Unit 27 · Capstone: Argue From Evidence

A fish kill at a park pond gets people talking. One neighbor says a factory dumped poison; another says the heat wave did it. Each of these is a claim: a statement that answers a question about what happened or why. A claim is not right or wrong just because someone says it with confidence. It becomes worth believing only when it is tied to evidence, meaning observations and measurements that anyone could check.

Evidence by itself does not settle anything. A dissolved-oxygen reading of 1.5 milligrams per liter is just a number until someone explains why it matters. That link is the reasoning: the scientific principle that connects the evidence to the claim. Warm water holds less oxygen, decaying algae use up what is left, and most fish suffocate below about 2 milligrams per liter. Now the number supports the heat-wave claim, and the poison claim needs its own evidence.

Good scientific claims share one more feature. They are falsifiable: there is some observation that would show them to be wrong. ’The fish died because of low oxygen’ can be checked against oxygen readings, fish species and the timing of the deaths. ’The fish died because the pond was unlucky’ cannot be checked at all. A claim that no possible evidence could ever count against is not a scientific claim, however attractive it sounds.

Throughout this unit you will use this claim, evidence and reasoning pattern to take arguments apart and to build your own. The questions to ask are always the same. What exactly is being claimed? What evidence is offered, and could I look at it myself? What reasoning links the two, and does that reasoning hold up? And what would it take to show the claim is wrong?

In most American cities, ice cream sales and drowning deaths rise and fall together across the year. Nobody believes that ice cream causes drowning. Both rise in summer, when people swim more and eat more cold treats. Summer is a confounding variable: a hidden factor that drives both things and creates a correlation between them. A correlation means two quantities move together. Causation means one actually produces the other.

There are three main ways a correlation can fool you. First, a confounder like the season may cause both. Second, the cause may run backward: hospitals with more patients in intensive care have more deaths, but the care is not killing them; the sickest patients are sent there. Third, with enough pairs of numbers, some will line up by pure chance. If you compare a hundred unrelated things, a few will match closely by accident, and a careless researcher will report those few.

So how did scientists ever conclude that smoking causes cancer, when no one can run an experiment forcing people to smoke? In 1965 the British statistician Austin Bradford Hill laid out a set of questions. Is the association strong? Does more of the cause bring more of the effect? Does the cause come before the effect in time? Does the link show up in many different studies and populations? Is there a known mechanism? No single answer proves causation, but together they build a case that a chance correlation cannot match.

The strongest tool is still the controlled experiment, where you change one factor on purpose and hold the rest steady. When that is impossible or unethical, scientists lean on the Bradford Hill questions, on natural experiments, and on looking hard for confounders. When you read that ’a study links’ one thing to another, your first question should be: what else could explain this?

In 1936 the magazine Literary Digest mailed about ten million ballots asking Americans how they would vote for president. More than two million came back, an enormous sample, and the magazine confidently predicted that Alf Landon would beat Franklin Roosevelt. Roosevelt won in one of the biggest landslides in history. The mailing list had been built from telephone directories and car registrations, and in the middle of the Depression the people who owned phones and cars were richer than average and leaned against Roosevelt. A young pollster named George Gallup, using a far smaller but carefully chosen sample, got it right.

The lesson is that size does not cure bias. Bias is any systematic way that a sample differs from the population it is supposed to represent. Surveying only people who answer the phone, only patients who came back to the clinic, or only the mice that survived the first week all build bias into the result. The cure is random selection, where every member of the population has an equal chance of being picked, so that the sample’s differences from the population are only due to chance, and chance shrinks as the sample grows.

A second kind of fairness is the control group. In 1747 the ship’s surgeon James Lind took twelve sailors sick with scurvy, kept them on the same diet in the same quarters, and gave each pair a different treatment. The pair who got oranges and lemons recovered within days; the others did not. Without the comparison, a recovery proves nothing, because people often recover on their own. Modern trials go further: patients are assigned to groups at random, the control group gets a placebo that looks identical, and neither patients nor doctors know who got which until the end.

