence is structural rather than incidental, which is what makes it durable: the aggregate is a lagging summary dominated by an inertial stock of past contributions, while the margin is a leading indicator carried by the latest flow. When the two point in opposite directions, the aggregate can keep moving favourably for as long as the past stock outweighs the new flow — which can be years — and during that whole window the decision-relevant quantity is invisible on the headline metric.

Structural Signature

the heterogeneous population (the mix being integrated over) — the aggregate metric (a lagging integral dominated by an inertial stock) — the marginal metric (the leading-edge contribution of the next unit) — the opposite-direction trends — the masking duration set by stock-over-flow — the integration-over-a-shifting-mix invariant

The pattern is present whenever these components are configured together:

  • The heterogeneous population (role). A mix of units — customers, cases, students, regions, years — whose characteristics shift under the average over time.
  • The aggregate (role). A widely tracked total or average: an integral over the mix, dominated by an inertial stock of past contributions, and a lagging summary.
  • The margin (role). The contribution of the next unit — the latest cohort, the leading edge — a leading indicator of the system's trajectory.
  • The opposite-direction trends (relation). The aggregate trends one way while the marginal contribution trends the other — sustainably, not as a transient or accounting artefact.
  • The masking duration (relation). The aggregate keeps moving favourably for as long as the past stock outweighs the new flow, which can be years — during which the decision-relevant quantity is invisible on the headline metric.
  • The integration invariant. The aggregate is integration over a shifting mix, and the trend in that integral can diverge from the trend at the leading edge — distinct from Simpson's paradox (cross-tabular reversal at one time-point) and from outlier leverage (no outliers required).

The components compose into the signature: reading a lagging aggregate as a proxy for the leading margin, when the two trend oppositely and the stock hides the marginal reversal until the flow overtakes it.

What It Is Not

  • Not aggregation. Aggregation is the operation of combining units into a total; this prime is the divergence between the aggregate's trend and the marginal unit's trend over time — a failure of reading the aggregate as a proxy for the margin, not the act of aggregating.
  • Not Simpson's paradox (simpsons_paradox). Simpson's paradox is a cross-tabular reversal of association at a single time-point under stratification; this is a temporal split between the trend of an integral and the trend at its leading edge, driven by a shifting mix.
  • Not variability. Variability is dispersion around a value; the divergence here is a directional opposition between aggregate and marginal trends, not a spread.
  • Not outlier leverage. Outlier leverage is a few high-influence points distorting an average; aggregate-marginal divergence requires no outliers and arises from a graceful shift in the population's marginal characteristics.
  • Not diminishing_returns alone. Diminishing returns is one generator of the divergence (the marginal curve falls while the integral grows), but the prime is the broader gap between any lagging aggregate and its leading margin, including mix-shift and compounding cases.
  • Not risk_pooling. Risk pooling concerns averaging away idiosyncratic risk across a population; this prime concerns the trend of the next unit diverging from the trend of the accumulated total, a temporal not a cross-sectional phenomenon.
  • Common misclassification. Steering off the aggregate dashboard — declaring health while the next unit decays — because the inertial stock keeps the headline favourable for years after the marginal reversal.

Broad Use

  • Unit economics. Total revenue grows while contribution margin per next customer turns negative — the unit-economics mirage — as past cohorts sustain the aggregate and the marginal customer costs more than they bring.
  • Epidemiology and healthcare. Total case counts fall while case-fatality in the most vulnerable subgroup rises, or hospital-wide readmission falls while the highest-acuity stratum worsens.
  • Education and macroeconomics. Average scores climb while the bottom decile falls, and aggregate productivity grows while the median firm stagnates, dragged up by superstar performance.
  • Ecology and climate. Total biomass holds steady while a keystone species crashes, and global mean temperature rises slowly while regional extreme-event frequency rises sharply.
  • Software performance. Average latency improves while tail latency worsens, so aggregate experience trends favourably while the worst-served users abandon the product.
  • AI compute and conservation. Aggregate model performance improves while marginal benefit per training unit collapses along the scaling curve, and protected-area acreage grows while each added hectare is more remote and less biodiverse.

Clarity

The construct sharpens four things ordinary metric-discourse fuses: the aggregate level (where we are), the aggregate trend (where the integral is heading), the marginal contribution (what the next unit adds), and the marginal trend (where the next unit's contribution is heading). Aggregate-trend and marginal-trend are routinely used interchangeably; the pattern insists they are distinct quantities that can move in opposite directions, and that confusing them is a systematic failure of dashboard-driven decision-making rather than an occasional slip.

