The Cost of Staying

1. Staying Can Change the Next Decision

Staying with something can change what staying is worth next time.

That sentence states the problem more directly than the vocabulary usually used to describe it. People do not always evaluate a fixed activity from a fixed psychological state. Reading an unfamiliar subject, learning statistics, understanding a musical system, acquiring legal doctrine, or developing a technical skill can change the learner who later encounters the same domain.

A beginner often confronts disconnected facts, unexplained terms, uncertain goals, and weak cues about what matters. An experienced learner may encounter the same material as structured relations. A historical dispute that initially looks like a collection of names and dates can later become a conflict among recognizable institutions. A harmonic event that initially sounds arbitrary can later become surprising because the listener has learned what was likely to happen instead. A mathematical procedure that initially requires deliberate attention to every step can later be organized into larger units.

The material may not have become simpler. The learner has changed.

Educational psychology already contains a strong developmental account of this process. Alexander’s Model of Domain Learning describes progression from acclimation through competence toward proficiency as changes in domain knowledge, strategic processing, and interest (Alexander, Jetton, & Kulikowich, 1995; Alexander et al., 1997; Alexander, 2003). The model matters here because these dimensions are not decorative properties of expertise. They can change the cost and expected return of continuing.

A learner with more domain knowledge may understand a difficult paragraph with fewer unresolved dependencies. A learner with better strategies may know when to reread, retrieve, summarize, or seek another explanation. A learner with developed interest may derive value from distinctions that previously appeared inert. A learner with better metacognitive calibration may recognize that apparent difficulty is still producing learning.

This creates a temporal asymmetry.

That asymmetry is especially important during early learning. The learner is asked to decide whether an activity is worth pursuing before possessing all of the knowledge required to experience what later competence can make available.

The present paper asks what happens when that developmental interval occurs inside an environment where competing activities are easy to reach.

The answer is not assumed. Easier exit can be beneficial. It can free people from badly designed material, low-value tasks, inefficient explanations, and cultural objects they simply do not value. A knowledgeable learner may rationally stop earlier than a novice because the knowledgeable learner has already extracted what matters. A student who abandons one poor explanation and finds a better one may improve learning.

The central outcome therefore cannot be “time spent.” It must be something stricter.

Learning-opportunity loss

An exit creates learning-opportunity loss when it reduces engagement with the underlying learning goal while meaningful attainable learning remains.

This definition differs from simple abandonment of a particular source.

A learner who leaves one article and immediately finds a clearer article on the same topic may not have lost developmental exposure. They have changed sources while preserving the learning goal.

A learner who closes the article and spends the rest of the intended learning period on an unrelated activity has.

This distinction produces three behaviorally different exits:

Within-source continuation or adjustment: the learner stays with the same source but changes strategy.

Domain-preserving switch: the learner leaves the current source but continues pursuing substantially the same learning goal elsewhere.

Domain abandonment: the learner reallocates the learning period away from that goal.

The theory is primarily concerned with the second and third categories only insofar as they reduce effective learning. Loyalty to one text is irrelevant. Continued development is not.

That is the paper’s contribution.

It does not discover switching costs. It does not discover opportunity costs. It does not discover metacognitive stopping. It does not discover that knowledge changes learning. It does not discover that digital environments can affect effort allocation.

Those ideas already exist.

The distinctive claim concerns their interaction across time.

2. Why People Leave

Several mature research traditions already explain why a person may stop doing something difficult.

Effort can enter the choice

Kool et al. (2010) found that people often avoided options associated with greater cognitive demand. Westbrook, Kester, and Braver (2013) showed that participants discounted monetary rewards associated with more demanding cognitive tasks. These findings support the proposition that cognitive effort can affect subjective value.

They do not establish that people generally dislike thinking.

Inzlicht, Shenhav, and Olivola (2018) describe an “effort paradox”: effort is often costly, but it can also increase value, become intrinsically meaningful, signal accomplishment, or be actively sought. Escobar and Mitchell’s (2025) systematic review further shows that effort-discounting studies differ substantially in definitions, procedures, and models.