When you read a study, ask three questions. How were the subjects chosen, and who was left out? Was there a control group, and was it treated exactly the same except for the one thing being tested? And how large was the sample, remembering that a small random sample beats a huge biased one every time?

The same data can be drawn to look like a crisis or like nothing at all. Suppose a city’s water use fell from 100 million gallons a day to 97 million over a decade. Plotted on an axis that runs from 0 to 100, the line is nearly flat. Plotted on an axis that runs from 96 to 100, the same line plunges toward the floor. Neither graph lies about the numbers, but one of them lies about their importance. Before you read the line, read the axes: where they start, what units they use and whether the scale is evenly spaced.

The second trick is the window. A stock that has fallen for five years can look like a winner if the chart shows only the last three months. Choosing the start and end points that make your case is called cherry-picking, and it is the most common way honest-looking graphs mislead. Ask what happened just before the graph begins and whether a longer view would change the story. A real trend holds up when you widen the window.

Watch also for two lines on two different vertical axes, which can be stretched to make anything appear to match anything; for bars whose width or picture size grows along with their height, so the eye sees a bigger change than the numbers show; and for lines with no error bars at all, which hide how uncertain each point is. A logarithmic scale, where each step up multiplies by ten, is not a trick, but it does make explosive growth look like a gentle slope, so check the labels.

None of this means graphs are untrustworthy. A well-made graph shows in one glance what would take a page of numbers to say. The habit to build is simply to read a graph the way you read a contract: the labels, the units, the range, the source and the sample size first, and the dramatic line last.

Measure a table with a meter stick and you might get 152.3 centimeters. Measure again and get 152.1. Neither is a mistake. In science the word error does not mean a blunder; it means the unavoidable gap between a measurement and the true value. Every instrument has a finest mark, every hand wobbles, every room changes temperature. The size of that gap is the uncertainty, and a measurement reported without it is only half reported.

Errors come in two kinds. Random error scatters readings above and below the true value with no pattern, like the wobble of a hand-held ruler. Repeating the measurement and averaging shrinks random error, because the highs and lows cancel. Systematic error pushes every reading the same way, like a scale that reads 2 grams heavy with nothing on it. Averaging does nothing for it; you find it only by checking the instrument against a known standard. A tight cluster of readings tells you the precision is high; it does not tell you their accuracy.

The measured speed of light shows how uncertainty gets smaller over time. Hippolyte Fizeau in 1849 used a spinning toothed wheel and got a value about 5 percent too high. Albert Michelson, using rotating mirrors and a long baseline, narrowed the range to a few kilometers per second by the 1920s. Each later scientist did not simply announce a number; they announced a number with a plus-or-minus, and the next experiment had to land inside or outside that range.

When you read a result, look for the plus-or-minus, the error bars on a graph or the words ’confidence interval.’ Two measurements that differ by less than their uncertainties agree; nothing has been discovered. A difference that is many times larger than the uncertainty is a real signal. Learning to tell those apart is most of what it means to read data.

A ruler marked in millimeters lets you read a pencil as 14.7 centimeters, with the last digit an honest estimate between marks. Writing 14.7000 centimeters would claim you know the length to a ten-thousandth of a centimeter, which the ruler cannot support. The digits that carry real information are the significant figures. In 14.7 there are three. The rule of the game is simple: report every digit you can defend and not one more.

Calculators are the main source of trouble. Divide a 14.7 centimeter pencil into 3 equal pieces and the screen says 4.9 centimeters; divide 14.7 by 7 and it shows 2.1, but the screen may first show 2.1000000. Those extra zeros mean nothing. A result cannot be more precise than the least precise measurement that went into it. When multiplying or dividing, keep the same number of significant figures as the measurement with the fewest. When adding or subtracting, keep the same number of decimal places as the roughest measurement. Counting numbers, like 3 pieces, are exact and do not limit anything.

