Selection-Visibility Gate¶

Version: v4 · 2026-08-30 · History
Prime #: 1450
Origin domain: Biology & Ecology
Subdomain: evolutionary selection → Biology & Ecology
Related primes: Natural Selection

Core Idea¶

A selection-visibility gate is the stage that decides which of a variant's consequences the selecting process is able to encounter. Selection is usually described as acting on variants; it acts on consequences, and only on those consequences that are expressed, exposed, or otherwise accessible at the moment selection operates. ^[1] A trait whose effects are real but never surface at that moment contributes nothing to its bearer's retention: for selective purposes, the variant does not have the trait.

Five roles carry the pattern. A generator supplies heterogeneous variants. Their consequences genuinely differ. An access stage — expression, exposure, representation, sampling, reporting — lets some of those consequences through and masks the rest. A selector acts differentially on what has been let through. Retention shifts accordingly, and the surviving pool ends up enriched for variants that cleared the gate rather than for variants that were best. ^[2]

The load-bearing word is upstream. Visibility here mediates selection; it is not missing measurement noticed afterwards by someone studying the survivors. The counterfactual makes the claim testable: hold the generator and the selection pressure fixed and change only the aperture. If the surviving composition moves, a gate was doing work. If it does not, what looked like a gate was ordinary attrition, noise, or an artefact of how the survivors were later described. ^[3]

What makes the prime worth having is the direction of the resulting error. Invisible variants are not selected against; they are not selected on at all. They drift, persist neutrally, or disappear for reasons unrelated to their value, so an absence in the survivor pool is ambiguous between was eliminated and was never legible. Telling those two readings apart is most of the analytical work the gate licenses.

Structural Signature¶

The pattern is: heterogeneous variants → an access stage that expresses some consequences and masks others → a selector acting only on the exposed subset → a survivor pool enriched for legibility as well as for merit.

Recurring features:

The access stage sits strictly upstream of the selector and is causally independent of it; the selector need not know the stage exists, and typically does not.
Expression is graded rather than binary — penetrance, sampling rate, reporting probability and coverage all set an aperture somewhere between zero and one.
Effective pressure on a trait is its true consequence discounted by its probability of expression, so a large effect that rarely surfaces behaves exactly like a small effect. ^[4]
Masked variants accumulate as a reservoir — unselected rather than purged, and available to be exposed in bulk if the aperture later opens.
The admission set is a property of the system, not of any variant: two variants with identical consequences fare differently when only one of them is legible.
The selector's own performance is scored on the exposed subset, so an aperture can be badly mis-set for a long time while producing no local signal of failure.
Changing the aperture redistributes survival without touching the generator — same variation, different survivors.

What It Is Not¶

This is not a claim about concealment or bad faith. Nobody need be hiding anything for a gate to operate; one exists wherever expression is partial, whether that partiality was designed, inherited or accidental. A recessive allele, an untested error branch, an unreported side effect and a shift no inspector happens to visit are all apertures, and only some of them have anyone to blame. ^[5]

Nor is it the general observation that information is incomplete. Everything is incompletely observed. The prime asks something much narrower: whether the incompleteness sits between the variation and the pressure, so that it changes which variants persist. Incompleteness downstream of selection changes what can be said about the survivors; it does not change who they are.

It is also not an assertion that the aperture is wrong or that it ought to be opened. Some closures are load-bearing — they shelter variation from premature elimination, keep cheap noise from consuming expensive attention, or hold a system steady while slow evidence accumulates. The prime is descriptive. It says the survivor pool carries the aperture's fingerprint; whether that fingerprint counts as a defect depends entirely on what the system is for.

Finally, it does not imply that the masked traits are the valuable ones. A gate enriches for legibility, and legibility is uncorrelated with merit by default. Sometimes the hidden variants are the good ones, sometimes the ruinous ones, and usually both in proportions nobody can state without opening the aperture and looking.

