From A to B

A ripe cacao pod contains sweet pulp around its seeds. The pulp offers food before anyone processes the seeds. Chocolate requires much more work. The familiar process includes fermentation and drying, followed by roasting and grinding. Looking backward from a chocolate bar, the sequence seems to demand someone who could foresee the result before any of the steps made sense.

That view gives the finished product too much authority over its history. Earlier users could have wanted something else. The pulp was food. A fermented drink could be desirable before anyone used the seeds to make a different drink. A useful intermediate product can pay for the experimentation that eventually replaces it. The archaeological case for such a path is suggestive, although it cannot recover a first inventor’s thoughts. Henderson et al., 2007.

The larger problem begins after a procedure works. People can preserve a route from A to B while losing much of the explanation between them. This is how knowledge becomes available to people who could never have discovered it alone. It also leaves them uncertain about what can safely change. A copied recipe may survive for centuries, then fail after an apparently harmless substitution.

The danger is greatest when a society retains the instruction but loses the ability to check whether its conditions still hold. Preventing that loss requires continued work. It does not require each generation to suffer the original disaster.

How could anyone discover chocolate without knowing the destination?

Cacao was used at Santa Ana-La Florida, in present-day Ecuador, about 5,300 years ago. Researchers identified starch grains, chemical residues, and ancient DNA. The finds establish early use of the plant, not the full process a modern chocolate maker follows. Zarrillo et al., 2018.

At Puerto Escondido, in present-day Honduras, residues in vessels support cacao beverages before 1000 BCE. Henderson and colleagues proposed that the earliest drinks there were fermented from the pulp. The chemical evidence cannot distinguish a pulp drink from a seed drink; vessel shapes contribute to the interpretation. The proposal supplies a plausible intermediate reward.

Fermentation itself does not wait for someone to understand it. In modern cocoa processing, yeasts consume pulp sugars and produce ethanol. Bacterial activity follows and overlaps with this work. Acetic acid bacteria oxidize ethanol, generating acid and heat. Conditions inside the seeds change, helping create flavor precursors used during later processing. The process depends on the organisms and their surroundings, including oxygen availability and how the mass is handled. De Vuyst and Leroy, 2020.

An accidental fermentation is therefore biologically plausible, though no specific ancient accident is documented. Microbes that transform an unattended material explain an opportunity for discovery, not the moment someone noticed it.

The difficult human achievement was making a desirable transformation recur. Farm practices still influence fermentation quality, and researchers develop starter cultures to make outcomes more consistent. Naming the microbes does not make the practical problem disappear. De Vuyst and Weckx, 2016.

This changes the invention question. Instead of asking how someone guessed the entire recipe, ask what made the next experiment worth attempting. A preparation method may begin with a benefit that is later secondary. Existing techniques can also be combined: the inventor of one stage need not invent every tool used by the next. Cultural evolution includes these changes of function and combinations of existing practices. Mesoudi and Thornton, 2018.

The finished sequence can consequently be harder to invent from scratch than any decision made during its development. Its complexity records accumulated opportunities. It does not prove that someone once understood the whole route.

What did early humans need to know about meat, marrow, and fire?

Eating an animal does not require first discovering the concept of protein. Cracking a bone does not require a theory of energy density. A behavior can be learned through its immediate consequences, or copied from someone already doing it. Asking how early humans knew marrow was nutritious can mistakenly require them to possess the explanation used today.

The archaeological evidence concerns behavior. At Kanjera South in Kenya, remains dating to about two million years ago show repeated processing of animal carcasses. Cut marks and percussion damage support access to flesh and within-bone resources. The assemblage contains evidence of different acquisition opportunities, rather than a single universal hunting or scavenging strategy. Ferraro et al., 2013.

A fatty edible substance provides a possible reward without any explicit nutritional judgment. Immediate preference and nutritional understanding are different things.

Fire poses several separate discovery problems. Encountering a natural fire differs from keeping one alive. Maintaining one differs from producing ignition. Cooking adds further control over where food sits and how long it stays there. These distinctions allow incremental paths without proving which path actually occurred.