The same idea guards you against false precision outside the lab. A headline that says the average adult walks 3,812 steps a day sounds exact, but a survey of a few hundred people with step counters of unknown quality cannot pin the number to the last step. A report that a hurricane has a 47.3 percent chance of landfall is claiming more than any forecast model can deliver. When a number carries more digits than its source could possibly justify, it is a sign that someone wants it to sound more certain than it is.

Rounding is therefore not laziness; it is honesty. Say about 3,800 steps, roughly a 50 percent chance, 4.9 centimeters. Keep the extra digits in your calculator during the work so rounding errors do not pile up, and round once at the end to the precision your data can actually support.

Before a scientific paper appears in a journal, editors send it to other scientists in the field for peer review. The reviewers look for weak methods, missing controls, overreaching conclusions and math errors, and they can demand changes or recommend rejection. Peer review catches a great deal, but it is not proof. Reviewers rarely see the raw data, cannot repeat the experiment, and are volunteers working for free. A peer-reviewed paper is a serious claim worth taking seriously, not a settled fact.

The deeper test is replication: can a different lab, following the published methods, get the same result? In March 1989 two chemists at the University of Utah announced at a press conference that they had produced nuclear fusion in a jar of water at room temperature. Labs around the world rushed to repeat it. Within months most had failed, the few early successes evaporated, and the claim collapsed. That is the system working. In 2015 a team of psychologists tried to repeat 100 published experiments and found that fewer than half held up, which has pushed many fields to raise their standards.

When a published paper turns out to be wrong through fraud or serious error, the journal issues a retraction, formally withdrawing it from the record. Retractions are rare and painful, but a field without them would be a field that never admits mistakes. Many scientists now also post preprints, drafts shared online before review, so that colleagues can criticize the work early. A preprint has not passed review at all, and news stories built on preprints deserve extra caution.

The lesson for a reader is that no single study is the final word. Ask whether the result has been repeated by independent groups, whether the raw data and methods are available, and whether the finding fits with the rest of what is known. A surprising result from one lab is a lead. A result that many labs have reproduced is knowledge.

Headlines compress. A study that found a small effect in 40 mice over eight weeks becomes ’Coffee Cures Memory Loss’ by the time it reaches your feed. The first questions to ask of any science story are the ones a scientist would ask of the paper behind it. Who did the study, and was it published in a peer-reviewed journal? What was studied, humans or animals or cells in a dish? How many subjects, and for how long? Was there a control group? Most stories that fall apart fall apart here.

The second trap is how risk is described. In 2015 a World Health Organization agency reported that eating about 50 grams of processed meat a day raised the relative risk of colorectal cancer by about 18 percent. Headlines compared bacon to cigarettes. But the lifetime absolute risk of colorectal cancer is around 5 or 6 percent, so an 18 percent increase moves it by about one percentage point. Both numbers are true. Only the absolute number tells you how much it matters to you, and stories almost always give you the relative one because it is bigger.

Third, follow the money and the megaphone. A university press release is written to attract attention, and a study of the topic found that exaggeration in news stories often begins in the release itself. A conflict of interest, such as a study of a sweetener funded by a sweetener company, does not prove the result is wrong, but it raises the bar for how carefully you should check it. Good journals require authors to declare funding; good stories mention it.

Finally, one study is one study. Science moves by the weight of many results, and a single paper that overturns decades of evidence is far more likely to be wrong than to be a revolution. The most reliable science stories are the least exciting ones: they report a body of work, quote scientists who were not involved, give absolute numbers and say plainly what is still unknown.

In the first half of the twentieth century, lung cancer went from a rare disease to a common one, and doctors argued about why. Cars, paved roads, factory smoke and cigarettes had all spread at the same time. In 1950 Richard Doll and Austin Bradford Hill published a case-control study from London hospitals. They compared hundreds of patients with lung cancer to similar patients with other diseases and asked about their habits. Almost every lung cancer patient was a smoker, and heavy smokers were far more common among the cases than the controls.

