When Science Results Disagree | A Worked Casebook for Choosing the Next Useful Test

This guide teaches readers what to do when two apparently careful investigations disagree. The purpose is not to decide which side “wins” by instinct, but to locate where the investigations stopped being comparable, preserve the original evidence, and choose the smallest next test that can discriminate among the remaining explanations.

Scope of this guide: this worked casebook connects the eduKate Yishun guides for measurement, experiments, models, evidence and explanation. It does not replace those foundation articles, nor does it replace school-level Science pages. It addresses a practical question: results disagree; what do we inspect next?

Reader-safety boundary: all people, investigations, numerical datasets and classroom exchanges below are fictional teaching constructions. They are not reports of eduKate students, evaluated interventions, official examination questions or real laboratory findings. Do not use these examples to judge drinking-water safety, diagnose illness, certify equipment or handle hazardous materials.

When Science Results Disagree

Two groups place their results on the same table.

The first says a treatment helped.

The second says it made no difference.

Both groups followed instructions. Both have numbers. Both can explain why their answer seems reasonable.

Someone suggests repeating everything.

Someone else suggests averaging the two answers.

A third person looks at the less tidy graph and decides that this must be the group that made the mistake.

None of those responses has yet identified the problem.

The useful question is smaller:

Where, exactly, did the two investigations stop being comparable?

Perhaps they measured different outcomes. Perhaps they used different units. Perhaps the treatment was applied to one whole tray although the analysis counted each seed as an independent experiment. Perhaps one instrument was read immediately and the other after it had settled. Perhaps one group excluded an awkward value. Perhaps both groups used the same faulty reference. Perhaps the apparent dispute is smaller than the uncertainty surrounding either estimate.

This casebook is for that moment. It does not replace a textbook on measurement, experimental design or statistics. It teaches the practical movement between them: preserve the record, define the disagreement, compare explanations, and choose an observation that would actually help.

Choose the disagreement you actually have

Expandable contents

1. First establish whether the two claims are the same
2. Preserve the notebook before repairing the answer
3. Check units, denominators and direction
4. Two instruments disagree: do not make either one the judge
5. Shared error can disappear from a difference
6. Ten readings are not necessarily ten experiments
7. The treatment followed the shelf
8. Use a crossed comparison to locate disagreement
9. The slower sensor may reach the same final state
10. Detected and not detected can be compatible reports
11. The sample changed before it reached the instrument
12. The inconvenient point is not automatically wrong
13. Missing readings can change the answer
14. Pooled rates can reverse subgroup comparisons
15. Same final result, different journey
16. The graph changed, but the observations did not
17. The model was asked outside its tested range
18. Perfect correlation does not mean methods agree
19. Five articles can still be one piece of evidence
20. Replay the calculation before rerunning the experiment
21. Case conference: which cloth should the class choose?
22. Case conference: the concentration doubled
23. Teach disagreement without teaching distrust
24. Write a result another reader can safely use
25. Choose the next test by the decision it could change

1. First establish whether the two claims are the same

The first report says, “Material A absorbs more water.” The second says, “Material B is more absorbent.” At first sight they contradict each other.

Before deciding who is right, ask each group to complete one sentence:

Our result is the amount of ___, measured in ___, for ___, after ___.

Suppose Group A tested equal-sized squares. Its square of Material A retained 12 g of water while Material B retained 9 g. Group B tested equal dry masses. Its 1 g sample of A retained 3 g while 1 g of B retained 4.5 g.

Both calculations can be correct.

One question is “how much water can this fixed area hold?” The other is “how much water can this material hold per unit mass?”

The disagreement was hiding in the denominator.

This is a common scientific failure mode. Two reports use the same everyday word—growth, strength, efficiency, temperature, improvement—while operationalising it differently.

The repair is not to force the values into one ranking. The repair is to name the measurand precisely, then decide which measure serves the practical question.

A useful conclusion would be:

In the supplied results, A held more water per tested square while B held more water per gram of dry material. These answer different questions. A common procedure is needed before ranking the materials for a specified use.