The vocabulary also separates the pattern from related but distinct ones. Simpson's paradox is about associations reversing under stratification — a within-group direction flipped from the aggregate direction — at a single time-point; aggregate-marginal divergence is about trends in level moving in opposite directions over time, a temporal rather than a cross-tabular phenomenon, and the two can co-occur but are mechanistically different. Outlier leverage is about a few high-influence points distorting an average; aggregate-marginal divergence requires no outliers and can arise from a slow, graceful shift in the population's marginal characteristics. Drawing these boundaries is what keeps the pattern from collapsing into its neighbours and losing its specific diagnostic content.

Manages Complexity

A wide family of decision failures — saturated sales channels, declining returns to compute, hollowed-out trophic structure, rising means that paper over inequities, tail-latency-blind optimisation — collapses to one diagnostic: report the marginal alongside the aggregate. The diagnostic is the same regardless of whether the aggregate is revenue, biomass, a mean score, or cumulative compute, and it converts a scattered set of "the dashboard looked fine until it didn't" failures into a single, checkable practice.

The intervention catalogue is portable. Disaggregate by cohort or subgroup and report marginal trends alongside aggregate trends, since the marginal is what is actually trending. Report leading-edge metrics — tail latency, worst subgroup, latest cohort, marginal-unit yield — as direct windows on the margin. Set guardrail metrics — thresholds on the subgroup whose movement the average suppresses, so the aggregate may improve only if the bottom does not regress. Reframe success criteria to target the marginal where the marginal is the decision-relevant quantity. And audit the aggregation cadence, since long windows mask recent marginal reversal that shorter rolling windows or explicit cohort-trend reporting would catch. Each move makes the marginal visible or decision-bearing, which is the single structural correction the pattern prescribes.

Abstract Reasoning

Recognising the divergence enables several distinct kinds of reasoning. Reasoning about the trajectory of the next unit, distinct from the trajectory of the cumulative aggregate, treats the marginal contribution as the leading indicator and the aggregate as a lagging summary. Reasoning about the duration over which an aggregate can mask marginal reversal: the lag depends on the size of the past stock relative to the new flow, so large stocks with small flows can hide a marginal reversal for years. Reasoning about scaling-law transitions: in any system with diminishing returns, the aggregate trend necessarily diverges from the marginal trend at some point, and naming the pattern pre-empts surprise at the transition.

Two further modes deepen the analysis. Reasoning about the politics of dashboards: which stakeholders benefit from aggregate-only reporting, since those who control the reporting cadence are often insulated from the marginal reversal until it is too late. And counterfactual reasoning: what would the aggregate look like in five years if the current marginal trajectory holds, since the marginal trend projected forward eventually overcomes any stock advantage. Together these convert a favourable headline number into a set of pointed questions about the leading edge, the masking duration, and the projected crossover — which is exactly what keeps a decision-maker from being reassured by an aggregate that is concealing the system's real direction.

Knowledge Transfer

Because aggregate-versus-marginal is mathematically primitive — the trend of an integral over a heterogeneous mix versus the trend at that integral's leading edge — the pattern and its repairs transfer across substrates that share no content. The startup discipline of asking "what is the next customer's contribution margin?" transfers to conservation audit as "what is the next hectare's biodiversity contribution?", both replacing an aggregate headline with marginal reality. The site-reliability discipline that worst-case latency matters more than average transfers to healthcare-quality reporting as worst-stratum outcomes reported alongside aggregate metrics. The scaling-law insight that the next training unit yields less than the last transfers to any R&D portfolio whose total spend grows while marginal return shrinks, where the aggregate "investment in innovation" misreads the trajectory.

Two transfers are especially clean. Cohort analysis from consumer products — measuring per-cohort retention — transfers directly to epidemiological surveillance measuring per-wave or per-variant fatality, where aggregate case counts hide the marginal vulnerability shift. And the microeconomic intuition that the marginal unit becomes inframarginal as a system saturates transfers to policy evaluation: a programme that has exhausted its easy-to-help population must be judged on its marginal reach, not its aggregate enrolment. A practitioner who has caught the unit-economics mirage in one business can recognise, in an ICU-versus- hospital rate or an AI scaling curve, the same lagging aggregate masking a reversed margin, and can reach for the same repair — report the marginal alongside the aggregate and set a guardrail on the leading edge. The substrates differ — revenue, fatality, latency, biomass, compute — but the integration-over-a-shifting-mix structure and its marginal-reporting repair are preserved, so the reasoning carries from one to the next without re-derivation.