A case-control study looks backward and can be fooled by memory and by how patients are chosen, so Doll and Hill launched a cohort study in 1951. They wrote to every doctor in Britain and recorded the smoking habits of about 40,000 who replied. Then they simply waited to see who died of what. Within a few years the pattern was unmistakable: the heavier the smoking, the higher the lung cancer death rate, a clear dose-response relationship. Doctors who quit saw their risk fall over time. Similar studies in the United States found the same thing.

No one could run a controlled experiment forcing people to smoke, so the case had to be built from the Bradford Hill questions. The association was extremely strong, it showed a dose-response, smoking came before the cancer, it appeared in every country studied, and tar from cigarettes caused tumors when painted on the skin of mice, giving a mechanism. In January 1964 the United States Surgeon General’s committee reviewed the evidence and concluded that cigarette smoking causes lung cancer in men.

The tobacco industry responded not by producing better evidence but by manufacturing doubt. It funded studies designed to muddy the picture and insisted the question was still open. That strategy is worth remembering, because it has been copied for other products since. The demand for more evidence is a good scientific instinct; the demand for evidence that could never be enough is not.

High in the stratosphere, roughly 15 to 35 kilometers up, a thin layer of ozone, a molecule of three oxygen atoms, absorbs most of the sun’s harmful ultraviolet light. In 1974 the chemists Mario Molina and Sherwood Rowland published a warning. CFCs, the chlorofluorocarbon gases then used in spray cans, refrigerators and foam, were so stable that nothing at ground level destroyed them. They would drift upward for decades until ultraviolet light broke them apart, releasing chlorine atoms. Each chlorine atom, they calculated, could act as a catalyst and destroy thousands of ozone molecules before being removed.

This was a claim built from laboratory rates and a model of the atmosphere, not from any observed loss. Critics, including the companies that made CFCs, said the effect was speculative. Then in 1985 three scientists from the British Antarctic Survey, Joseph Farman, Brian Gardiner and Jonathan Shanklin, reported that the ozone above their Halley Bay station every October had fallen sharply since the 1970s, to far below anything the models had predicted. A satellite had been measuring the same region, but its software had been flagging the extremely low values as instrument errors.

The Antarctic ozone hole was a stronger effect than Molina and Rowland had predicted, and the reason turned out to be ice clouds in the polar winter that speed up the chlorine chemistry. Within two years, in 1987, nations signed the Montreal Protocol to phase out CFCs, and the treaty has been tightened several times since. Molina and Rowland shared the 1995 Nobel Prize in Chemistry with Paul Crutzen. Measurements show the amount of ozone-destroying chlorine in the stratosphere peaking around the turn of the century and slowly falling, and the ozone layer is now recovering, with the Antarctic hole expected to heal around the 2060s.

The case shows the full arc of an argument from evidence. A mechanism and a prediction came first. Independent measurements confirmed the prediction, and the exceptions, the values too low to believe, turned out to be the most important data. The claim survived attack, the world acted, and later measurements tested whether the action worked.

In February 1998 the medical journal The Lancet published a paper describing twelve children with developmental problems, most of them autism, whose parents linked the symptoms to the measles, mumps and rubella vaccine. Twelve children, no control group, no comparison to unvaccinated children, and a hypothesis raised at a press conference by the lead author, Andrew Wakefield, who suggested the combined vaccine be split up. Vaccination rates in Britain fell. Measles, which had nearly vanished, came back.

The proper way to test the hypothesis is to compare large groups of vaccinated and unvaccinated children and see whether autism is more common in one. Researchers in Denmark, where national health records track every child, did exactly that. A 2002 study of more than 500,000 children found no difference in autism rates. A 2019 study of more than 650,000 children, including children considered at higher risk, found none either. Studies in the United States, Japan and elsewhere agreed. Autism diagnoses have risen, but the rise tracks broader definitions and better screening, and it appears in vaccinated and unvaccinated children alike.

Meanwhile an investigation found that the 1998 paper’s data had been altered, that the children had been recruited through lawyers preparing a lawsuit against vaccine makers, and that the lead author had been paid by those lawyers. In 2010 The Lancet retracted the paper, and Britain’s medical council struck Wakefield from the register of doctors. The paper’s central claim had failed every test that evidence can apply: it could not be replicated, its data were unreliable and its author had an undisclosed conflict of interest.