The first diagnostic move is therefore semantic before it is statistical: make sure the claims are actually about the same thing.

2. Preserve the notebook before you repair the answer

A group notices that one length is written as 43.2 cm while nearby readings are around 13 cm. A pupil reaches for an eraser.

Stop.

The number may be a transcription error. It may be a ruler-position error. It may refer to the wrong sample. It may be a real unusual observation. The current record does not yet tell us which.

Erasing it destroys part of the evidence required to diagnose the problem.

The correct habit is:

1. preserve the original entry;
2. record the suspected problem;
3. look for independent evidence such as a photograph, instrument log or raw export;
4. record any correction separately with the reason;
5. retain both the original and corrected versions.

If a photograph clearly shows 13.2 cm for that sample, a transcription correction is supported. If the only reason for changing 43.2 to 13.2 is that 13.2 makes the graph smooth, the correction is not justified.

This distinction matters because data cleaning can quietly become data rewriting.

A smooth graph is not the standard of truth. Traceability is.

3. Check units, denominators and the direction of comparison

Many disagreements disappear when units are made explicit.

One team reports 0.50 g/L. Another reports 500 mg/L. Those values are equal.

One team reports “a 20% decrease.” Another reports “80% of baseline.” Those statements can describe the same final state.

One team reports success per participant. Another reports success per attempt. The denominators differ even if both use percentages.

Direction matters too.

An “80% increase” from 10 to 18 is not reversed by a “80% decrease” from 18. A decrease of 8 from 18 is about 44.4%.

Percent change is asymmetric because the reference value changes.

Before treating percentages as contradictions, recover the numerator, denominator, baseline and direction.

A practical worksheet should therefore include raw counts wherever possible. “81 of 110” is more informative than “73.6%” when the reader needs to understand how the proportion was formed.

4. Two instruments disagree: do not make either one the judge

Two thermometers are placed in the same bath. One reads 20.4°C. The other reads 21.1°C.

Which is correct?

The answer is not “the newer one,” “the digital one,” or “the one with more decimal places.”

Two disagreeing instruments cannot resolve the dispute by comparing only with each other.

We need an external reference, calibration history, a third independent method, or another test that distinguishes plausible failure modes.

Suppose both instruments are checked against a traceable reference at several temperatures. Thermometer A differs by +0.1°C across the range. Thermometer B differs by +0.8°C. That evidence supports treating A as closer under those conditions.

But one reference point is not always enough. An instrument can be correct at one temperature and nonlinear elsewhere.

When instruments disagree, ask:

• Were they exposed to the same physical location and time?
• Were both allowed to settle?
• Do they have the same response time?
• Are they within their calibrated range?
• Do they share the same reference or manufacturer calibration?
• Could both be wrong in the same direction?

The aim is not to crown a favourite instrument. It is to locate the disagreement.

5. A shared error can disappear from a difference without disappearing from the measurements

Suppose a balance has a stable +0.5 g offset.

Sample A reads 10.5 g. Sample B reads 14.5 g.

The absolute masses are biased. Yet their difference is still 4.0 g if the same offset applies equally to both.

This is a useful reminder: an error can cancel in one calculation while remaining present in the underlying measurements.

The opposite also occurs. Two measurements may each have small errors that amplify when subtracted.

If A = 10.0 ± 0.5 and B = 10.4 ± 0.5, the difference of 0.4 is not automatically meaningful just because the subtraction is exact.

Shared errors also complicate independence. If two laboratories use standards prepared from the same faulty stock, agreement between them is weaker evidence than it appears.

When interpreting agreement or disagreement, ask whether the measurements have independent error sources.

6. Ten readings are not necessarily ten experiments

A group applies one treatment to one tray containing ten seedlings. It then measures ten seedlings and reports n = 10.

But what was independently treated?

If the entire tray received one shared treatment environment, the tray may be the experimental unit. The ten seedlings are subsamples within that unit.

This matters because counting subsamples as independent experiments can make the evidence look more precise than it is.