Examples

Formal/abstract

A neural-network scaling law is the formal worked instance, because the divergence is mathematically forced rather than incidental. The heterogeneous population is the sequence of training units (compute, data, or parameters) added over a project's life. The aggregate is total model capability — a lagging integral over all units invested so far, dominated by the large inframarginal stock of earlier, high-yield training. The margin is the capability gained from the next unit of compute — the leading edge. The opposite-direction trends are the crux: aggregate capability keeps climbing (the integral still grows) while the marginal return per unit falls along the diminishing-returns curve, so the headline "the model is getting better" coexists with "each new unit buys dramatically less than the last." The masking duration is set by stock-over-flow: because the accumulated capability stock is huge relative to any single increment's contribution, the aggregate can keep rising impressively for a long time even as the marginal return approaches the floor. The integration invariant is what the example proves with force: in any system with diminishing returns, the aggregate trend necessarily diverges from the marginal trend at some point — so naming the pattern pre-empts surprise at the scaling transition, where teams that tracked only aggregate capability are blindsided by the collapse in marginal efficiency. This is distinct from Simpson's paradox (a cross-tabular reversal at one time-point) and from outlier leverage (no outliers required). The repair is to report the leading-edge metric — marginal capability per unit compute — alongside the aggregate. Mapped back: the training units are the heterogeneous population, total capability is the lagging aggregate, capability-per-next-unit is the leading margin, and the diminishing-returns curve forcing their divergence is the integration invariant.

Applied/industry

The unit-economics mirage in a subscription business is the applied worked case, exercising a commercial domain. The heterogeneous population is the firm's customers, acquired in successive cohorts whose characteristics shift over time. The aggregate is total revenue — a lagging integral dominated by an inertial stock of early, cheaply-acquired, loyal cohorts. The margin is the contribution margin of the next customer: what the marginal acquired customer brings in minus what they cost to acquire and serve. The opposite-direction trends are the mirage: total revenue grows quarter over quarter (the dashboard looks healthy) while the marginal customer's contribution margin has turned negative, because the easy-to-reach market is exhausted and each new customer now costs more to acquire and churns faster than they pay back. The masking duration is the danger: the large stock of profitable legacy cohorts sustains the aggregate for years after the marginal economics inverted, during which the decision-relevant quantity is invisible on the headline revenue chart. The integration invariant explains why this is durable rather than a transient accounting artefact — the aggregate is integration over a shifting customer mix, and its trend can diverge sustainably from the leading edge. The repair is the pattern's prescription: disaggregate by cohort and report marginal contribution alongside aggregate revenue, set a guardrail metric (the aggregate may grow only if marginal contribution stays positive), and project the crossover — what the aggregate looks like in five years if the current marginal trajectory holds, since the marginal trend eventually overcomes any stock advantage. Two further genuine domains share the structure: software performance, where average latency improves while tail latency worsens and the worst-served users abandon the product, and education, where average test scores climb while the bottom decile falls. Mapped back: the customer cohorts are the heterogeneous population, total revenue is the lagging aggregate, the next customer's contribution margin is the leading margin, and the legacy-cohort stock hiding the marginal reversal is the masking duration the integration invariant produces.

Structural Tensions

T1 — Lagging Aggregate versus Leading Margin (temporal). The aggregate is a lagging integral dominated by an inertial stock; the margin is the leading edge carried by the latest flow. The decision-relevant quantity is the margin, but the visible one is the aggregate. The failure mode is steering off the dashboard — declaring health while the next unit decays, because the stock keeps the headline favorable for years after the marginal reversal. Diagnostic: alongside the aggregate trend, always compute the direction of the latest cohort's contribution, and ask whether the two point the same way.

T2 — Aggregate-Marginal Divergence versus Simpson's Paradox (scopal). The prime is explicitly distinct from Simpson's paradox (a cross-tabular reversal at one time-point) and from outlier leverage (no outliers required) — its divergence is a temporal trend split driven by a shifting mix. The failure mode is misdiagnosing the mechanism: applying a Simpson's-style subgroup decomposition to a margin-versus-stock dynamic, or vice versa, and reaching for the wrong correction. Diagnostic: ask whether the reversal is between subgroups at one time (Simpson) or between the integral and its leading edge over time (this prime).