The defensible conclusion is therefore narrow. Effort can matter to choice, but its meaning depends on the person, activity, expected reward, and surrounding context.

Alternatives matter

Kurzban et al. (2013) propose that subjective effort partly reflects opportunity costs: allocating limited cognitive processes to one task prevents their simultaneous use elsewhere. Shenhav, Botvinick, and Cohen’s (2013) expected-value-of-control framework similarly asks whether the expected benefit of deploying cognitive control justifies its cost.

These theories differ, but both make the same point relevant here: a difficult activity cannot always be evaluated without considering what else the person could be doing.

The cost of moving matters

Information-foraging theory comes even closer to the present problem. Pirolli and Card (1999) model information seeking as movement among information patches that differ in expected yield and in the cost of finding or entering another patch.

The theory predicts that people should remain longer in a current source when movement elsewhere is more costly, all else equal.

Liu et al. (2016) tested this logic directly in an interactive reading environment. Increasing the delay required to switch among text sources increased persistence within the current source and improved immediate and delayed recall.

That finding establishes something the present paper does not need to rediscover: environmental switching costs can change reading behavior and learning.

But “switching cost” means two different things

An environmental transition cost is the friction required to reach another activity: delay, navigation, authentication, physical movement, additional clicks, or some other barrier.

A cognitive switch cost is the mental cost of changing task sets and reconstructing context after a switch (Kiesel et al., 2010).

These costs can move in opposite directions.

A smartphone can make the environmental transition from reading to another activity extremely cheap while the human cognitive system still pays a cost when returning to the interrupted material.

The distinction matters because only an anticipated cost can directly influence the initial decision. If returning to a text later turns out to be cognitively expensive but the learner did not expect that cost, it cannot explain why the learner left. It can, however, make the switch more damaging after it occurs.

People also decide whether continued study is still useful

Metcalfe and Kornell’s (2005) Region of Proximal Learning model makes a different contribution. Learners decide not only what to study but how long to continue.

They do not have direct access to their true knowledge state. They use judgments.

A learner may stop because they believe the material is already known. They may continue because they believe they are still making progress. They may stop because the perceived rate of learning has fallen.

This distinction is central.

More knowledge does not necessarily imply more persistence.

A knowledgeable learner may correctly stop early because little additional learning remains. A novice may persist for a long time without learning efficiently. A learner may also stop prematurely because they believe progress has stalled even though another interval would produce substantial durable learning.

The important variable is therefore not simply knowledge. It is what the learner believes another interval of effort will produce.

Difficulty can be useful while feeling unproductive

The desirable-difficulties literature sharpens the problem. Some learning conditions reduce subjective fluency or immediate performance while improving longer-term retention and transfer. De Bruin et al.’s (2023) Start and Stick to Desirable Difficulties framework asks why learners often avoid or abandon such methods.

The answer partly concerns a mismatch between actual learning and experienced learning.

A strategy can work while feeling slow. An easy strategy can feel productive while producing less durable learning. The learner must decide using imperfect evidence.

Failure may matter more than the alternative

Embrey et al. (2025) provide important adverse evidence against a simple opportunity-cost account. Across three experiments, increasing the value of alternatives did not reliably increase effort avoidance. In their third experiment, a more enjoyable alternative did not reliably increase switching away from a demanding N-back task. Recent errors were more predictive.

This matters because two superficially identical exits can have different causes.

A person can leave because elsewhere became more attractive. They can leave because the present task produced repeated evidence of failure. Or both can happen at once.

A serious theory must allow those explanations to compete.

3. Learning Changes the Learner

The theory requires one developmental premise: learning can change the functional problem presented by later learning.

This premise has deep roots.

Chase and Simon’s (1973) chess studies showed that expert players encoded meaningful configurations differently from novices. Gobet and Simon (1998) refined the chunking account while preserving the broader insight that stored domain structure changes how information is represented and retrieved.