Broad Use¶

Genetics and development: recessivity, incomplete penetrance and condition-dependent expression all mean an allele's fitness consequence reaches selection only some of the time, which is why populations carry large reservoirs of standing variation that no pressure has ever evaluated. ^[6]

Software engineering and reliability: an automated test suite, a review surface and a monitoring stack jointly define which defects can express as a visible failure. Behaviour on uncovered paths, in unlogged states and under unsimulated load is outside the loop entirely.

Organizations and careers: promotion acts on artefacts — shipped features, closed deals, named launches. Work whose success consists of an incident that never happened, a dependency that never broke, or a conflict resolved before it reached anyone's queue produces no artefact to be promoted on.

Safety, health and regulation: harms enter the record only through a reporting channel, and enforcement acts on inspected sites. A hazard concentrated in populations that do not report, or in facilities the sampling frame rarely reaches, is not lightly penalized — it is unpriced.

Recommendation and marketplace systems: an item accrues feedback only where it is shown, and is shown on the strength of feedback it has already accrued, so the exposure policy determines which items ever become eligible to be judged good. ^[7]

Model evaluation: a benchmark is an aperture. Systems are retained, tuned and funded on the slice of behaviour the evaluation surface renders, and behaviour outside that slice varies freely without cost.

Cultural transmission: a practice is copied only if it is performed where copiers can watch. Tacit, private and preparatory work — the rehearsal, not the performance — transmits far worse than its contribution warrants.

Risk and finance: exposures that express only in a regime not yet sampled survive every review conducted in the regimes that have been.

Clarity¶

The confusion this dissolves is the habit of reading a survivor pool as a ranking — it survived, so it was better; it vanished, so it was worse. The gate inserts a third and usually neglected possibility: it was never in the competition, because the property that would have decided the competition never reached the judge.

That reframing does specific work on a specific recurring puzzle: if this trait is so costly, why has a long-running, competent process not eliminated it? The answer is frequently that the process cannot see it, and no amount of additional pressure will help, because pressure multiplies a term that is already near zero.

It also separates two things routinely collapsed into each other: a selector's values and a selector's reach. A promotion process that never rewards reliability work is not necessarily one that disvalues reliability work; it may be one in which reliability work leaves nothing behind to reward. The distinction is not academic, because the remedies do not overlap. Re-weighting the criteria addresses a values problem and leaves a reach problem untouched, which is one reason a carefully designed rubric can leave promotion outcomes unchanged. ^[8]

Manages Complexity¶

You can stop enumerating the variant distribution and stop reconstructing the selector's full preference ordering. To get the direction in which a population's composition will move, two objects mostly suffice: what the generator produces, and what the aperture admits. The selector's sophistication, the strength of the pressure and the number of rounds mainly govern how fast and how far the shift goes — with the exception that where the pressure surviving the aperture is small relative to drift there is no direction left to sharpen, as the worked example below shows. ^[4]

You can also stop tracking traits one at a time. What matters is the interaction of a trait with its expression probability, which collapses a long list of individually distinct properties into two classes — those that register at the gate and those that do not — with only the first needing detailed modelling.

And when the aperture already accounts for an outcome, you can stop auditing the selector. Investigating judgment is slow, contested and personal; reading an admission set is fast, and it is written down somewhere. That asymmetry is what makes the gate a good first hypothesis rather than a last resort: it is the cheapest of the available explanations to check.

Abstract Reasoning¶

Working forwards, the procedure has four steps. Name the generator. Name the selector and the pressure it applies. Name the access stage and write out its admission set explicitly — what reaches the selector, in what units, at what rate. Then rank candidate traits by the ratio of consequence to expression probability; the top of that list is where the gate is costing the most, and it is rarely where attention is.