Likewise:

• ten readings from one water sample are not ten independent water samples;
• ten photographs of one specimen are not ten independent specimens;
• ten pupils from one classroom are not automatically ten independent classroom interventions;
• ten technical assay wells prepared from one biological sample may not be ten biological replicates.

The key question is:

At what level was the condition independently assigned or generated?

Repeated measurement is useful. It estimates measurement repeatability. It does not create new independent experimental units by arithmetic.

7. The treatment followed the shelf

Imagine all treated plants are placed on the upper shelf and all controls on the lower shelf.

The treated plants grow more.

Did the treatment work?

Perhaps. But shelf position is now perfectly aligned with treatment. If the shelves differ in light, temperature or airflow, those factors are confounded with treatment.

No statistical technique can reconstruct what would have happened to treated plants on the lower shelf because that condition was never observed.

A stronger design could randomise pots across positions or block by shelf and place both treatment and control within each block.

This chapter teaches a powerful diagnostic rule:

If treatment follows another factor perfectly, the experiment cannot separate them.

The same problem occurs when all controls are measured in the morning and all treatments in the afternoon, when all samples from one group are processed in one batch, or when each condition is handled by a different operator.

Good design breaks unwanted alignments before data collection.

8. Use a crossed comparison to locate the disagreement

Suppose two instruments and two samples disagree. Instead of asking which instrument is “bad,” cross the comparison.

Measure Sample A and Sample B on Instrument 1 and Instrument 2, preferably in balanced order.

Four outcomes appear.

If both instruments agree that A > B but differ by a constant offset, the disagreement may lie in calibration rather than relative ordering.

If Instrument 1 says A > B while Instrument 2 says B > A, the problem is more serious. It may involve selectivity, nonlinearity, matrix effects or sample handling.

If disagreement appears only for Sample B, ask what makes B different.

Crossed designs are powerful because they separate factors that were previously entangled.

The general form is:

vary one suspected source of disagreement across the other so that their effects can be distinguished.

This is often more informative than repeating the same original comparison ten more times.

9. The slower sensor was not necessarily measuring a different final state

Two sensors are placed into a new environment.

Sensor A reaches 25.0 quickly. Sensor B reads 22.0, then 23.5, then 24.6, then 25.0.

If the instruments are compared after ten seconds, they disagree. After two minutes, they agree.

The disagreement was about response time, not final equilibrium.

This distinction matters whenever instruments have different dynamics.

Thermometers, humidity sensors, dissolved-gas probes, biological assays and imaging systems can all respond over different timescales.

Therefore record when a reading was taken relative to the change being measured.

An instrument can be accurate at equilibrium yet unsuitable for capturing a rapid transient.

Conversely, a fast sensor can have greater noise while a slower sensor smooths variation.

The right instrument depends on the time structure of the scientific question.

10. “Detected” and “not detected” can be compatible reports

Laboratory A reports that a substance was detected. Laboratory B reports “not detected.”

This sounds contradictory.

But suppose A has a lower detection limit than B.

If the true concentration lies above A’s detection capability but below B’s, both reports can be consistent with the same sample.

“Not detected” does not mean zero. It means the method did not provide evidence above its detection criterion under those conditions.

Likewise, “detected” does not automatically mean accurately quantified.

Near the detection threshold, a method may distinguish signal from background without estimating quantity precisely.

When reports differ, compare:

• detection limits;
• quantification limits;
• sample preparation;
• instrument selectivity;
• blank values;
• confirmation methods;
• units and reporting conventions.

The vocabulary can create a dispute that the underlying measurements do not.

11. The sample was changed before it reached the instrument

Two teams collect water from the same source.

Team A filters before analysis. Team B analyses the unfiltered sample.

Their reported concentrations differ.

Which is correct?

Possibly both.

Filtering can remove particle-bound material. The filtered sample may represent a dissolved fraction while the unfiltered sample represents a broader total fraction.

Sample preparation is part of the measurement chain.

Other preparation differences include drying, grinding, dilution, extraction, digestion, storage temperature, preservation, centrifugation and elapsed time before analysis.