Ericsson and Kintsch’s (1995) long-term working-memory theory extends related logic to skilled cognitive performance. The point is not that expertise creates unlimited working memory. Ricks and Wiley (2009) show that knowledge benefits depend on domain relevance. Lai, Huang, and Gershman (2025) similarly model action chunking as use of learned structure to reduce policy complexity under limited resources.

Learning changes organization, not human finitude.

For present purposes, Alexander’s Model of Domain Learning provides the stronger developmental foundation. It explicitly links changes in knowledge, strategies, and interest across increasing domain competence.

Actual knowledge

Let K_t represent what the learner actually knows at time t.

Actual domain knowledge can improve comprehension by supporting inference, helping distinguish central from peripheral information, and allowing new propositions to attach to existing conceptual structures. Smith et al. (2021) review evidence showing substantial effects of background knowledge, although those effects vary with text, misconceptions, and reader ability.

Knowledge can also change strategy.

Moos and Azevedo (2008) found that prior domain knowledge was associated with differences in planning, monitoring, and strategy use during hypermedia learning.

The environment that helps a beginner may therefore become redundant or even counterproductive as knowledge grows. That is the central lesson of the expertise-reversal effect (Kalyuga et al., 2003).

Perceived mastery

Let K̂_t represent what the learner believes they know.

Actual knowledge and perceived knowledge need not match.

A learner can know little and feel confident. Another can know a great deal and remain uncertain.

Because stopping decisions depend on beliefs rather than omniscient access to actual state, this distinction is essential.

Perceived marginal learning

Let ĝ_t represent what the learner expects to gain from another interval of effort.

This is not the same as perceived mastery.

A person can know little while believing that another five minutes will teach them almost nothing. Or know a great deal while recognizing that an unresolved distinction is still worth working through.

Metcalfe and Kornell’s work implies that judgments of learning rate can govern perseverance even when actual learning differs from those judgments.

Interest and value

Let I_t represent interest or subjective domain value.

Interest can precede learning, result from learning, or participate in reciprocal development.

Hidi and Renninger (2006) describe interest developing from situational engagement toward more stable individual interest. Rotgans and Schmidt (2017), testing competing causal accounts, found evidence in their samples that knowledge acquisition preceded later individual interest. Sansone, Renninger, and Harackiewicz (2026) emphasize the broader interaction among interest, utility value, self-regulation, effort, and persistence.

The direction should therefore be measured rather than assumed.

Knowledge does not automatically produce staying

Babineau et al. (2022) provide an important limit. In their geology-learning study, students’ domain knowledge increased, but measured prior knowledge did not predict stay-versus-switch decisions. Retrieval practice changed switching more clearly.

This is exactly the kind of result the present theory must survive.

The claim cannot be: more knowledge → more persistence.

Knowledge may reduce processing cost. It may improve competence signals. It may change perceived marginal learning. It may create interest. It may also allow a knowledgeable learner to recognize that little remains to be gained and leave sooner.

The behavioral effect depends on the state produced by learning and on how accurately that state is understood.

4. A Dynamic Model of Staying and Leaving

The model can be stated with three equations and one diagram.

The equations are not intended as a complete computational theory. They identify the quantities that must be separated experimentally.

Current choice

At time t, let the learner’s subjective value of continuing be:

where expected reward or utility and expected subjective cost can depend on perceived mastery, perceived marginal learning, recent success or failure, interest, goals, and task difficulty.

The best available alternative has subjective value:

T̂ is expected environmental transition cost and Q̂ any anticipated cognitive or resumption cost. Exit becomes more likely as the alternative becomes comparatively more valuable.

This is not a new discovery. It is a compact representation of existing allocation logic.

Learning changes the next state

Actual knowledge changes through engagement:

E_t is the amount of effective engagement, q_t the quality or continuity of that engagement, and δ_t forgetting.

This equation distinguishes two ways exit can matter.

A learner may simply spend less total time learning. Or the same total learning time may be broken into fragments that reduce learning per unit of time.

Those are different mechanisms.

