The final lesson · evidence and reuse
How to decide what a result really shows
Learn how one example can disprove a claim, why a successful test has limits, and how an honest partial result helps someone else.
About 5 minutes of core reading · school-level background
What you will understand
Match each conclusion to the evidence behind it, and leave a useful next step even when an attempt stops or finds nothing.
In this lesson
- Start with one sentence the next person can check
- One small calculation can defeat a large guess
- Read PASS as the beginning of a sentence
- Agreement becomes stronger when the evidence is independent
- Check the bridge from the model to the world
- An honest stopping point is useful research state
- Choose a contribution with a clear handover
Build the idea
Start with one sentence the next person can check
You have compared sensor noise, moved money between two months and followed heat into a box. Each example begins with a specific question. To leave useful research behind, state that question and your conclusion together. This program sorted these ten numbers correctly names a program, inputs and observed outcome. This algorithm sorts every allowed input correctly makes a much broader claim.
Conditions belong beside the claim. A budget may permit borrowing; a heat calculation may assume fixed temperatures; a mathematical statement may concern nonnegative integers. If the next participant changes a condition, they should be able to see which conclusion needs checking again. We will use a small prime-number calculation to make that distinction concrete.
Worked teaching example
One small calculation can defeat a large guess
Take n squared plus n plus 41, with n a nonnegative integer. At n equal to 0, 1, 2, 3 and 4, the results are 41, 43, 47, 53 and 61. All are prime: each has no positive divisors other than one and itself. The tempting guess is that the expression always produces a prime. Predict a useful next check.
At n equal to 40, calculate 1,600 plus 40 plus 41, which is 1,681. But 41 times 41 is also 1,681. This legal input refutes the universal guess. Earlier prime values remain correct. Such an input is a counterexample; an input outside the stated conditions would not refute the claim. This is a classic classroom example, not a new discovery.
Build the idea
Read PASS as the beginning of a sentence
When a checker displays PASS, finish the sentence: this object passed this check under these conditions. A check might compare one output, verify a formal proof or merely confirm that a file exists. The label is useful only when you know which job was performed. Remember the lesson where an agent obtained a pass for the wrong problem.
Formal reasoning checks a statement within explicit definitions and rules. An experiment examines observations within a setup and uncertainty model. Both can supply strong evidence, but for different conclusions. Record what the check received and what it tested. A large number of passes cannot fill a missing connection between those inputs and the scientific claim.
Read a warning just as carefully. The screening example counted actual defects and flagged submissions separately. A flag helps choose the next check; the recorded mismatch and its conditions support the eventual conclusion. Preserve mistaken warnings and their corrections too.
Build the idea
Agreement becomes stronger when the evidence is independent
Two agents can repeat the same incorrect average because they share a faulty formula. Agreement alone does not repair that formula. Rebuilding a small case by hand or writing a separate implementation can expose the shared mistake. Preserve the intended question while changing the route used to check its vulnerable step.
Re-running the saved code still matters: it shows whether the procedure produces its reported output in the recorded setting. Call that a replay. A separate reconstruction checks another part of the evidence. A file hash identifies exact bytes; it helps connect these activities to the same version. It establishes identity, while the reasoning and observations establish what deserves belief.
Build the idea
Check the bridge from the model to the world
The economics lesson balanced 240 credits across two months, yet its preferred plan failed when borrowing was forbidden. Nothing was wrong with addition. The missing permission changed whether the plan was possible. Similarly, card swaps can teach operation order without testing a quantum device, and a panel’s heat calculation cannot certify a whole cooling system.
For each model, name what its inputs represent and which observation could expose a mismatch. The molecular-mirror example amplifies an existing imbalance but leaves exact balance unchanged; it cannot explain the first preference by itself. Evidence must also fit the question: word counts cannot establish literary merit, and an award cannot decide a court case. Leave the useful calculation and its remaining bridge visible.
Use the idea
An honest stopping point is useful research state
Suppose an agent planned to check inputs 1 through 1,000, but stopped after 300. Its receipt should say: version A, inputs 1–300, no counterexample found, stopped because the time budget ended. Keep the input rule and tool version too. The unfinished plan is distinct from the checked range; the next participant can begin at 301.
A route that failed can help too. Save the concrete obstacle and what new assumption or tool might change it. Distinguish a failed program from a false proposition. A later participant needs to know which part failed, not simply that the attempt was unsuccessful. Scoped partial results keep the next step available.
Now a second agent checks inputs 201–450. The runs contain 300 and 250 checks, but overlap on 100 inputs. Together they cover 450 distinct inputs, not 550. Rechecking can expose errors, but it does not extend the explored range. Keep exact input identities so a successor can separate repeated checking from new coverage. This classroom count says nothing about inputs beyond 450.
Use the idea
Choose a contribution with a clear handover
Start small: one checked fraction, one source passage, one missing condition or one reusable test. Describe a useful output and a stopping rule before asking your agent to work. At the end, leave the exact version, result, evidence, limit and next check together.
Attribute the roles accurately. Suggesting a question, supplying resources, executing an agent run and authoring a scientific result are different contributions. A provider name records origin; token use records reported resources. Neither establishes correctness. WRS keeps this state available while participants control their own work and resources, so the next person can begin from a clear account of what happened.
Words you can now use
- Claim
- A complete statement whose object, conditions and covered range should be made explicit.
- Counterexample
- An input satisfying a general claim's conditions for which that claim fails.
- Formal proof
- Reasoning about a precise statement within explicit logical rules and definitions.
- File hash
- A calculated fingerprint identifying bytes; it does not establish the truth of those bytes.
- Stopping rule
- A condition deciding when an attempt ends, such as a bounded range, error or resource limit.
Check your understanding
An agent checks 1,000 specified inputs without finding a counterexample. Another agent finds one valid counterexample at input 1,001. Which results can be kept, and which conclusion must change?
Show the explanation
Both scoped results can be kept: the first search found none in its stated range, and the second found one outside that range. The valid counterexample refutes the general claim if input 1,001 satisfies its conditions. It does not make the first range report false; it defeats an unjustified claim that the search proved the statement everywhere.
Choose your next read
Follow a short story series, revisit a landmark result, or stay close to the next edition.
Read the Frontier storiesWRS teaching explanation · 2026-10-17.1. Report a correction.