The case matters not because scientists were wrong for a while, which happens, but because of how the two kinds of evidence compare. On one side is a story about twelve children told by a person with a financial stake. On the other are cohort studies of more than a million children with no link found. Learning to weigh those two against each other, rather than treating them as two equal sides, is the whole skill this chapter has been about.

Why does bread rise? Are plants happier in the sun? Does music help you study? Each of these is a real curiosity, but none of them is yet a testable question, one that an investigation could actually answer. ’Happier’ is not something a plant shows on a dial. ’Help you study’ could mean a dozen things. The first job in any investigation is to sharpen the curiosity until it points at something you could measure and something you could change.

The sharpening step is to write an operational definition for every fuzzy word. Instead of ’happier,’ choose ’height in centimeters after three weeks’ or ’number of leaves.’ Instead of ’help you study,’ choose ’score on a 20-word recall test after ten minutes.’ Now the question can be rewritten: how does the number of hours of daily light affect the height of bean seedlings after 21 days? It names the thing you will change, the thing you will measure and the time frame. Someone else could run it.

From the question comes a hypothesis, a proposed answer with a reason behind it: seedlings with more daily light will grow taller, because light supplies the energy for photosynthesis. And from the hypothesis comes a prediction that could fail: if we give one group 4 hours and another 12 hours, the 12-hour group will be taller at 21 days. If they are not, the hypothesis is in trouble. A question whose every possible answer would leave you believing the same thing is not worth the time.

McClintock’s question was of this kind. Not ’why are these kernels strange?’ but ’where on the chromosome is the element that switches the color gene off, and does it stay put across generations?’ Each cross she planned was a prediction that the next season’s ears could confirm or wreck. Your capstone question needs the same shape, whatever its subject.

Every experiment has three kinds of quantities. The independent variable is the one you deliberately change, such as the hours of light a seedling gets. The dependent variable is the one you measure to see the effect, such as height. The controlled variables are everything else that could matter and that you hold the same for every group: the kind of seed, the soil, the pot size, the water, the temperature. A careless experiment lets a controlled variable drift, and then you cannot tell whether the light or the drift caused the difference.

The comparison is what makes a test fair. If every seedling gets 12 hours of light, you learn what 12-hour seedlings look like but nothing about the effect of light. You need at least two levels of the independent variable, and it is usually best to have several, so that you can see whether the effect grows steadily or levels off. A control group, given the ordinary or zero condition, anchors the comparison. In a medical trial the control gets a placebo; in a plant experiment the control might be the light level seedlings normally get.

One variable at a time is the beginner’s rule, and it is a good one. It is not the only design, though. Statisticians led by Ronald Fisher in the 1920s and 1930s worked out how to vary several factors at once in carefully arranged patterns and then untangle their separate effects, which is how modern agricultural and industrial experiments are run. What never changes is the principle: the only difference between the groups should be the one you put there, and chance should decide which unit goes into which group.

That last point, random assignment, deserves emphasis. If you put the healthiest-looking seedlings in the 12-hour group because you expect them to do well, you have built your expectation into the result. Number the pots, flip a coin or use a random number list, and let chance sort them. Then a difference at the end can only come from the treatment or from chance, and you can calculate how likely chance is.

A protocol is the written procedure for an investigation, detailed enough that a stranger could follow it and get the same kind of data. It lists materials, the exact steps, the order of steps, the timing, what will be measured and how. Writing it forces decisions you would otherwise make on the fly, and those on-the-fly decisions are where bias creeps in. It also lets a reader judge the work later, which is why published papers include a methods section.

Two numbers matter most in the plan: how many subjects and how many trials. One seedling per group tells you almost nothing, because seedlings differ from one another for reasons that have nothing to do with light. Ten per group lets the individual differences average out. Repeating each measurement, and repeating the whole experiment if you can, shows how much scatter to expect, and without knowing the scatter you cannot say whether a difference is real. As a rule, plan for more subjects than you think you need, because some will be lost.