The causal structure

Three rival paths must be tested alongside this one: recent failure → exit; stable preference → both early and later exit; effort recalibration or habit → later exit independent of domain knowledge.

That distinction is the theoretical center of the paper.

Moderator versus mediator

Knowledge can play two different roles.

Existing knowledge can moderate the immediate effect of the environment. A knowledgeable learner may be less affected by easy alternatives because the focal activity is already intelligible and rewarding—or more willing to leave because they can quickly judge that nothing useful remains.

Knowledge acquired during the experiment can also mediate a historical effect. Earlier exit conditions can reduce later knowledge, and the resulting knowledge difference can alter later behavior.

These claims should never be treated as equivalent.

One concerns who is susceptible now. The other concerns how earlier conditions create later differences.

Predictions at a glance

Prediction 1. Environmental transition cost should affect learning-opportunity loss under at least some conditions, but not necessarily all exit.

Prediction 2. Actual knowledge, perceived mastery, perceived marginal learning, and recent failure should explain different portions of current stopping behavior rather than collapsing into one “engagement” variable.

Prediction 3. Different exit environments during acquisition can produce later behavioral differences after those environments are made identical, if acquisition conditions changed relevant learner states.

Prediction 4. If a particular learner state carries the historical effect, experimentally repairing that state should reduce the later difference.

The fourth prediction is the strongest.

It turns the theory from an account of association into a causal rescue problem.

5. Reading as the Main Test Case

Long-form reading provides the strongest empirical domain because each major component can be measured.

Information-foraging research already shows that environmental switching costs can affect reading allocation. Liu et al. (2016) found that increasing the delay associated with switching sources increased within-source persistence and improved recall.

The knowledge literature supplies the developmental edge. Background knowledge can improve comprehension by supporting inference and integration (Smith et al., 2021). Stanovich’s (1986) account of Matthew effects in reading describes reciprocal processes in which reading exposure, knowledge, vocabulary, and reading ability can become mutually reinforcing.

Hwang, McMaster, and Kendeou (2023) provide more recent longitudinal evidence. Earlier science-domain knowledge predicted later reading performance, and earlier reading predicted later domain knowledge, even after accounting for several cognitive, linguistic, literacy, and demographic variables.

Those findings do not establish the present mechanism. They establish that reading and knowledge can develop together.

Why long-form reading is especially useful

Long-form reading contains delayed returns.

An early distinction may matter only later. A definition can remain cognitively expensive until subsequent material reveals why it was necessary. Characters, historical actors, premises, causes, and concepts can acquire meaning across pages rather than sentences.

This makes learning opportunity measurable over time.

Researchers can estimate the incremental learning produced by another standardized reading interval at different points in the text.

But the strongest version should not rely solely on matched participants to estimate another person’s counterfactual gain.

Micro-randomized continuation probes

At selected decision points, participants who would otherwise be free to leave could be randomly assigned either to receive the opportunity to exit immediately or to complete one additional short interval before choice resumes.

Across many participants and many decision points, these randomizations would estimate the causal learning yield of an additional interval at different positions on the learning curve.

The logic is simple: rather than guessing what another five minutes would have taught someone, randomly determine whether some people receive those five minutes.

The average difference in later knowledge provides an estimate of what that interval was worth under those conditions.

This produces a stronger empirical definition of remaining learning opportunity.

Source switching is not domain abandonment

Reading also makes another distinction especially important.

Leaving one source may improve learning if the learner finds a better source.

Experiments should therefore track what happens after exit.

A participant who leaves the focal article might choose another explanation of the same concept, a video teaching the same material, a retrieval exercise on the same domain, or an unrelated activity.

Only the last clearly abandons the learning goal.

The second and third may improve it.

The relevant outcome is therefore not focal-source persistence but domain-level learning opportunity preserved or lost after switching.

That turns the theory into an account of developmental allocation rather than document loyalty.

The paper-versus-screen literature becomes secondary

Delgado et al. (2018), Li and Yan (2024), and Clinton-Lisell and Litzinger (2026) show that differences between paper and digital comprehension depend substantially on moderators such as scrolling and task conditions.