A sample can change chemically or physically before the instrument ever sees it.

Therefore a disagreement between final numbers may originate upstream.

When comparing reports, reconstruct the sample’s entire journey.

12. The inconvenient point is not automatically the wrong point

Five readings are 10.1, 9.9, 10.0, 10.2 and 13.8.

The fifth value looks unusual.

Should it be removed?

Not because it is inconvenient.

Investigate.

Was the sample label wrong? Did the instrument alarm? Was the procedure different? Is there an independently documented reason to exclude it?

If the observation is valid, it may reveal genuine heterogeneity, a regime change or a process the model does not capture.

If there is a clear procedural failure, exclusion may be justified—provided the rule and evidence are documented.

Outlier treatment should not be determined by whether the point improves the desired conclusion.

A strong report can show results both with and without a disputed point and explain why the main conclusion does or does not depend on that choice.

13. Missing readings can change the answer without appearing in the graph

Suppose six samples were assigned to Group A, but only four have recorded outcomes.

The observed values total 40, giving an observed mean of 10.

If the two missing values could each plausibly lie between 4 and 16, the complete-group mean could range from:

(40 + 4 + 4) / 6 = 8

to

(40 + 16 + 16) / 6 = 12

This is not a confidence interval. It is a simple sensitivity bound based on supplied limits.

If Group B’s mean is 9, the missing outcomes could reverse the ranking.

If Group B’s mean is 15, the ranking would remain the same under the stipulated bounds.

This gives us a useful principle:

The value of missing information depends on whether plausible values could change the decision.

Missing data also have causes. Samples may be lost randomly, or failure may occur more often in one condition. The second situation is more dangerous because missingness can carry information about the outcome.

Keep the original denominator visible. Do not let a graph of observed cases silently pretend that unobserved cases never existed.

14. Both within-group comparisons can point one way while the pooled comparison points another

Consider this fictional table:

Task category	A	B
Easy	9/10 = 90%	80/100 = 80%
Difficult	20/100 = 20%	1/10 = 10%

Within both categories, A has the higher success rate.

But pooled across categories, A has 29/110 ≈ 26.4% while B has 81/110 ≈ 73.6%.

The overall ranking reverses because the methods were tested on very different task mixtures.

A faced mostly difficult tasks. B faced mostly easy tasks.

The arithmetic is correct. The question changed.

The pooled result answers, “How did these methods perform on the actual mixtures they happened to receive?”

The stratified result answers, “How did they compare within each difficulty category?”

If the intended deployment population is half easy and half difficult, a standardised comparison might weight the categories equally. But that target distribution must be justified rather than chosen after seeing which method wins.

The lesson is broader than this table:

Averages carry weights. Inspect who or what received the weight.

15. Same final result, different journey

Two processes both finish at 20 units after twenty minutes.

One rose rapidly to 30 and then fell. The other rose gradually from 0 to 20.

If we measure only the endpoint, they look identical.

If the scientific question concerns peak exposure, rate of change or mechanism, they are very different.

This appears in temperature histories, chemical concentration curves, learning trajectories, population dynamics and physiological responses.

Endpoint agreement can hide pathway disagreement.

Likewise, equal averages can hide different variability. Two classes may have the same mean but very different distributions. Two instruments may have the same mean bias but one may drift over time.

When final values agree unexpectedly, ask whether the journey matters for the question.

16. The graph changed, but the observations did not

The same data can look calm or dramatic depending on the graph.

A truncated vertical axis can make a small difference appear enormous. A very wide axis can make a substantial proportional change look flat.

Logarithmic scales can reveal multiplicative structure but confuse readers who expect equal vertical distances to represent equal absolute changes.

Smoothing can reveal broad trends while hiding local variation.

Normalising each series to its own baseline can help compare relative change while removing absolute differences.

None of these transformations is automatically dishonest.

The danger appears when the reader is not told what transformation occurred or when the chosen view is used to imply a claim the raw quantities do not support.

A useful disagreement protocol is to replot the same data in several defensible ways and ask which scientific question each representation answers.