Burned bone and plant ash at Wonderwerk Cave support burning inside the cave around one million years ago. That is evidence of a close association between hominins and fire, not a dated record of the first fire-making technique. A review of European sites found widespread evidence for habitual fire use much later, around 300,000 to 400,000 years ago. These findings concern different places and different strengths of claim. Berna et al., 2012; Roebroeks and Villa, 2011.

Richard Wrangham’s cooking hypothesis proposes that cooking helped support major changes in human biology. A mouse experiment found increased energy gain from cooked food under its conditions. Dating habitual cooking and estimating its contribution to human evolution require other evidence. Carmody, Weintraub, and Wrangham, 2011.

Cooking also changes more than digestibility. A later mouse experiment reported no energetic advantage from cooked meat under its conditions and discussed fat lost during heating. The amount of usable food remaining after preparation matters alongside the ease of digestion. These experiments support checking the particular food and method, rather than assigning cooking one universal calorie multiplier. “Human Brain Expansion during Evolution Is Independent of Fire Control and Cooking,” 2016.

Different questions therefore need different evidence. A damaged bone can establish processing. A feeding experiment can test an energetic benefit. Neither recovers the discovery episode. Combining them helps explain how a practice could be useful before its users had a theory, while leaving the original sequence open.

Can a technique improve while understanding stays still?

In an experiment by Maxime Derex and colleagues, participants adjusted weights on a wheel to make it travel faster down a track. They worked in chains of five. Each successor inherited the previous participant’s final configurations and results. Average final-trial speed rose from 123.6 to 145.7 metres per hour. Scores on a separate understanding test scarcely changed. Derex et al., 2019.

This was improvement without a matching improvement in the measured causal model. Participants were not devoid of useful intuition. They explored some possibilities more often than others. But the experiment demonstrates that a succession of people can preserve successful adjustments without converging on a complete explanation of why the device works.

A second condition let people transmit theories too. Partial theories directed later exploration toward the variables they mentioned. An inherited explanation could therefore make other possibilities harder to notice. This was a constrained laboratory task, but it identifies a risk in teaching: an incomplete reason may be treated as a complete account.

Historical technology also puts practical achievement ahead of mature theory. Newcomen’s atmospheric engine appeared in 1712. Watt patented a separate condenser in 1769. Carnot’s general analysis of heat engines followed in 1824. The engineers plainly had causal ideas. What they lacked was the later theoretical system into which their machines would fit. “Some Early Heat Engine Concepts and the Conservation of Heat,” 1974.

Medicine offers a similar ordering. Aspirin was used long before John Vane’s 1971 work identified inhibition of prostaglandin synthesis as a mechanism.Prostaglandins are chemical signals involved in inflammation. Aspirin’s history connects to salicylates in willow, although willow bark and pharmaceutical aspirin are not interchangeable substances. Variolation, deliberate exposure to material from smallpox lesions, likewise preceded modern germ theory and immunology. It could confer protection while also causing disease and transmitting infection. Effective practice and complete safety did not arrive together. Vane’s Nobel lecture; Mahdi et al., 2006; Riedel, 2005.

A procedure can have evidence behind it without having a complete mechanism. An explanation can also sound coherent without having good evidence behind it. These are separate evaluations. Refusing every unexplained result would discard useful knowledge; accepting every plausible rationale would preserve mistakes.

What must travel with a recipe?

A crop is easier to transport than the whole system that makes it a reliable food. Seeds can cross a border while preparation, complementary foods, and skilled judgment remain behind.

Nixtamalization treats maize with an alkaline material, commonly lime (calcium hydroxide). Among its effects is increased availability of niacin, vitamin B3. A heavily maize-dependent diet can otherwise provide too little available niacin and too little of its precursor, tryptophan. Historical work on maize processing connects the technique to the nutritional viability of maize as a staple. Pellagra’s history also involves poverty and restricted diets; it cannot be reduced to a single missing culinary step. Katz, Hediger, and Valleroy, 1974; “An Outbreak of Pellagra in the Kasese Catchment Area,” 2017.