T3 — Sustainable Divergence versus Transient Artefact (measurement). The structural claim is that aggregate and margin can diverge sustainably — not as a transient or accounting artefact — because the mix shifts and past contributions compound. The failure mode is dismissing a genuine marginal reversal as noise that will mean-revert, waiting for the aggregate to confirm what the margin already showed. Diagnostic: ask whether a mechanism (mix shift, compounding stock) makes the divergence structural and durable, or whether it is a one-period blip that the next aggregate reading will absorb.

T4 — Masking Duration versus Stock/Flow Ratio (scalar). How long the aggregate hides the marginal reversal is set by the ratio of inertial stock to new flow — a large legacy base masks a marginal decline for years; a small one exposes it fast. The failure mode is assuming a fixed lag, being lulled by a slow-moving aggregate in a stock-heavy system and surprised by a fast crossover in a flow-heavy one. Diagnostic: estimate the stock-over-flow ratio to bound how long the headline can stay favorable after the margin turns, rather than treating the masking window as constant.

T5 — Margin as Leading Indicator versus Margin Noise (sign/direction). The margin leads the system's trajectory, but the leading edge is also the noisiest, smallest-sample slice — the latest cohort has the least data. Over-reading the margin trades the aggregate's lag for the margin's variance. The failure mode is whipsawing on a noisy marginal signal, reorganizing strategy around a few recent units whose trend is sampling noise. Diagnostic: ask whether the marginal trend is statistically distinguishable from noise given the leading cohort's small size, before treating it as the system's true direction.

T6 — Which Margin (scopal). "The margin" presumes a single well-defined next unit, but a system has many possible marginal slices — next customer, next region, next dollar, next time-period — and they can trend differently. The failure mode is picking one marginal definition that confirms a prior and calling it the trajectory, when a different equally-valid margin tells the opposite story. Diagnostic: ask which marginal unit is decision-relevant for the choice at hand, and whether the divergence holds across the marginal definitions that matter, not just the convenient one.

Structural–Framed Character

Aggregate-marginal divergence sits at the structural pole of the structural–framed spectrum — a paradigm structural prime, aggregate 0.0 with every diagnostic reading zero. Its content is mathematically primitive: the trend of an integral over a heterogeneous mix versus the trend at that integral's leading edge, with a stock-over-flow masking duration. The whole pattern is a fact about lagging integrals and leading margins, and it applies wherever a heterogeneous population is summed — which the entry exhibits in revenue, case-fatality, biomass, test scores, latency, and compute scaling curves.

Every diagnostic points one way. The pattern carries no home vocabulary that must travel with it: the identical structure is told as the unit-economics mirage in business, the keystone-crash-under-stable-biomass in ecology, tail-latency-under-improving-average in software, and the scaling-law transition in AI, each in its own field's words — vocab_travels is 0. It carries no inherent approval or disapproval; a divergence between aggregate and marginal trends is a value-neutral mathematical fact until you specify which direction is desired (evaluative_weight 0). Its origin is formal — the aggregate-versus-marginal distinction is primitive calculus over a population, owing nothing to a human institution (institutional_origin 0). It runs in physical and biological substrates indifferently: total biomass can hold while a keystone species crashes, and global mean temperature can rise while regional extremes accelerate, with no reasoning agent required for the divergence to obtain (human_practice_bound 0). And invoking it merely recognizes a relationship already present between an integral and its leading edge, rather than importing an interpretive frame (import_vs_recognize 0). Even where the entry reaches into the "politics of dashboards" — who benefits from aggregate-only reporting — the structure is the same integration-over-a-shifting-mix fact; the stakes attach to the reading and the decision, not to the pattern, which is exactly why it grades the same as feedback at the structural pole.

Substrate Independence

Aggregate-marginal divergence is about as substrate-independent as a prime can be — composite 5 / 5 on the substrate-independence scale. Its domain breadth is maximal: the pattern in which an aggregate trends one way while the next unit's contribution trends the other recurs with identical force across unit economics (total revenue growing while contribution margin per next customer turns negative), epidemiology and healthcare (total cases falling while case-fatality in the most vulnerable subgroup rises), education and macroeconomics (average scores climbing while the bottom decile falls; aggregate productivity growing while the median firm stagnates), ecology and climate (total biomass steady while a keystone species crashes; global mean temperature rising slowly while regional extremes spike), software performance (average latency improving while tail latency worsens), and AI compute and conservation (aggregate model performance improving while marginal benefit per training unit collapses) — physical, biological, economic, and computational substrates without exception. Its structural abstraction is maximal: the distinction between an integral over a heterogeneous mix and the trend in the integrand at the margin is mathematically primitive, a purely relational fact carrying no normative or institutional content. Transfer evidence is maximal and concrete: seven-plus domain instances are each load-bearing and the same formal relation (level versus derivative-at-the-margin over a heterogeneous population) is what is being read in every one, so the diagnosis is recognized rather than translated. Maximal breadth, maximal abstraction, and heavily documented transfer all line up, making this one of the catalog's canonical 5s.