Before committing weeks to the full experiment, run a pilot study: a small, quick version to find out what goes wrong. The pots may dry out faster than expected; the recall test may be too easy; the microscope slide may need a different stain. Pilots are cheap and save you from discovering a broken procedure at the end. Many scientists now also write down their hypothesis and analysis plan before collecting data, a practice called preregistration, so that they cannot quietly change the question to match the result.

Finally, plan for safety and ethics. Chemicals, heat, sharp tools and electricity need specific precautions in writing. Any investigation involving people, even a survey of classmates, needs their informed consent and a way to keep their answers private. Research on animals and people at universities must be approved by a review board before it starts. A plan that skips these is not ready.

Data collection begins before the first measurement, with the instrument. Calibration means checking the instrument against a known standard: a balance against a certified mass, a thermometer in an ice-water bath that should read 0 degrees Celsius. An uncalibrated instrument can be perfectly precise and perfectly wrong. Know its resolution too, the smallest change it can show. A ruler marked in millimeters cannot report tenths of a millimeter, no matter how carefully you squint.

Design the data table before you collect anything. Each column gets a label and a unit; each row is one trial or one subject. Leave room for the date, the time, the conditions and a notes column. Record raw data, the numbers as the instrument gave them, and never overwrite them. Do calculations in separate columns later. A smudged, corrected or ’improved’ raw number is worthless, because no one, including you, can tell what was actually observed. Scientists keep bound notebooks with numbered pages for exactly this reason.

Write down everything unusual. A pot got knocked over; the room was unusually warm on day 9; the microscope lamp flickered. When a strange value, an outlier, appears in the data, the notes tell you whether it came from a real event or from a known accident. Throwing out an outlier because it spoils the pattern is not allowed. Throwing it out because the notes say the sample was contaminated, and saying so in the report, is good practice.

If you use this site’s microscope for a project, the same rules apply: note the magnification, the stain, the time since the slide was made and the field of view you counted. Two counts of cells in a drop of pond water made under different conditions are not the same measurement, and the notes are what let you tell the difference.

With the data in hand, the first step is to summarize each group. The mean, the sum divided by the count, is the everyday average. The median, the middle value when the data are sorted, is better when a few extreme values would drag the mean around. The range, or a more careful measure of spread like the standard deviation, tells you how scattered the values are. Report both a center and a spread; a mean without a spread is a claim without an uncertainty.

Next, plot the data. Put the independent variable on the horizontal axis and the dependent variable on the vertical, with every point shown, not just the averages. A picture reveals what a table hides: whether the relationship is a straight line, a curve that flattens out, or no relationship at all. If the points fall roughly along a line, a best-fit line drawn through them gives a rate, and its slope has units, such as centimeters of growth per hour of light per day.

The key question is whether the pattern is larger than the noise. If the 12-hour seedlings average 14 centimeters and the 4-hour seedlings average 12, but each group’s values range from 9 to 17, the difference is inside the scatter and you have not shown anything. If the ranges barely overlap, the difference is real. Statistical tests put a number on this: how often would chance alone produce a difference this big? A result that chance would produce fewer than 1 time in 20 is usually called significant, but that is a convention, not a law of nature, and a tiny significant difference may not matter.

Beware of the pattern that fits too well. With enough freedom you can draw a wiggly curve through any set of points, and it will predict nothing. Prefer the simplest relationship that fits the data within its uncertainty, and check whether it fits new data you did not use to draw it.

A model is a stand-in for something too big, too small, too slow or too complicated to work with directly. A globe is a model of Earth. An equation relating a pendulum’s length to its period is a mathematical model. A computer simulation of the atmosphere, dividing the air into millions of boxes and stepping their temperature and wind forward in time, is a model too. Every one of them leaves things out on purpose. The globe has no weather; the pendulum equation ignores air resistance; the climate model cannot track every cloud.