The important lesson is methodological.

“Digital reading” is too coarse a treatment.

Scrolling, alternative availability, notification salience, transition friction, search access, and resumption cues can be manipulated separately.

The primary experiment should therefore use the same text on the same device while changing only the causal variable under study.

6. One Transfer Test: Music

Music supplies one useful test beyond reading.

Its purpose is not to demonstrate that every demanding cultural activity follows the same psychology. It tests whether learning can change subjective value through a different mechanism.

Predictive theories of musical cognition propose that musical events are experienced relative to learned expectations about rhythm, melody, harmony, timing, and form.

Gold et al. (2019) found systematic relations among predictability, uncertainty, and musical pleasure. Cheung et al. (2019) found that uncertainty and surprise jointly predicted pleasure and responses in auditory and limbic-related regions. Vuust et al. (2022) place prediction centrally within a broader account of music perception and reward.

The relevant point is straightforward.

Learning can change what counts as surprising. That can change what is rewarding.

A listener who has learned a musical grammar can experience an event as delayed, violated, resolved, or structurally significant in a way unavailable to a listener who has not learned the relevant probabilities.

Savage and Fujii (2022) provide an important limit: musical prediction is culturally structured. Models derived from Western tonal systems cannot simply be treated as universal accounts of musical expectation.

Music therefore supplies an independent test of the broader principle: engagement can change a learner state that changes what future engagement yields.

That is enough. The paper does not need additional analogies to cooking, poetry, or other cultural domains.

7. Digital Environments Are an Application, Not the Theory

The theory does not require smartphones.

People have always had alternatives. Libraries, newspapers, conversations, television, work, family obligations, and physical environments all create competing demands.

Digital environments matter because they can combine several relevant features at unusually high intensity: many alternatives, very low transition cost, persistent cues, personalization, rapid feedback, and portable access.

A reader can move from a difficult paragraph to communication, entertainment, shopping, news, music, games, or search within seconds.

That does not prove that the person will leave. It changes the environment in which leaving is evaluated.

Wiradhany, Parry, and Aru’s (2026) effort-recalibration framework already proposes that repeated engagement with low-friction digital activities can change how effort is valued and allocated. Kushlev’s (2026) framework similarly describes displacement, substitution, and interference as ways technology can alter well-being and behavior.

The present account differs in the proposed carrier of historical effects.

Effort recalibration predicts that repeated environments change how effort itself is valued.

The present theory predicts that earlier environments may also change domain-specific knowledge and metacognitive state.

These mechanisms can coexist. The experiments should force them apart.

If later persistence differences remain after actual knowledge and metacognitive calibration are experimentally repaired, effort recalibration, habit, or another persistent process becomes more plausible.

If the difference disappears after repairing domain knowledge, the developmental mechanism gains support.

Acute evidence should remain in its proper place. Chen et al. (2025) found a medium negative effect of mobile-phone distraction on immediate recall across randomized experiments. Fournier et al. (2026) found transient slowing after social-media-style notifications. Tam and Inzlicht (2024) found that boredom can increase switching while switching can itself increase boredom.

These studies show that interruption and switching matter.

They do not establish that digital environments permanently reduce cognitive capacity.

The theory does not require that claim.

8. What Else Could Explain the Effect?

A theory becomes useful when neighboring explanations are allowed to win.

Rational exit

Leaving can be optimal.

Information foraging predicts that a person should leave when another source offers greater expected return.

A knowledgeable learner may stop sooner because they know the material.

A person may leave one source because a better explanation exists.

No experiment should count those cases as developmental failure.

Failure or frustration

Embrey et al. (2025) show that recent errors can predict leaving more strongly than manipulated alternative value.

If failure signals explain exit and environmental transition cost contributes little after those signals are manipulated independently, the present theory should concede explanatory territory.

Stable preference

People differ before the experiment begins.

Need for Cognition, reading ability, conscientiousness, interest, education, available leisure, and socioeconomic resources can all predict both exposure and later competence.