  • Composite substrate independence — 5 / 5
  • Domain breadth — 5 / 5
  • Structural abstraction — 5 / 5
  • Transfer evidence — 5 / 5

Relationships to Other Primes

One-hop neighborhood: parents above, mutual partners to the right, children below.Aggregate-MarginalDivergencecomposition: AggregationAggregation

Parents (1) — more general patterns this builds on

  • Aggregate-Marginal Divergence presupposes Aggregation

    The divergence is a diagnostic about READING an aggregate: it presupposes aggregation (the collapsing operation) and adds a heterogeneous mix, a stock/flow masking duration, and the opposite-direction-trends invariant. The file: 'presupposes aggregation but adds...'.

Path to root: Aggregate-Marginal Divergence → Aggregation → Micro Macro Linkage

Neighborhood in Abstraction Space

Aggregate-Marginal Divergence sits in a moderately populated region (47th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Throughput, Efficiency & Distribution (14 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-06-14

Not to Be Confused With

The embedding-nearest neighbour is aggregation (similarity 0.86), and the relationship is operation-to-failure-mode. aggregation is the act of combining many units into a summary total or average — a constructive operation. Aggregate-marginal divergence is a diagnostic about reading that summary: it names the condition in which the aggregate's trend and the marginal unit's trend move in opposite directions over time, so that the aggregate (a lagging integral dominated by an inertial stock) is a misleading proxy for the margin (the leading edge that is the system's true trajectory). The prime presupposes aggregation but adds a heterogeneous mix, a stock/flow masking duration, and the opposite-direction-trends invariant — none of which belong to aggregation itself. A practitioner who collapses the prime into aggregation treats the headline total as the object of interest, exactly the error the prime exists to flag: steering off a favourable aggregate while the next unit decays.

The most important and explicitly-flagged confusion is with simpsons_paradox (T2). Both are "the aggregate misleads" patterns driven by a population mix, and both can co-occur, which is why they are routinely conflated. But the mechanisms are distinct. Simpson's paradox is a cross-tabular phenomenon at a single time-point: an association that holds within every subgroup reverses when the subgroups are pooled, because a confounder is distributed unevenly across them. Aggregate-marginal divergence is a temporal phenomenon: the trend of an integral over a shifting mix diverges from the trend at its leading edge, as a large inertial stock masks a reversal in the marginal flow. The distinction is decision-relevant: a Simpson's-style remedy is to stratify and inspect within-group associations at the same moment, while the aggregate-marginal remedy is to report the leading-edge metric (latest cohort, marginal yield, tail) alongside the aggregate over time. Applying a subgroup decomposition to a margin-versus-stock dynamic, or vice versa, reaches for the wrong correction.

A third worth separating is diminishing_returns. The two are tightly linked — in any system with diminishing returns the aggregate trend necessarily diverges from the marginal trend at some point, so diminishing returns is one generator of the pattern (the formal example turns on exactly this). But aggregate-marginal divergence is the broader prime: it also arises from a shifting population mix and from compounding past contributions that have nothing to do with a falling marginal-return curve (the unit-economics mirage is driven by an exhausted easy-to-reach market and adverse cohort selection, not a smooth diminishing-returns function). The contrast tells the practitioner that recognising diminishing returns is sufficient to predict the divergence in some cases but not necessary for it — and that the repair (report the margin alongside the aggregate, set a leading-edge guardrail) is the same regardless of whether the generator is a diminishing-returns curve or a mix shift.

For a practitioner these distinctions decide the diagnostic move. aggregation is merely the operation producing the headline; simpsons_paradox calls for cross-sectional stratification at one time; diminishing_returns is one generator predicting the split; and aggregate-marginal divergence calls for the temporal move — compute the direction of the latest cohort's contribution alongside the aggregate trend, estimate the stock/flow ratio bounding the masking window, and set a guardrail on the leading edge.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.