Back to the on-screen lesson ·
distinguish observation, model and testable prediction
Paper packet. Every task here also exists on screen, where it is checked automatically; answers written on paper are not assessed by Nydus. When you are back at a device, enter your answers there.
You will distinguish observation, model and testable prediction, explaining the inference and its limits in a supplied case.
Several models can fit one observation. A useful test derives a prediction and identifies what result would conflict with it under stated assumptions.
| Term | What it means |
|---|---|
| Model | A representation that organizes and predicts selected features of a system. |
| Prediction | An expected observation derived from a model and conditions. |
| Falsifiable | Exposed to possible observations that would conflict with its stated commitments. |
| Corroboration | Support from surviving relevant tests, without a deductive guarantee of truth. |
Scientific inquiry connects observations with models that organize and explain them. A model can be mathematical, mechanical, conceptual, or computational. It usually simplifies the system by retaining selected features and setting others aside. A straight-line model of a spring may ignore friction and assume a limited range of loads. These simplifications do not automatically make the model useless. They define the conditions under which its predictions should be assessed.
An observation is a recorded result under a procedure. A prediction is what a model says should be observed under specified conditions. A test compares the prediction with the result. Keeping these roles separate prevents a model from being treated as merely a list of observations, or an observation from being treated as an entire explanatory theory. A measured value can support or challenge a model without uniquely determining it, as the underdetermination lesson showed.
Suppose model M states that an output equals twice the input plus one. For input three, it predicts seven. An independent measurement at that input gives nine. Under an exact version of the classroom model, with the input and measurement stipulated accurate, the result conflicts with M. The comparison is informative because the prediction was specific enough to risk failure. A model compatible with every possible output at every input would not make the same discriminating commitment.
Scientific models do not all use exact predictions. Some give probability distributions or ranges. In those cases, a single unusual result may not logically refute the model. The test must be appropriate to the type of prediction. Here we begin with exact toy models to understand the inference structure, then identify the qualifications needed for realistic cases. The toy model's simplicity is an instructional assumption rather than a claim about all science.
Another way: Falsifiability and logical asymmetry
A universal statement says that every case in a specified domain has a property. If one established member of the domain lacks that property, the universal statement is false. No finite collection of positive cases deductively establishes an unrestricted universal statement, because an unobserved exception remains possible. This asymmetry helps explain why falsification became an influential idea in philosophy of science: a theory should expose itself to observations that would count against it.
Falsifiability is a property of an account's empirical commitments. It means that some possible observation would conflict with what the account says under the stated conditions. It does not mean that the account has already been falsified, that a critic dislikes it, or that it is necessarily false. A currently successful model can be falsifiable precisely because it makes risky predictions that have not failed. The distinction between testability and actual test outcome is essential.
Consider 'all tiles produced by this process are square'. Finding one round tile demonstrably produced by the process would refute the universal claim. Finding one hundred square tiles would be consistent with it and could provide inductive support, but would not deductively prove that every future tile will be square. The quantifier and domain do the work. A round tile from another process is not a counterexample unless the universal claim includes it.
An account that is adjusted to accommodate every possible result loses this kind of risk. If a prediction succeeds, it is called confirmation; if it fails, an uncheckable exception is invented solely to preserve the account. This does not mean every revision is illegitimate. Scientific models improve through revisions. The important question is whether the revision has independent motivation and produces new commitments, or merely prevents any observation from counting against the preferred view.
Another way: Hypothesis plus auxiliaries
A prediction usually follows from a package: a central hypothesis, initial conditions, instrument assumptions, and background principles. If the prediction fails, logic tells us that this package cannot all be correct under the exact comparison. It does not automatically identify the central hypothesis as the sole culprit. An incorrect input setting or a faulty measurement can also explain the mismatch. This is why real testing requires calibration, controls, and independent checks.
Write the package as H and A, where H is the hypothesis and A collects the relevant auxiliaries. If H and A together imply observation O, and not-O is established, then not-(H and A) follows. That is not the same as not-H. To isolate H, investigators need additional support for A or further tests that distinguish the possible sources of error. The simple propositional pattern captures an important caution without reducing scientific judgment to a mechanical formula.
A toy exercise may explicitly hold A fixed to focus on testing H. Respect that stipulation. If the problem says the measurement and input are exact, do not invent a broken instrument to prevent the stipulated model from being challenged. Conversely, outside the exercise, do not silently assume every auxiliary has been settled. A clear report states which conditions were checked and which remain open. Models are evaluated in evidential contexts, not in a vacuum.