Randomization is therefore essential.

Interest selection

People often become knowledgeable about things they already like.

Interest can cause exposure. Knowledge can also create interest. Reciprocal processes are possible.

Repeated measurement is necessary to determine direction.

Strategy change

Leaving the focal source can be intelligent.

A learner may replace rereading with retrieval practice, switch to an instructor, use a diagram, consult a glossary, or ask an AI system for another explanation.

The model should therefore treat stopping as one action inside a broader learning policy.

Cultural status

Expertise can be real while prestige remains socially structured.

Bourdieu (1984) makes it impossible to treat cultivated taste as transparent evidence of intrinsic superiority. Thomas (2022) shows that highbrow cultural preferences can affect perceptions of class and competence. Breinholt and Jæger (2020), however, find little support in their educational data for a simple mechanism in which cultural capital raises performance because teachers mistake it for ability.

The theory therefore has no preferred cultural hierarchy.

A programming language, bluegrass idiom, video game, mathematical system, literary tradition, or professional field can instantiate the mechanism.

The criterion is structured learning, not prestige.

9. Experiments That Could Prove the Theory Wrong

The empirical program should proceed in four stages.

Each stage should eliminate an ambiguity left by the previous one.

Study 1: What causes current exit?

The first study should independently manipulate the major current-state causes.

Participants undertake the same difficult learning task on the same device.

The experiment manipulates alternative reward, environmental transition cost, focal-task failure, and expected resumption cost.

Alternative activities are individually rated as high or low value. Alternatives are either immediately accessible or require a modest delay or additional steps. Task difficulty or feedback is manipulated to generate more or fewer genuine errors while preserving the broader learning target. Some participants receive accurate information about how disruptive returning after a switch is likely to be; others do not.

This creates a cleaner causal design than simply measuring many correlates at once.

Exit must be real. Participants cannot be required to repay time later.

After leaving, their next action is classified as domain-preserving or domain-abandoning.

The primary outcome is not simply time to first exit. It is learning-opportunity loss following exit.

Micro-randomized continuation probes estimate whether meaningful learning remained at the stopping point.

The study can then ask: Did low transition cost increase exit? Did it increase learning-truncating exit? Did high alternative reward matter independently? Did recent failure matter more? Did people underestimate resumption costs?

Those answers determine which mechanism deserves weight.

Study 2: Separate what people know from what they think they know

The second study independently manipulates actual knowledge and metacognitive information.

Participants first receive either substantive domain instruction or a matched control. This changes actual knowledge.

Separately, participants receive calibrated or experimentally distorted information about their current mastery and expected learning rate. This changes perceived mastery and perceived marginal learning without necessarily changing underlying content knowledge to the same degree.

The design can therefore create theoretically revealing states: high actual knowledge with low perceived mastery; low actual knowledge with inflated mastery; high actual knowledge with accurate calibration; low actual knowledge with accurate recognition of substantial remaining learning.

Participants then make continuation decisions under different transition-cost conditions.

This directly tests whether current stopping is driven more strongly by actual knowledge, subjective mastery, expected marginal learning, errors, or outside-option value.

A sophisticated theory should allow greater knowledge to produce less persistence when little useful learning remains.

That is not failure. It is calibrated stopping.

Study 3: Does history remain after the environment is gone?

This is the central experiment.

Participants learn an unfamiliar structured domain across repeated sessions.

During acquisition, the environment varies in transition cost and access to alternatives.

One group can leave for valued alternatives with almost no friction. Another has access to the same alternatives with modest friction.

The first question is whether this changes total effective learning.

But a second mechanism must also be separated.

Did people learn less because they spent less total time? Or because the same amount of time was broken into pieces?

A yoked design answers this.

For every free-choice participant, matched participants receive equivalent total focal-learning time. One receives that time in relatively continuous blocks. Another receives the same amount in a fragmented pattern matched to the switching history.

This distinguishes less exposure from equal exposure with greater fragmentation.

At the end of acquisition, all participants enter the same environment.