The graph is an interface to the evidence. It is not the evidence itself.

17. The model was asked a question outside its tested range

A calibration relationship is validated from 0 to 100 units. A sample produces a signal corresponding to 140 if the line is extended.

Can we trust 140?

Not automatically.

Sensors can saturate. Chemical reactions can become nonlinear. Detector response can change. The model may no longer represent the system.

The safest move, if the method allows, is to bring the sample into the validated range—for example by an appropriate dilution—and remeasure.

This is not merely a laboratory lesson. It applies to any model.

A relationship observed between 10°C and 30°C may fail at 300°C. A learning model developed for one age group may not transfer unchanged to another. A traffic model calibrated under ordinary flow may fail during evacuation conditions.

Extrapolation is a new scientific claim. It needs evidence.

18. Perfect correlation does not mean the methods agree

Imagine Method B always reports exactly twice Method A.

The correlation between them can be perfect.

But the methods do not agree numerically.

Correlation measures co-variation, not equivalence.

Two methods can also have strong correlation while one has a constant offset or proportional bias.

For method comparison, inspect differences, ratios, calibration relationships and error across the range rather than relying on correlation alone.

This distinction matters whenever one method is proposed as a replacement for another.

A method can track ranking beautifully and still require correction before values are interchangeable.

19. Five articles can still be one piece of evidence

A literature review finds five papers supporting the same conclusion.

That sounds like five independent lines of evidence.

But suppose all five analyse the same public dataset.

Or three papers come from the same laboratory using the same instrument, sample bank and preprocessing pipeline.

Or several reviews all cite the same original experiment.

The number of publications is not the same as the number of independent evidence sources.

This is evidence ancestry.

When studies agree, trace where their data, methods, assumptions and references come from.

Convergence is strongest when methods have genuinely different failure modes.

This principle also applies inside the eduKate ecosystem. Internal links are learning routes, not independent corroboration. A linked eduKate article can teach a concept but does not become a second scientific source merely because it is another page.

20. Replay the calculation before you rerun the experiment

Two groups have the same raw data but different conclusions.

Before repeating the experiment, replay the analysis.

Did one group average replicates before calculating the final quantity while the other pooled everything first?

Did one exclude missing values differently?

Did one convert units incorrectly?

Did one subtract a blank before normalisation while the other subtracted afterward?

Did one calculate percent change from baseline and the other from final value?

A reproducible analysis trail can resolve such disputes without collecting a single new sample.

This is an efficiency principle:

Before acquiring more data, determine whether the disagreement already exists in the transformation from raw record to reported result.

Versioned scripts, saved spreadsheets, formulas and analysis notebooks make this easier.

Science becomes more auditable when the path from raw observation to conclusion can be replayed.

21. Case conference: which cloth should the class choose?

A class must choose a cloth for a fictional spill-response exercise.

Group A says Cloth X is best because a 10 cm square retained 18 g of water while Cloth Y retained 14 g.

Group B says Cloth Y is best because 1 g of Y retained 5 g of water while 1 g of X retained 3 g.

Do not ask the class to vote.

Ask what “best” means.

If each pad must fit into a fixed area, the area-based measurement may matter.

If supplies must be carried with a strict mass limit, the mass-based measurement may matter.

If cost matters, neither result is enough.

If repeated use matters, durability and drying may matter.

The scientific disagreement becomes a decision problem.

The next useful test is therefore determined by the decision variable that remains unresolved.

This is an important transfer: science does not always deliver one universal ranking. It often delivers a map of trade-offs.

22. Case conference: the concentration doubled

A solution’s measured concentration rises from 2 mg/L to 4 mg/L.

One pupil says twice as much substance must have been added.

Another says the volume could have halved through evaporation.

Both recognise that concentration is a ratio.

The measurement alone does not identify which part of the ratio changed.

To distinguish the explanations, measure total volume, total amount, mass balance or another relevant quantity.

This case teaches a general principle:

A derived quantity can change because its numerator changed, its denominator changed, or both changed.