The verification step is the aperture counterfactual. Vary expression while holding generator and pressure fixed — across sub-populations that happen to differ in reporting rate, across periods before and after a coverage change, or by direct intervention — and check whether survivor composition moves with it. Dropout uncorrelated with expression is attrition, not a gate. ^[9]

Working backwards is the diagnostic that earns the prime its keep. A costly trait that persists over many rounds under an apparently competent selector admits exactly three explanations: the trait is not actually costly, the selector is not actually competent, or the cost does not express at the gate. The third is the only one that is cheap to test, and it is tested by finding a corner of the system where expression happens to be higher and asking whether the trait is rarer there.

Knowledge Transfer¶

Four things travel intact across substrates: the role skeleton, the direction of enrichment, the aperture counterfactual as a test, and the reservoir prediction — that masked variation accumulates rather than dissipating, and can express in bulk when conditions change. Those four support real inference in genetics, in engineering and in institutions without adjustment.

Several things do not travel, and assuming they do is where the analogy breaks. Whether the aperture has an author is substrate-specific: a buffered genome has no designer,^[3] an evaluation suite does, and only the authored case supports the question who chose this admission set and can they be made to change it. Whether variants can act on their own visibility is likewise specific — molecules cannot manage their penetrance, but employees, firms and models can,^[10] so the entire family of gaming, performance and strategic legibility effects transfers only to substrates with strategic variants. Reversibility differs too: cryptic genetic variation remains in the population,^[11] while items removed from a moderation queue are gone — that second half of the contrast is asserted here rather than documented. And the timescale is not portable at all, since the same structure runs over generations in one substrate and over hours in another; that contrast too is this entry's own observation, offered without a source.

Examples¶

Formal/abstract¶

Take a single locus with alleles A and a in a finite population. A carries a fitness consequence of size s relative to a, but that consequence is expressed only in environment E, which obtains with probability p in any given generation; otherwise A and a are phenotypically identical. The pressure selection actually applies is not s but the product s·p. With s = −0.10 and p = 0.02, the realized coefficient is −0.002, which in a population of moderate size falls below the drift barrier: A is effectively neutral, and its frequency wanders on demographic noise rather than on its consequence. ^[12] It is not being maintained by balancing selection and it is not being purged. It is unselected.

Now let the environment shift so that p approaches 1. The realized coefficient jumps fiftyfold, and the population pays a cost proportional to the frequency the allele reached during the era when nothing could see it. The reservoir expresses all at once, and the size of the bill is set not by the shift but by how long the aperture stayed shut.

The model has a second-order move. Introduce a modifier locus M that lowers p for alleles of this class — a buffering or canalizing function. M is itself under selection, and it is favoured precisely when the alleles it masks are deleterious, because masking removes their expressed cost in the present generation. So M spreads, the aperture narrows further, and the reservoir deepens. Selection on M is selection on the visibility set itself.

Mapped back: the generator is mutation, the access stage is conditional expression, the selector is fitness, and the survivor pool is the allele frequency distribution. The model makes the central asymmetry quantitative — a large-consequence trait behaves like a small-consequence trait whenever the discount factor is small enough, and no increase in selection intensity repairs this, because intensity multiplies a term already near zero. The modifier locus shows the further point that an aperture can itself be a selected object, and that it can be pushed in a direction which improves every current generation while enlarging a liability none of them will pay. ^[3]

Applied/industry¶

A platform team merges around two hundred changes a week. Each is reviewed by a peer and must pass an automated suite before it lands. Defects on covered paths express as a red build within minutes and are removed the same day. Defects on uncovered paths — the error branch of a rarely-hit endpoint, the migration that runs once per release, the timeout handler behind a mocked dependency — pass review, pass the suite, and land. Over a year the covered surface converges towards defect-free while the uncovered surface silently accumulates. Every dashboard the team watches improves throughout, because pass rate, escape count and time-to-green are all computed on the exposed subset. ^[13]

Two second-order effects follow. Management, wanting a handle on the problem, sets a coverage percentage target — and coverage is visible at the gate while correctness is not, so the suite fills with tests that execute lines without asserting behaviour, and the number rises while nothing improves. Meanwhile engineers quietly stop maintaining the uncovered regions. Not from negligence: effort there returns no signal of any kind, so it drifts to the places where effort registers, exactly as it should under the incentives actually in force.