Cassava illustrates why a step can matter even when its benefit is hard to see. In a 1992 study, a population affected by konzo, a disease that causes permanent spastic paralysis, consumed cassava processed with shorter soaking than an unaffected comparison population. Measurements supported much greater cyanide exposure in the affected population. The researchers identified food shortage and poverty as underlying causes. Subsequent research continues to connect konzo with reliance on insufficiently processed cassava, while leaving aspects of its detailed pathogenesis unresolved. Tylleskär et al., 1992; Kashala-Abotnes et al., 2019.

It would be misleading to turn this into a simple story of descendants forgetting ancestral wisdom. A person can know that preparation matters and lack the time or resources to complete it. Hunger can make a dangerous shortcut immediately attractive. The distinction changes the remedy: an explanation alone cannot supply water or replace a missing meal.

A recipe therefore depends on more than its written steps. Someone must be able to recognize the starting material and judge the result. The necessary inputs must be available. Conditions that were constant in one place may vary in another. In fermentation, even the organisms doing the work belong to the practical arrangement that must recur.

A counterfactual makes the resource problem clearer. If two households receive identical instructions but only one has enough water and food to wait through preparation, equal information will not produce equal choices. Calling the resulting difference ignorance would misdiagnose the cause. The household must be able to carry out the procedure under its constraints.

Transferring a method means enabling another person to produce its result under their conditions. Handing over instructions is one part of that job. A technically accurate document can still fail if it assumes resources or judgments its reader does not possess.

How can the instruction survive while the knowledge fails?

Scurvy shows how an apparently faithful continuation can conceal a material change. Citrus had practical value against the disease before vitamin C supplied the explanation. Yet the later substitution of lime juice for lemon juice around 1860 was followed by renewed scurvy, particularly on polar expeditions. The category “citrus juice” preserved a verbal resemblance while losing a dependable equivalence in protection against scurvy. Baron, 2009.

Preparation and storage mattered too.Heating juice to concentrate or preserve it could damage its vitamin content. Long storage further complicated what a nominal daily ration delivered. Historical study of Arctic crews documents the difficulty of maintaining effective protection under these conditions. Carrying something called lemon or lime juice did not establish that the necessary active substance remained in sufficient quantity. “The health of nine Royal Naval Arctic crews, 1848 to 1854,” 2016.

The error was possible because the practical category was broader than the causal requirement. Sourness and botanical resemblance do not measure vitamin C. Once the relevant mechanism is known, substitutions can be evaluated against it. Before then, maintaining the observed effect requires closer attention to the original material and its handling.

This is a different kind of loss from forgetting the recipe. Several failures can hide behind the same phrase, “lost knowledge.”

Even successful production does not make knowledge permanent. Argote, Beckman, and Epple found that industrial learning depreciated, and that cumulative production overstated how much learning persisted. A record of past output was an unreliable measure of current capability. Argote et al., 1990.

Preservation has to include periodic reproduction of the result. An archive can retain instructions after the ability to execute them has disappeared. A continuing practice supplies evidence that a document alone cannot.

Why do useless steps survive too?

If a learner cannot distinguish necessary actions from irrelevant ones, copying the whole sequence can be reasonable. Removing an unexplained step risks removing the protection that makes the process work. This is a plausible benefit of conservative imitation. The same policy preserves excess detail.

Horner and Whiten demonstrated the problem with a puzzle box. Children copied causally irrelevant actions even when a transparent version made the mechanism visible. Chimpanzees omitted more of those actions when the box was transparent. The test compared imitation strategies in this particular task. Horner and Whiten, 2005.

Lyons, Young, and Keil found that children could continue such over-imitation despite efforts to warn them against unnecessary actions. Copying can be unusually persistent. That persistence offers no guarantee that every copied step is useful. Lyons et al., 2007.

The tempting generalization is that any long-lived custom must contain a hidden survival advantage. The spice hypothesis shows why that inference needs testing. One proposed explanation for spicier cuisines in hotter places is that antimicrobial spices protected food where spoilage was more dangerous. It is an attractive connection between culinary preference and an invisible biological benefit.