Whenever two teams disagree about a derived metric—density, concentration, rate, efficiency, incidence—open the ratio.

23. Teach the disagreement without teaching children to distrust everything

A poor lesson from scientific disagreement is: “Scientists are always changing their minds, so nobody really knows anything.”

An equally poor lesson is: “The textbook answer is correct, so disagreement means one group was careless.”

A better lesson is that confidence has structure.

Some results are highly stable across many independent methods. Others are preliminary. Some disagreements come from different definitions. Others reveal genuine uncertainty. Some are resolved by better measurement. Others require new theory.

Teach pupils to ask:

• What do both groups agree on?
• What exactly differs?
• Is the disagreement about data, method, model or interpretation?
• What evidence would change either conclusion?
• How strong should the wording be right now?

This creates calibrated trust rather than cynicism.

Science is trustworthy not because disagreement never occurs, but because disagreement can be localised, tested and revised.

24. Write a result that another reader can safely use

A good scientific conclusion should not force the next reader to reverse-engineer hidden assumptions.

Instead of:

Method A is better.

write:

In the supplied easy and difficult task categories, Method A had the higher observed success rate within both categories. The pooled overall rate favoured Method B because the methods received very different task mixtures. The data do not establish which method would perform better under a common target mixture without additional design information.

This conclusion is longer because it carries the conditions that make it safe to interpret.

A useful reporting structure is:

1. Observation: what was measured?
2. Comparison: against what?
3. Scope: under which conditions?
4. Uncertainty: what remains unresolved?
5. Interpretation: what does the evidence support?
6. Boundary: what does it not support?
7. Next test: what would discriminate among the remaining explanations?

Strong scientific writing is not timid. It is bounded.

25. Choose the next test by the decision it could change

When results disagree, the instinct to “collect more data” is too vague.

More data of the same kind may simply reproduce the same ambiguity.

A better next test is one whose possible outcomes separate the leading explanations.

Suppose two sensors disagree because either one is biased or one has a slower response.

Repeating the same ten-second reading does little.

Instead, compare both against an external reference over time. If they converge to the same value but at different rates, response time is implicated. If one remains offset, calibration is implicated.

Suppose treatment and shelf are confounded. Measuring more plants on the same arrangement does little. Randomise treatment across shelf positions.

Suppose missing outcomes could reverse the ranking. Recover the missing records or design a follow-up that measures the failure-prone cases.

Suppose two models fit old data but predict different responses under a new condition. Test that condition.

The general decision card is:

Our decision concerns ___.

The evidence currently establishes ___.

The main unresolved alternative is ___.

The next check is ___ because its possible outcomes would distinguish ___.

We will preserve the original records and report any changes by ___.

We will not claim ___ unless additional evidence establishes ___.

This is how disagreement becomes productive.

The goal is not to make one group surrender. It is to move the boundary of what can be said responsibly.

Fast Diagnostic Table

If the disagreement looks like…	Check first	Useful next test
Same word, different values	Measurand, unit, denominator	Rewrite both claims as full quantity definitions
Two instruments differ	Reference, range, response time	Cross both instruments against an independent reference
Treatment results differ	Assignment, confounding, experimental unit	Break the alignment with randomisation or blocking
One graph looks dramatic	Axis, transformation, smoothing	Replot the same raw data under transparent alternatives
One result is “not detected”	Detection limits and preparation	Compare methods at known low concentrations or with appropriate controls
Pooled ranking reverses	Subgroup counts and weights	Define a target mixture before standardising
One point changes the conclusion	Raw record and exclusion rule	Investigate provenance; show sensitivity with and without it
Model predicts outside observed range	Domain and assumptions	Bring the case into validated range or collect boundary data
Several papers agree	Evidence ancestry	Seek genuinely independent data or methods
Same raw data, different conclusion	Analysis pipeline	Replay each calculation step from raw records