The leaving-out is the point, and it is also the danger. Each model rests on assumptions, and it is only reliable where those assumptions hold. The pendulum equation works beautifully for small swings and fails for wide ones. The bead-on-a-string picture of genes that McClintock’s colleagues held was a useful model that made good predictions for decades, right up until her corn showed genes moving. A model that has never been wrong may just never have been tested outside its comfortable range.

So the test of a model is its record. Does it predict measurements it was not built from? Where does it start to fail, and does the failure make sense? Weather models are checked every single day against what actually happened, which is why a five-day forecast today is about as good as a three-day forecast was decades ago. When a model and a careful measurement disagree, the measurement usually wins, and the disagreement is where the next discovery hides.

For your own project you will build at least a simple model: a line through data points, a diagram of a food web, a formula relating your variables. State its assumptions. Say where you expect it to hold and where you do not. A reader who sees that you know the limits of your model will trust the parts inside them.

Science asks what is true; engineering asks what will work. A school wants to cut the flooding on its playing field after storms. Before anyone sketches a solution, the problem needs a sharp definition. The criteria are what a successful design must do, written so that they can be measured: standing water gone within 12 hours of a 2-centimeter rain; no damage to the running track; safe for students. Vague criteria like ’better drainage’ cannot be tested and cannot settle an argument between two designs.

The constraints are the limits every design must stay inside: a budget of a certain amount, a completion date before the fall season, no digging near the buried gas line, compliance with the city’s stormwater rules. A design that fails a constraint is out, no matter how well it scores on the criteria. Listing constraints early saves everyone from falling in love with a design that could never be built.

Almost no design wins on every criterion, so engineers face trade-offs. A deep gravel trench drains fastest but costs the most and takes longest to build. A raised field is cheap but changes the playing surface. A decision matrix makes the trade-offs visible: list the designs as rows and the criteria as columns, score each cell, weight the criteria by how much they matter, and add up. The matrix does not decide for you, but it forces you to say out loud which criteria you value most and why.

This is the same claim, evidence and reasoning habit in a new dress. The claim is that a design meets the criteria within the constraints; the evidence is test data and cost figures; the reasoning shows how the trade-offs were weighed. Notice that criteria and constraints usually come from people, not from nature, and that different stakeholders, the coach, the neighbors, the budget office, will weight them differently.

Engineering rarely gets a design right the first time, and it does not expect to. The design cycle runs: define the problem, brainstorm, build a prototype, test it, study what went wrong, revise, and test again. A prototype is a working version made quickly and cheaply enough that you are willing to break it. A cardboard model of the drainage trench, a scale bridge made of craft sticks, a spreadsheet of the costs: each is a prototype of a different part of the design.

The test must be against the written criteria, not against a feeling that it looks good. Load the craft-stick bridge with weights until it breaks and record the mass. Pour a measured amount of water on the model field and time how long it stands. Comparing prototypes on the same test is what turns opinion into evidence. Numbers from a fair test also settle arguments between team members far more peacefully than volume does.

The most valuable step is failure analysis: figuring out why a prototype failed rather than just that it did. Did the bridge break at a joint or in the middle of a stick? Did the water pool because the slope was wrong or because the gravel clogged? In 1901 the Wright brothers found that the published tables of lift they had trusted were wrong, built their own wind tunnel, and tested more than a hundred wing shapes before the 1903 flight. In 1940 the Tacoma Narrows Bridge twisted itself apart in a moderate wind, and the study of that failure changed how every long bridge since has been designed.

Each cycle is an iteration, and the number of iterations is not a sign of poor work. Commercial products often go through hundreds or thousands of prototypes. The record of what was tried, what failed and why is as important as the final design, because the next engineer, or you next year, will need it.

The claim comes first, in one sentence, answering the question you asked. ’Bean seedlings given 12 hours of daily light grew taller over 21 days than seedlings given 4 hours.’ Not ’light affects plants,’ which is too vague to test, and not ’light is essential to all life,’ which your data cannot support. State it in the past tense about what you found, and keep it inside the range you tested.