Replication and varied methods can help. If several independently calibrated instruments record the same mismatch under carefully reproduced conditions, a one-device failure becomes less plausible. If the mismatch appears only beyond a particular load, the model may have a limited domain rather than be useless everywhere. The appropriate revision follows the pattern of evidence. Protecting a model at all costs and abandoning it at the first unexplained result are both poor substitutes for locating the discrepancy.
Another way: Confirmation without overclaiming
A successful prediction can provide support, especially when it concerns a result not used to construct the model and when rivals would predict differently. Yet the inference 'if H, then O; O; therefore H' is not deductively valid. Another hypothesis may also predict O. This is the error of affirming the consequent when presented as a deduction. Scientific confirmation can still be informative as a non-deductive comparison, provided its strength and alternatives are made explicit.
Novel tests matter because an account tailored to already observed data may fit by construction. Suppose a curve is drawn through every available measurement using many adjustable parameters. Its fit alone tells us less about unobserved inputs than a successful prediction at a new input could. Complexity, independent motivation, and error estimates all matter to the comparison. The goal is not always the fewest parameters, but an account whose explanatory and predictive success survives beyond the data that suggested it.
Distinguish a model's empirical performance from a claim that it is the final complete description of reality. A model may predict accurately within a range while idealizing objects or ignoring interactions. That can make it useful and scientifically important. A later model may explain both its successes and its limits. Treating any simplification as immediate disproof misunderstands what the model claims; treating every success as proof of literal completeness makes the opposite mistake.
A good test report contains the model statement, the input and background conditions, the predicted output, the observed output, and the comparison rule. If measurement tolerance matters, state it before deciding whether the outputs conflict. An exact difference of one unit can be decisive in a stipulated exact case but irrelevant when the instrument's uncertainty is several units. The evidential conclusion should follow the rule appropriate to the actual procedure.
End by separating three claims: the tested prediction succeeded or failed; the evidence changes support for the model under stated assumptions; and further work may be needed to choose among rivals or locate a failure. This separation prevents both triumphal claims after one matching result and indiscriminate rejection after one mismatch. Scientific inference is rigorous when it makes its commitments, assumptions, and revision conditions inspectable.
Another way: Distinguish a limit from an excuse
A model may explicitly claim accuracy only within a stated range. A result outside that range does not directly contradict its bounded claim, though it can motivate a broader model. By contrast, adding a new range restriction only after every inconvenient result, without an independent reason, weakens the evidential test. Write domain restrictions before testing when possible, and document why any later restriction was introduced. This makes the model's changing commitments available for criticism.
Another way: Keep the comparison rule fixed
If a test allows a stated measurement tolerance, apply that tolerance to favorable and unfavorable outcomes alike. Changing the acceptance range only after seeing an awkward result makes the comparison asymmetric. A revised tolerance may be justified by a new calibration study, but the report must distinguish that new evidence from the original test rule. Readers should be able to reconstruct how each verdict was reached.
A museum builds an interactive device whose dial controls a row of lights. An engineer proposes the exact teaching model y = 3x + 2, where x is the dial setting and y is the number of lit elements under the specified demonstration mode. At x = 4, the model predicts fourteen lights. A verified observation shows fourteen. This result fits the model, but another model might also predict fourteen at that one setting.
The team therefore tests a new setting, x = 6. The same model predicts twenty. The verified result is nineteen. For the exact classroom comparison, the dial setting and counted output are stipulated correct, so the prediction fails. The team records the mismatch rather than modifying the prediction after seeing the result. Recording both matching and mismatching trials preserves the evidential value of the test sequence.
In an actual device investigation, the team would also check the auxiliaries: whether one light has failed, whether the demonstration mode changed, and whether the dial truly reached the intended setting. A burned-out light could explain why the underlying control rule and visible count diverge. The mismatch alone identifies a problem in the model-and-conditions package; locating the defective component requires further evidence.
Suppose an independent electrical check finds one failed light and replacing it yields twenty at setting six. That gives a motivated auxiliary revision with a new successful check. It differs from inventing an invisible exception solely because the model was favored. The case illustrates both sides of disciplined testing: models should risk observable failure, and apparent failures should be investigated with enough care to identify what actually failed. Neither automatic protection nor automatic rejection captures the reasoning required.