Bromham and colleagues tested the account with 33,750 recipes from 70 cuisines containing 93 spices. Once shared history, geographic proximity, and covarying conditions were considered, the results did not support the simple infection-mitigation explanation. This does not establish that spices have no antimicrobial properties. It challenges the proposed explanation of their geographic use. Bromham et al., 2021.

A tradition’s persistence is evidence that people transmitted it. Why they transmitted it remains a further question. A practice may have material benefits, social value, or simply low enough costs to escape correction. Several can coexist. Successful transmission alone cannot tell which explanation applies.

There is consequently no general permission either to delete every unexplained step or to defend every old one. The useful question is narrower: what changes when this particular step is altered, and what evidence would reveal the difference?

Does experience teach what explanation cannot?

An instruction can transfer a conclusion without transferring the ability to recognize when it applies. Knowing the words for a texture differs from learning to feel it. A warning can also be understood without acquiring the urgency it has for someone who experienced the danger. These limits affect what another person can learn from an account. Personal experience has limits of its own.

Hertwig and colleagues compared decisions based on described probabilities with decisions based on sampled outcomes. Participants could behave as though rare events mattered less when learning from experience. Small samples often contain no rare event to remember.Later analysis emphasized how much of the apparent psychological difference can come from the samples people actually saw. Hertwig et al., 2004; Fox and Hadar, 2006.

A person who has done something many times without disaster may therefore have strong conviction supported by weak evidence about rare danger. Experience can supply detail and still misrepresent frequency. A larger record may deserve more weight than a vivid personal history.

Nor does failure automatically teach its lesson. In five studies involving 1,674 participants, Eskreis-Winkler and Fishbach reported poorer learning after personal failure feedback than after success, while people learned from others’ failures. A later methodological critique showed that response consistency could produce apparent learning differences in one of the tasks. The broad claim that failure reliably makes people disengage is therefore less secure than the original interpretation suggested. Eskreis-Winkler and Fishbach, 2019; 2025 commentary.

A separate obstacle arises before a successor gets any chance to learn. Experiments on “hidden failures” found that people undershared informative failures. The inherited record can become biased because unsuccessful attempts are missing. This differs from an inability to grasp another person’s experience. The information was never passed on. Eskreis-Winkler and Fishbach, 2020.

The practical response is to transfer more than the conclusion where the missing context matters. A worked case can show the mistaken choice and its consequence. Supervised practice can expose a learner’s misread cues. A simulation can make a decision consequential without reproducing the original harm. Their effectiveness needs testing in the domain where they will be used.

The assessment can then ask whether the learner recognizes the relevant danger and chooses an appropriate response. Feeling the same fear as a predecessor is a different objective. It may be impossible to reproduce, and may add nothing to the decision being taught.

Why can an explanation feel complete before it is understood?

A named concept can bring scattered observations into a recognizable pattern. That is useful. It can also create more confidence than the explanation warrants. The feeling that several things now fit together does not show that the proposed mechanism can predict what happens next.

Rozenblit and Keil asked people to assess their understanding of familiar mechanisms and then explain them in detail. The attempt exposed gaps, and self-ratings fell. People could recognize an object and know what it did while overestimating their understanding of how it worked. Rozenblit and Keil, 2002.

Sloman and Rabb found a related effect involving other people’s expertise. Participants rated their own understanding more highly when told that scientists understood a phenomenon, even though they had received no explanatory information. Access to a knowledgeable community can be confused with knowledge already inside one’s own head. Sloman and Rabb, 2016.

This helps distinguish two defensible positions. Someone may trust a result because a reliable specialist has checked it. Someone may personally understand the mechanism well enough to evaluate a change. Both are useful, but they license different actions. Trusting an aircraft engineer does not qualify a passenger to redesign a wing.

An explanation earns practical value when it supports discrimination. It should help predict the effect of changing a material, or identify an observation that would contradict the account. A label may organize several examples without making either test possible. Asking for those predictions exposes the missing understanding more directly than asking whether the explanation feels convincing.

How can a society function on borrowed wisdom?