The evidence follows, with numbers. Means with their spread, the number of subjects, and a graph or table the reader can inspect. ’The 12-hour group averaged 14.2 centimeters, ranging 12.5 to 16.0, across 10 seedlings; the 4-hour group averaged 9.8 centimeters, ranging 8.1 to 11.4.’ Then the reasoning: the scientific principle that explains why this evidence supports the claim, such as light supplying the energy for photosynthesis, which builds the sugars used for growth. Reasoning is where you connect your small experiment to what is already known.

The strongest explanations then turn on themselves. Name the alternative explanations you considered and say what evidence rules them out: could the 12-hour lamp also have warmed the pots? If you measured temperature and it was the same, say so; if you did not, say that too. State the limitations: one bean variety, three weeks, indoor conditions. A reader who sees you searching for your own weaknesses is a reader who begins to trust you.

Finally, give credit. A citation for every fact, value or method that came from someone else tells the reader where to check and shows that you built on real work rather than invention. McClintock’s papers were dense with cross-references to decades of maize genetics, which is part of why they held up when they were finally read carefully.

A scientific presentation, whether a talk, a poster or a written report, is an argument, and arguments draw fire. Prepare for it. Before you present, list the five hardest questions someone could ask and work out your answers. Most objections come in a few forms: your sample was too small, you did not control something, there is a simpler explanation, or your numbers do not show what you say they show. Each of these is a counterclaim, and each deserves a rebuttal that points at evidence, not at your effort or your confidence.

Some objections will be right. The honest response to a good objection is to say so, and to say what you would do about it: ’That is a fair point; we did not measure temperature, and the lamp could have warmed the pots. The next run should include a thermometer in each pot.’ Scientists who cannot say ’you are right’ in public are not trusted in private either. The goal of a defense is not to win but to find out, together with your critics, what the evidence actually supports.

Learn to separate two kinds of disagreement. Whether CFCs destroy ozone is a question of fact, settled by measurement. Whether the cost of phasing them out was worth it is a question of values, on which measurement informs but does not decide. In a poster session or a public meeting, people mix the two constantly. Naming which kind of question is on the table is often the most useful thing a presenter can do.

Two cautionary tales frame the skill. The cold fusion scientists of 1989 were confident, presented dramatically and were wrong. McClintock was quiet, thorough and right, but her audience was not ready and she stopped pushing. The lesson is not that confidence is bad or that being ignored proves you correct. It is that a claim stands or falls on the evidence and on how carefully it was gathered, and the presenter’s job is to make that evidence as clear and as checkable as possible.

Your capstone project is one investigation, chosen by you, carried through every stage this unit has covered. It can be an experiment, a field study, an analysis of existing data or an engineering design, but it must produce evidence and end with a defended claim. Start with what you are curious about near you: the temperature of Chicago’s lakefront versus an inland neighborhood, the number of organisms in a drop of pond water under this site’s microscope, the craters visible on the moon with its telescope on different nights, the drainage on your own school’s field. Local questions are easier to measure and harder to fake.

The written plan has a fixed set of parts. The testable question and hypothesis. Background: what is already known, with citations. The variables, identifying which is independent, which dependent and which controlled, or the criteria and constraints if it is a design. The protocol, step by step, with materials. The data table, empty, with units. The analysis plan: what you will calculate and what graph you will draw. Safety and ethics. And a line stating what result would make you abandon the hypothesis.

Then the timeline. Work backward from the presentation date and set milestones: plan approved, pilot run done, data collection finished, analysis done, draft written, practice defense. Plants take weeks to grow and pond water changes with the seasons, so the calendar is a real constraint. Build in slack, because a pilot will reveal something that needs changing.

Everything in the plan is a promise to your future self and to your readers about how the work will be done, made before the results can tempt you to bend it. That is preregistration in miniature, and it is the single best protection against fooling yourself. When the data come in, you will analyze them the way you said you would, report what you find whether or not it is what you hoped, and stand up to defend it. That is the whole practice of science in one project.