Falsifiable does not mean already false. A successful prediction does not deductively prove a model, and a failed prediction can implicate auxiliaries as well as the main hypothesis. State the observation rule and domain before judging a result. Otherwise the claim can move whenever the evidence becomes inconvenient.
Write the exact model.
y = 2x + 1
The equation states a testable relation rather than merely naming a trend.
Insert the specified input.
x = 3
The prediction must use the same condition as the observation.
Calculate the predicted output.
y = 2 × 3 + 1 = 7
Substitution derives a definite commitment from the model.
Compare with the verified observation.
Observed y = 7: match
The observed value equals the exact prediction in this case.
State the limited result.
The tested prediction succeeds
A matching result does not deductively eliminate every rival model.
Write the proposed relation.
y = 3x + 2
This is the hypothesis being assessed under exact conditions.
Insert the new input.
x = 6
The test uses a condition distinct from the earlier matching trial.
Compute the commitment.
y = 20
Three times six plus two gives the model's output.
Record the verified result.
Observed y = 19
The observation is stipulated accurate in the simplified test.
Evaluate the prediction.
Mismatch of 1 unit
Under exact assumptions, nineteen is incompatible with the predicted twenty.
State the central hypothesis.
H: the control rule is y = 3x + 2
The rule is one component of the prediction package.
State a relevant auxiliary.
A: every commanded light functions
Visible counts depend on this condition as well as the control rule.
Derive the packaged prediction.
H and A imply twenty visible lights at setting six
Both components are needed to connect the rule with the observation.
Record the conflict.
Nineteen lights are visible
The predicted visible count fails.
Infer only the logical result.
H and A are not both correct in this test
The failure does not by itself identify which component is wrong.
Choose a discriminating follow-up.
Independently test each light
This directly investigates the auxiliary rather than assuming the rule failed.
Use y = 4x + 1 at x = 2.
Predicted y = 9
Substitute the input into the model before examining the observed value.
The verified observed output is eleven.
Absolute difference = 2
The distance between eleven and nine measures the exact mismatch.
Interpret the test under its stipulations.
Exact model: y = 2x + 3. At x = 4, the verified observed y is 11. Produce the prediction and absolute difference. Inputs and observations are stipulated accurate.
| Response | |
|---|---|
| Predicted | |
| Difference |
The exact model is y = 2x + 1 at x = 4. The verified observed output is twelve. Complete the comparison.
Substitute the input into the model.
Predicted y = prediction
The equation supplies a definite output before the observation is compared.
Calculate the absolute mismatch.
Absolute difference = difference
Compare the observed output with the predicted output without changing either.
State the bounded test result.
The exact prediction conflicts with the observation
Locating the real cause would require examining the model and its auxiliaries.
Exact model: y = 3x + 1. At x = 5, the verified observed y is 18. Produce prediction and absolute difference under accurate-input and accurate-measurement assumptions.
Predicted output: predicted. Absolute difference from the observed output: difference.
Exact model: y = 4x + 2. At x = 3, the verified observed y is 10. Produce prediction and absolute difference, not a signed difference.
Predicted output: predicted. Absolute difference from the observed output: difference.
An exact toy model is y=4x+1. Accurate input is x=5 and verified observed y=24. Write prediction, absolute mismatch and Y/N: does this observation match the stipulated exact prediction? Use the following symbols in your response fields; these codes replace word answers and remain the same in every language. Response symbols: Y = yes; N = no.
| Your reconstruction | |
|---|---|
| Prediction | |
| Mismatch | |
| Matches |
For a museum demonstration, the stated exact control model is y = 5x + 2 lights. At verified dial setting x = 4, observers count 19 functioning lights. Produce the model prediction and its absolute mismatch with that observation; do not assume which component caused it.
Predicted output: predicted. Absolute difference from the observed output: difference.
Lesson test: one question per skill, one attempt each, no hints. Your answers are checked when you submit.
An exact toy model predicts y = 3x + 4. At input x = 6, an independent verified measurement gives 25. Produce the prediction and absolute difference; a mismatch concerns the stipulated prediction package.
Predicted output: predicted. Absolute difference from the observed output: difference.
Without rereading, explain how to distinguish observation, model and testable prediction. Give a fresh case and identify what would change your analysis.
10. Finish an exact model test, step 3
The exact prediction fails
The conclusion remains conditional on the stipulated input and observation accuracy.