A society does not need every person to rediscover its knowledge. The ability to use discoveries made by others is a source of collective competence. Henrich’s account of cultural evolution places this accumulation at the center of human achievement: individual ingenuity operates within an inheritance that no individual assembled. Henrich, 2015.

The difficult question is how that inheritance stays connected to current conditions. Alan Rogers developed a model in which social learning could spread without increasing the population’s mean fitness, its average success in survival and reproduction. When copying is rare, copyists benefit from others’ costly information gathering. As copying becomes common, reliable new information becomes scarcer. Under the model’s assumptions, the advantage disappears at equilibrium. Rogers, 1988.

Rogers’ result isolates a maintenance problem: copying a correct answer does not produce new evidence that the answer remains correct. The environment may change during the chain of transmission. A learner can then copy faithfully and inherit an answer to an older problem.

Enquist and colleagues modeled one resolution. A learner first copies, then tries independent learning if the acquired behavior proves unsatisfactory. This “critical social learning” can outperform pure copying and, under many conditions, pure individual learning. Its success depends on costs, transmission quality, and how the environment changes. Enquist, Eriksson, and Ghirlanda, 2007.

Laland’s review makes the same problem concrete through selective strategies: copy when uncertain or when discovering an answer independently is costly, and attend to whom the information comes from. These rules let learning combine information from others with direct checking. Laland, 2004.

For institutions, this suggests a division of checking labor. Most people can follow a procedure while some people maintain the competence to test it and revise it. A useful system makes that competence reachable. It records which results justify the rule and who can investigate an exception. The appropriate organization depends on what must be checked and how expensive errors are.

The independence of checks matters too. Ten people repeating the same successful claim may all depend on one trial. Agreement then adds less evidence than ten independently observed results. This is a logical property of the information chain: counting endorsements does not count independent tests. A record of where a claim came from helps distinguish widespread knowledge from widespread repetition.

It is tempting to call this inheritance unearned wisdom. The label obscures the arrangement: the recipient did not pay the discovery cost, but costs still exist in teaching, verification, and maintenance. Refusing the inheritance would force needless repetition. Treating it as permanently self-validating would let its reliability decay.

Can successful protection make its own lesson disappear?

A levee changes what residents experience. Frequent smaller floods stop reaching protected land. Development can expand behind the barrier, while the consequences of a rare failure grow. Di Baldassarre and colleagues modeled these interactions between protection, settlement, and social memory. The model describes a feedback mechanism, not an inevitable outcome of every flood defense. Di Baldassarre et al., 2013.

Residents observe the outcomes after protection has changed them. Few floods may mean that protection is working. The same observation can be interpreted as evidence that the danger was exaggerated. To distinguish these explanations requires attention to the protection itself, not just to the outcomes observed behind it.

Insurance behavior supplies empirical evidence of changing attention to risk. Gallagher found that flood insurance uptake rose after a flood and then declined toward its previous level. Nonflooded communities in the same television market also responded, at about one-third the rate of flooded communities. Direct experience was influential, but information about others’ experience mattered too. Gallagher, 2014.

Longer records do not automatically remain prominent. Candia and colleagues studied attention to cultural products through citations and other measures. Their two-stage model distinguished faster communicative decay from slower persistence through recorded culture. Those measurements do not provide an expiry date for disaster wisdom. They show why storing something and keeping it active in attention are different achievements. Candia et al., 2019.

Financial experience provides another scale. Malmendier and Nagel found that people who had lived through lower market returns reported less willingness to take financial risks and invested differently. The history experienced during a lifetime influenced judgment even though older market history was available. Malmendier and Nagel, 2011.

The popular claim that hard times create strong people, who create good times, which create weak people, compresses these problems into a cycle of character. The evidence does not establish that cycle. Research associating moderate adversity with later resilience concerns individual outcomes, and also distinguishes moderate exposure from severe adversity. It cannot validate a law about the rise and decline of societies. Seery, Holman, and Silver, 2010.

Systems that suppress bad outcomes also suppress opportunities to observe them. Their users need evidence about what the protection is doing, including conditions under which it would fail. Otherwise successful prevention can gradually weaken its own justification.