Mapped back: the variants are code changes, the access stage is the suite plus the review surface, and the selector is the merge-or-revert decision applied repeatedly. The suite is not a measurement of quality — it is the operational definition of which quality differences are permitted to affect retention, which is why an outage in an uncovered region is not evidence that the process degraded but the reservoir expressing on schedule. The coverage target is the proxy failure in miniature: a property was chosen because it could pass the gate, and selection promptly optimized the property instead of the thing it stood for. ^[14]

Structural Tensions¶

T1 — Widening the aperture makes it gameable. Every widening announces what now counts, and any variant able to act on its own visibility will invest in clearing the new aperture rather than in producing the consequence the aperture was built to reveal. Narrow gates hide real differences; wide, published gates get performed to. A system can choose which of the two failures it prefers, but it cannot escape both — and the choice is normally made once, at design time, then inherited by people who no longer know a choice was made.

T2 — Proxies displace the traits they stand for. A selector can act only on representations, so someone must pick one: coverage for correctness, citations for insight, response time for care, tenure for judgment. The proxy is legible and the construct is not, so the proxy absorbs the entire selective pressure. Effort reallocates towards it, the correlation that justified it decays, and the gate goes on operating faithfully on a signal that has stopped standing for anything. The tighter the selection, the faster the decoupling — so the best-run process corrodes its instrument soonest.

T3 — Invisible load-bearing traits decay unselected. Traits whose function is to prevent bad outcomes leave no positive trace; their success is an absence, and absences do not clear apertures. Because nothing registers them, they are neither rewarded nor defended, and they erode through ordinary attrition — the maintainer who retires, the inspection that gets trimmed, the review ritual quietly dropped. The erosion is silent until a regime arrives that would have needed them. Systems therefore shed their safeguards fastest during their calmest years, which is exactly when the shedding looks like efficiency.

T4 — The visibility set has authors. Every admission set was written by someone with interests: which incidents are reportable, which work is recorded, which metrics reach the committee, which tests are mandatory. That authorship is rarely contested, because an aperture presents itself as infrastructure rather than as policy, and its authors are usually not the people whose variants it filters. Arguing about selection criteria while the aperture stays fixed relitigates a decision already taken upstream — and whoever set the aperture wins the argument without having to appear in it.

T5 — Gate cadence versus consequence horizon. A gate admits whatever has expressed by the moment selection acts, so consequences slower than the selection cycle are structurally invisible no matter how large they eventually become. Shortening the cycle sharpens the pressure and shrinks the admissible horizon; lengthening it admits slow effects but weakens and blurs the pressure acting on fast ones. No single cadence sees both. The standard repair — running two selectors at different tempos — buys horizon coverage at the price of two admission sets that will disagree.

T6 — Masking as protection versus masking as debt. A closed aperture is not simply a defect. It shelters variation from premature elimination, spares expensive attention from cheap noise, and lets a system hold steady while slow evidence accumulates; these are real functions and the reason many closures were built deliberately. But the same closure that preserves optionality also accumulates unpriced liability, and from inside the system the two are indistinguishable, because both present as quiet. Opening the aperture releases the reserve and the debt together, rarely in the order anyone planned for.

Structural–Framed Character¶

Selection Visibility Gate sits at the structural end of the structural–framed spectrum with an aggregate of 0.0 — all five criteria read exactly zero. The skeleton is two-stage: a generating process supplies heterogeneous variants; an expression or accessibility stage upstream of selection lets only some of their selection-relevant consequences through; and the selector retains differentially on what it can encounter, without reconstructing hidden state. The consequence is a retained record enriched for variants that clear the gate, relative to the distribution generated.