The original transition-cost difference disappears.

They encounter new material requiring what they previously learned.

If the groups now behave identically, the effect was current and environmental.

If differences remain, history has entered the learner.

The study then measures which learner state differs: actual knowledge, perceived mastery, perceived marginal learning, interest, strategy, or something else.

That is the result on which the theoretical paper ultimately stands.

Study 4: Try to rescue the effect

The final study tests causation by repairing the proposed mediators.

Participants first acquire different learning histories through the Study 3 manipulation.

They then enter a randomized 2 × 2 intervention.

One factor is knowledge restoration. Participants receive targeted instruction designed to close the experimentally produced gap in actual domain knowledge.

The other factor is metacognitive calibration. Participants receive structured evidence comparing predicted and actual performance and accurate information about the learning yield of another study interval.

The goal is not simply to help participants. It is to repair different proposed causal carriers.

Then everyone faces another identical continuation problem.

If restoring actual knowledge reduces the historical difference while calibration does little, knowledge is a major carrier.

If calibration reduces the difference while knowledge restoration does not, subjective state matters more.

If both matter, they transmit independent parts of the effect.

If restoring knowledge improves performance but the behavioral difference persists until metacognition is calibrated, what people believe about learning mediates the consequences of what they know.

If neither intervention rescues the difference, the present model has identified the wrong carrier.

Effort recalibration, habit formation, interest, avoidance learning, or another persistent mechanism then deserves greater weight.

That is not a secondary result. It is a failure of the theory’s central causal claim.

Methodological commitments

The studies should preregister minimum effects of theoretical interest rather than allowing any statistically significant effect to count as support.

Equivalence testing or Bayesian model comparison should be used where appropriate to determine whether theoretically important effects are absent or negligible.

If transition friction changes switching but produces trivial differences in later knowledge or later common-environment behavior, the developmental claim has failed even if the immediate behavioral effect is statistically detectable.

Knowledge measurement also requires care. Repeated quizzes can themselves teach through retrieval practice.

The experiments should therefore use parallel forms, sparse randomized probes, or latent-variable approaches such as item-response models to estimate knowledge while minimizing measurement-induced learning.

The same problem applies to subjective measurement.

Repeatedly asking participants whether they are interested, confident, or still learning can change the behavior being studied.

Metacognitive and interest probes should therefore be sampled intermittently or randomized so measurement reactivity can itself be estimated.

The studies should use hierarchical models for repeated observations and survival or hazard models for exit timing.

But the statistical sophistication is subordinate to the causal logic.

If those steps fail, the theory should fail with them.

10. Conclusion

People do not decide whether to continue difficult activities from outside their own developmental histories.

They decide using whatever knowledge they possess, whatever progress they think they are making, whatever competence or failure the task is currently signaling, whatever interest has developed, and whatever alternatives the environment makes available.

Existing research explains many pieces of that decision.

Effort can influence value.

Information-foraging theory explains leaving declining information patches.

Metacognitive research explains stopping when perceived learning has stalled.

Domain-learning research explains changes in knowledge, strategy, and interest.

Digital-cognition research explains why contemporary environments may make some alternatives unusually immediate.

The unresolved question is what happens when those processes operate across time.

By changing exposure, fragmentation, domain knowledge, perceived mastery, expected marginal learning, or interest, an earlier choice environment may alter later stopping even after that environment is gone.

The theory makes no virtue of staying for its own sake.

The rational learner should leave when useful learning is exhausted or when another route offers a better return.

The important case is narrower: a person leaves while substantial learning remains, and repeated exits alter the learner state that will govern the next decision.

This is why the strongest empirical test is historical.

Change the environment during acquisition.

Then remove the difference.

If behavior remains different, identify what changed inside the learner.

Restore knowledge.

Calibrate beliefs about learning.

If the behavioral difference contracts, the proposed developmental pathway becomes credible.

If it does not, a better theory should replace it.

The cost of staying is real. So is the cost of leaving. The scientific problem is to determine when either cost is being judged from a state that continued learning itself would have changed.