Is machine learning another recipe without reasons?

A learned model can map inputs to useful outputs without providing a human-scale explanation of the relationship. Its computation is specified. The difficulty is understanding which patterns it uses and whether they will hold in a new setting. This resembles inherited practical knowledge at one precise point: observed success can travel farther than understanding of its conditions.

Geirhos and colleagues call one version of this problem shortcut learning. A model learns a decision rule that works well on standard data but fails when tested under different conditions. High performance within the original setting does not establish that the model learned the relationship its users intended. Geirhos et al., 2020.

Caruana and colleagues describe a pneumonia prediction problem with a revealing pattern. Patients with asthma appeared at lower risk in historical data because they had received aggressive care. A rule-learning system exposed that association. If used to reduce hospital admissions, such a rule could withdraw the care responsible for the favorable outcome. The investigators identified the problem before using the proposed opaque model for that purpose. Caruana et al., 2015.

The pneumonia example has the same causal structure as the flood-protection problem. An intervention helps produce safe outcomes. A learner sees the outcomes and mistakes them for properties of the unprotected situation. Removing the intervention then changes the process that generated the evidence. This is a causal error, whether made by a person or a statistical system.

Rudin argues that high-stakes decisions should use inherently interpretable models where suitable, rather than relying on explanations fitted afterward to opaque ones. An explanation that approximates a model is not necessarily a faithful account of its decisions. Her argument also challenges the assumption that opacity always buys necessary accuracy. Rudin, 2019.

Transparency alone would not solve the pneumonia problem. A perfectly readable rule can still confuse treatment with underlying risk. Interpretation makes the mistake easier to inspect; causal evaluation determines why it is a mistake. The model’s apparent knowledge remains dependent on the circumstances that produced its data.

What should survive the handover?

A reliable handover preserves the parts needed to recognize failure as well as to reproduce success. Some of those parts fit in a document. Others require practice or continued access to someone who can make the relevant judgment.

The cost of checking also needs a place in the arrangement. If ordinary users must sacrifice scarce time to investigate every exception, the written duty may accomplish little. Maintenance needs resources just as food preparation does. Assigning responsibility without the means to perform it repeats the cassava problem at an institutional scale.

The instruction should specify what counts as success. When the consequence is delayed, an earlier measurable indicator may be needed. A food process that looks normal can still differ in a chemically important way. A safety system that produces few incidents can still be accumulating exposure. The endpoint must match the purpose of the rule closely enough to reveal deterioration.

The explanation should distinguish established mechanism from useful conjecture. A conjecture can guide a test without becoming a prohibition on other possibilities. Where the mechanism remains unknown, the original materials and conditions deserve more careful preservation. Substitution is an experiment even when the replacement has a familiar name.

The record should also include failed alternatives. Without them, successors may repeat a discarded approach because they see only the successful result and cannot tell which details earned their place. A short account of a failure can be more valuable than another polished description of normal operation. Keeping such records usable requires a way to find the relevant case when a decision arises.

There is also a practical test for whether the inheritance remains alive: can someone reproduce the result, diagnose a deliberate variation, and locate help when the explanation runs out? Their purpose is to expose a gap while it is still possible to repair it cheaply.

A successor should therefore inherit permission and means to investigate a failing rule. Otherwise a discrepancy becomes a choice between obedience and unsupported improvisation. Records of earlier tests, access to skilled help, and a way to try a change safely give that successor a better option: establish which condition changed before revising the procedure.

Appendix

Method

Sources were checked in September 2026 against the papers, publisher pages, university copies or reference works linked in the text; numbers are quoted from them. Some questions stay open: the exact moments when fermentation, cooking or marrow extraction were first discovered, which no evidence can recover; the full role of steamships in the return of scurvy; and why Tasmania lost technologies, which is disputed and left out. The popular cycle in which hard times create strong people is not supported as a historical law.

Related

• Red Trees shares history of science, medicine
• Business Shape Morphology shares evolution
• Functional Anatomy of Systems shares evolution
• Stories and the AT Field shares memory