No criterion lifts this prime off zero, so what matters is why the likeliest candidate does not. Human-practice-bound reads 0.0 because moderation, testing and promotion are only some instances; phenotypic expression is equally one, and there nobody administers the gate. Invisible candidates persist neutrally, disappear stochastically, or fail to propagate — none of which requires a practice.

Domain vocabulary reads 0.0: variant, expression, selector and retention are the pattern's own terms, not a field's. Evaluative weight is 0.0 — the gate describes what a selector can act upon, with no verdict on the enrichment. Institutional origin is 0.0: no publisher or review body is constitutive, only a masking stage. Import-vs-recognize is 0.0, since genetics and algorithm evaluation alike recognize a real two-stage mechanism.

A clean zero means the roles survive substrate substitution with no imported frame. The care it demands is keeping it distinct from its neighbours: statistical selection bias concerns inference from a conditioned sample, while observability asks whether internal state is reconstructible from outputs. Here the visibility stage must be real and upstream.

Substrate Independence¶

Selection-Visibility Gate is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. The invariant is an access stage sitting upstream of a selector: variants differ in their consequences, an aperture expresses some of those consequences and masks the rest, and the survivor pool is enriched for legibility as much as for merit. Incomplete penetrance in genetics, what a culture happens to transmit, moderation queues, test coverage, promotion systems and benchmark suites all carry the same roles. What keeps it below the ceiling is where the clean cases sit: the defining counterfactual needs an aperture someone can actually move while holding variation and pressure fixed, and outside settings offering that handle, gates are usually inferred from survivors after the fact.

Composite substrate independence — 4 / 5
Domain breadth — 4 / 5
Structural abstraction — 4 / 5
Transfer evidence — 4 / 5

Relationships to Other Abstractions¶

Current abstraction Selection-Visibility Gate Prime

Parents (1) — more general patterns this builds on

Selection-Visibility Gate is a kind of Natural Selection Prime

A selection-visibility gate is natural selection specialized by an upstream access stage that determines which variant consequences the selector can act on.

Children (2) — more specific cases that build on this

Haldane's Sieve Domain-specific is a kind of Selection-Visibility Gate

Haldane's sieve is selection visibility specialized to a new beneficial allele whose heterozygous expression scales with dominance while it is rare.
Baldwin Effect Domain-specific is part of, conditional Selection-Visibility Gate

In the exposure branch of Baldwin dynamics, plastic expression makes otherwise latent heritable acquisition differences visible to selection.

Hierarchy path (1) — routes to 1 parentless root

Selection-Visibility Gate → Natural Selection → Selection

Neighborhood in Abstraction Space¶

Selection-Visibility Gate sits among the more crowded primes in the catalog (32^nd percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.

Family — Stimulus-Response Adaptation & Conflict Timing (22 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With¶

Selection-Visibility Gate must first be distinguished from Natural Selection, its parent. Natural selection is the general engine: a population of varying, heritable variants, a pressure that retains some differentially, and composition shifting over rounds. It says nothing about whether the pressure can reach a given variant's consequences, and in its ordinary statement that reach is silently assumed. The gate is that assumption converted into a variable. Every gate case is a case of natural selection; the converse fails whenever expression is uniform across variants, which is precisely the regime in which the parent's standard predictions hold. Reach for the parent when the live question is which trait wins, and for the child when the live question is which traits were ever in the running.

It is not Selection Bias, though the names invite the collapse. Selection bias is about an analyst's sample: a dataset conditioned on some variable supports conclusions the underlying population would not. The conditioning sits between the world and the observer, and the repairs are statistical — reweight, model the conditioning, sample differently. A visibility gate sits between the world and the selector, and no statistical repair touches it, because the population itself has already been reshaped. The two look identical in a plot and diverge completely in remedy: correcting selection bias tells you what was true, while opening an aperture changes what will be true. A study with impeccable sampling can describe, faithfully, a population that a gate has thoroughly deformed.