Misconceptions Worth Hunting

• “One group must be wrong.” The groups may have measured different quantities or asked different questions.
• “More repeats solve disagreement.” Repeating a confounded or misdefined procedure can reproduce the same problem.
• “The tidier graph is more trustworthy.” Visual smoothness is not evidence of correctness.
• “The newer instrument is right.” Instrument age is not a calibration argument.
• “Not detected means none exists.” It means the signal was not detected under that method and criterion.
• “Ten readings mean ten experiments.” Independence depends on how treatment or sampling was generated.
• “An outlier should be removed.” Exclusion requires a defensible reason independent of the desired conclusion.
• “Correlation proves methods agree.” Methods can correlate perfectly while differing systematically.
• “Five papers mean five independent confirmations.” They may share data, samples, methods or assumptions.
• “More data is always the next step.” The next useful observation is the one that can discriminate among remaining explanations.

Transfer Check

Question 1 — Two reports say “growth” but one uses height and the other dry mass. Do they necessarily disagree?

No. They have different operational definitions. Compare like with like before interpreting the difference.

Question 2 — A result looks like a transcription mistake because it ruins the trend. Can you correct it?

Not on that basis alone. Preserve the original record and seek independent evidence for a correction.

Question 3 — Ten seedlings share one treated tray. Are there ten independent treatment replicates?

Not necessarily. The independently treated unit may be the tray.

Question 4 — Two sensors disagree at 10 seconds but agree at 2 minutes. What is a plausible explanation?

Different response times rather than different equilibrium values.

Question 5 — One laboratory detects a substance and another does not. Must one be wrong?

No. Different detection limits, preparation or selectivity can produce compatible reports.

Question 6 — A pooled percentage reverses both subgroup comparisons. What should you inspect?

The distribution of observations across subgroups and the weights implicit in the pooled average.

Question 7 — Two methods have correlation 1.0, but one always reports twice the other. Do they agree?

No. Correlation measures co-variation, not numerical agreement.

Question 8 — What makes a next experiment informative?

Its possible outcomes would distinguish among serious remaining explanations or change the decision.

Continue through the eduKate science library

Use these routes when one missing step—not the entire casebook—is the next useful lesson.

• How Scientific Measurement Works — quantities, sensors, calibration, sampling and uncertainty.
• How Scientific Experiments Work — variables, controls, randomisation, replication and experimental units.
• How Scientific Models Work — assumptions, parameterisation, prediction and boundary failure.
• How Scientific Evidence Works — claims, alternatives, causal inference, replication and confidence.
• How Scientific Explanations Work — mechanisms, causes, scale and competing explanations.
• Science Tuition & Learning Hub — main Yishun science entrance.
• Science Directory — broader navigation through the science estate.
• Primary 4 Science Experiments in Yishun — school-level variables, data and clear answers.
• Primary 5 Science Open-Ended Questions in Yishun — evidence-to-explanation practice.
• eduKate Sengkang: Missing Data, Blanks, Zeros and Exclusions — focused companion to Chapter 13.
• eduKate Sengkang: No Change, Null Results and What You Can Conclude — focused companion for null-result reasoning.
• eduKate Sengkang Science Hub — broader connected science library.

These links are intentionally complementary. This casebook owns disagreement diagnosis; the linked works own the deeper learning jobs they were built to teach.

Source Notes and Boundaries

The methodological distinctions in this article align with established measurement and scientific-method practice, including the Joint Committee for Guides in Metrology vocabulary for measurands, calibration and detection, NIST statistical guidance on experimental design and model checking, and National Academies discussions of reproducibility and replicability.

The numerical examples are original fictional teaching constructions. They are designed to expose reasoning structure, not to provide universal laboratory acceptance thresholds or professional safety protocols.

The Quiet Ending

Two careful investigators may finish with different levels of confidence.

That is not automatically failure.

Ask whether they now agree on the observations, the assumptions and the test that would change their minds.

A disagreement with those features is much more productive than a contest of preferred conclusions.

A strong scientific reader can say:

This part is arithmetic. This part is observation. This part is a conditional explanation. This part still needs a test.

That sentence is not a retreat from knowledge.

It is a map of where knowledge has actually reached.

The next useful test begins at the edge of that map.