It is not Observability, which asks whether a system's internal state can be reconstructed from its outputs — a question about the sufficiency of information for an inference, settled by the structure of the mapping. The gate asks something coarser and more consequential: whether a consequence is accessible enough to affect retention at all. A state can be perfectly observable in principle and still fall outside the gate, because nobody performs the reconstruction at the moment selection acts, or because performing it takes longer than the selection cycle allows. Observability concerns what could be known; the gate concerns what actually gets acted upon.

Gatekeeping is closer in imagery and further in structure. It describes an actor at a choke point exercising selective passage, with the arbiter's judgment as the mechanism and the arbiter's criteria and incentives as the interesting variation. A visibility gate needs no arbiter and usually has none — recessivity, sampling rates and untested branches gate nothing on purpose and answer to nobody. Where an arbiter does exist the two can co-occur, but even then they sit at different stages: gatekeeping controls which candidates pass, while the visibility gate controls which of a candidate's properties are present to be judged at all, including for the candidates who sail through.

Invisible Affordance shares the masking but not the consequence. There, a capability genuinely exists while the surfaces a prospective user samples carry no signal of it, so the user's effective option set excludes something real. Both patterns involve a real thing hidden behind the surfaces on offer. What differs is who is deprived and what follows from the deprivation: an invisible affordance shrinks an agent's choices, and the cost is use that never happens, whereas the gate shrinks a selector's grounds for retention, and the cost is a survivor population assembled on the wrong basis. Revealing an affordance changes what one user does next; opening an aperture changes what an entire population comes to consist of.

Goodhart's Law and Proxy-Target Divergence are downstream of the gate rather than alternatives to it. Goodhart describes the collapse of a proxy's correlation with its construct once the proxy is placed under binding optimization pressure; proxy-target divergence describes an apparatus still operating on a proxy after that relationship has quietly decoupled. Both concern a measure standing in for something else, and the gate explains why there is always a measure to corrupt: a selector can act only on representations, so some proxy is inevitably chosen, and that choice fixes which corruptions are even available. Goodhart tells you what happens to the proxy under pressure; the gate tells you why a proxy had to exist.

Finally, Signaling and Screening are interventions on an aperture, not the aperture itself. Signaling is a variant making its hidden quality visible at a cost that lower-quality variants could not bear; screening is the selector restructuring the choices on offer so that variants reveal themselves. Both presuppose strategic agents on at least one side. The gate is the underlying condition that makes either worth attempting, and it is present wherever expression is partial — including in populations of molecules that can signal nothing to anyone. Signaling and screening are two techniques among many for widening an aperture; neither is required for a gate to exist, and neither is required for one to bite.

Solution Archetypes¶

No catalogued solution archetypes reference this prime yet.

Notes¶

The prime's sharpest use is diagnostic rather than descriptive. Given a persistent, costly trait under an apparently competent selector, it supplies a third hypothesis alongside "the trait is not really costly" and "the selector is not really competent" — and that third hypothesis is the only one of the three that is cheap to test, since it predicts the trait will be rarer wherever expression happens to run higher.

A practical corollary is worth stating separately: the aperture is almost always easier to change than either the generator or the selector, which makes gate interventions the highest-leverage move available in many stuck systems. It also makes them the most frequently made by accident. Any change to reporting, measurement, coverage, sampling or exposure is a change to what the system is able to select on, whether or not anyone framed the change that way at the time.

References¶

[1] Lewontin, R. C. "The Units of Selection". Annual Review of Ecology and Systematics, 1970. States the conditions for selection as phenotypic variation, differential fitness and heritability of fitness, so selection acts on expressed phenotypic differences rather than on variants as such. registry ↩

[2] Haldane, J. B. S. "A Mathematical Theory of Natural and Artificial Selection, Part I". Transactions of the Cambridge Philosophical Society, 1924; reprinted in Bulletin of Mathematical Biology, 1990. Shows that selection is ineffective on rare recessive characters because their effects are masked in heterozygotes, so which variants the surviving pool is enriched for depends on which effects are expressed. registry ↩

[3] Rutherford, Suzanne L., and Susan Lindquist. "Hsp90 as a Capacitor for Morphological Evolution". Nature, 1998. Reduces a buffering modifier while leaving the standing variation and the selective regime in place, and finds previously masked phenotypic variation expressed and available to selection. registry ↩a ↩b ↩c

[4] Van Dyken, J. David, and Michael J. Wade. "The Genetic Signature of Conditional Expression". Genetics, 2010. Shows that the strength of selection on a conditionally expressed gene is discounted by the fraction of individuals expressing it, so a rarely expressed gene behaves like a weakly selected one and accumulates deleterious mutations accordingly. registry ↩a ↩b

[5] Hazell, Lorna, and Saad A. W. Shakir. "Under-Reporting of Adverse Drug Reactions: A Systematic Review". Drug Safety, 2006. Finds a median under-reporting rate of 94 percent across 37 studies in 12 countries, an aperture produced by ordinary reporting practice rather than by concealment. registry ↩

[6] Gibson, Greg, and Ian Dworkin. "Uncovering Cryptic Genetic Variation". Nature Reviews Genetics, 2004. Documents large reservoirs of standing variation shielded from selection by recessivity, incomplete penetrance and condition-dependent expression. registry ↩

[7] Chaney, Allison J. B., Brandon M. Stewart, and Barbara E. Engelhardt. "How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility". Proceedings of the 12^th ACM Conference on Recommender Systems, 2018. Simulates the exposure-feedback loop in which what a system shows determines what accrues evidence of being good. registry ↩

[8] Star, Susan Leigh, and Anselm Strauss. "Layers of Silence, Arenas of Voice: The Ecology of Visible and Invisible Work". Computer Supported Cooperative Work, 1999. Analyses how work becomes structurally invisible to the systems that record and reward work, independently of how highly that work is valued. registry ↩

[9] Rubin, Donald B. "Inference and Missing Data". Biometrika, 1976. Separates missingness independent of the underlying value from missingness that depends on it, which is what distinguishes ordinary attrition from a selective gate. registry ↩

[10] Hardt, Moritz, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. "Strategic Classification." ITCS '16 (2016): 111–122. Supports only the strategic-variant half of the contrast — that classified individuals "may manipulate their attributes in order to obtain a better classification outcome" — and neither the paired claim about molecules nor any account of which effects fail to transfer. registry ↩

[11] Paaby, Annalise B., and Matthew V. Rockman. "Cryptic Genetic Variation: Evolution's Hidden Substrate." Nature Reviews Genetics, vol. 15, no. 4 (2014): 247–258. Supports only the genetic half of the reversibility contrast: cryptic genetic variation "is invisible under normal conditions" — unexpressed rather than absent, and so still standing in the population. It carries nothing about irreversible removal in non-genetic substrates such as a moderation queue. registry ↩

[12] Kimura, Motoo. The Neutral Theory of Molecular Evolution. Cambridge University Press, 1983. Establishes the drift barrier: a variant whose realized selection coefficient is small relative to the effective population size behaves as effectively neutral. registry ↩

[13] Inozemtseva, Laura, and Reid Holmes. "Coverage Is Not Strongly Correlated with Test Suite Effectiveness". Proceedings of the 36^th International Conference on Software Engineering, 2014. Finds that a suite's coverage is a poor predictor of its ability to detect faults, so metrics computed on the covered surface do not track quality off it. registry ↩

[14] Campbell, Donald T. "Assessing the Impact of Planned Social Change". Evaluation and Program Planning, 1979. States that the more a quantitative indicator is used for social decision-making the more it is subject to corruption pressures and the more it distorts the process it was meant to monitor. registry ↩