Laboratory safety, reproducibility, and replication

Start with the task and its hazards

A laboratory task begins with a question, a proposed activity and the people it can affect. A hazard is a source or situation with potential to cause harm. Risk concerns the likelihood and severity of harm in the actual circumstances, including existing controls. CERN's risk-assessment guideline makes this distinction. A moving apparatus, its possible contact with a person and the circumstances of that contact are different parts of the assessment.

In our original planning case, a class wants to calculate a cart's average velocity from an already supplied record. That question requires analysis of the record; it does not require launching another cart. A different question—whether a newly measured run gives a consistent result—would require new data and its own approved plan. Changing the activity changes what needs assessment.

Every numerical record below is fictional and supplied as exact teaching input. No physical trial has been performed for this lesson. The planning cases identify missing information and decisions; they do not certify an apparatus or provide a universal operating limit.

Choose controls before starting

NIOSH's hierarchy of controls orders five approaches: elimination, substitution, engineering controls, administrative controls and personal protective equipment, or PPE. The table states their planning roles. A substitute needs assessment too; being different does not establish that it is safer.

For the record-analysis question, avoiding a new cart launch removes exposure to that proposed launch. It does not establish that every hazard of the analysis activity has disappeared, and it cannot supply new experimental evidence. If a physical trial is required, the responsible local staff must assess the actual setup and controls.

Controls need to work in practice. A reminder does not replace an appropriate barrier, and PPE does not establish acceptable remaining risk. The hierarchy assigns no numerical safety guarantee to our apparatus.

Check readiness and reassess changes

Before a real trial, use the institution's approved procedure for that activity and the required training and supervision. Know whom to contact and the locally taught emergency arrangements. Japan's JSPS research-integrity guide advises planning with experienced research and technical staff and checking institutional safety rules. Reading this lesson cannot replace that preparation.

In the planning case, the proposed cart setup has been modified, its protective barrier is damaged, and the learners do not know its approved procedure. Those facts are reasons not to start the proposed operation. Report them to the responsible teacher or safety staff for reassessment; do not improvise a repair or override protection to obtain a matching result.

A previous successful run does not establish that a changed apparatus is ready. A different attachment, location or person can alter the activity's hazards or controls. A replication plan must accommodate the locally approved safety requirements and document relevant method changes.

Name what is being repeated

We use the National Academies' stated convention: computational reproducibility concerns consistent results from the same input data and computational methods, code and analysis conditions. A replication attempt addresses the same scientific question with newly obtained data. Terminology varies between communities, so declare the convention.

These uses differ from metrology's vocabulary. Measurement repeatability is precision under repeatability conditions, including the same procedure, operators, measuring system and location over a short period. Measurement reproducibility is precision under reproducibility conditions that include different locations, operators and measuring systems; state what changed and what stayed the same.

Copying a result is not a new measurement. Rerunning an analysis can check the reported computation, but cannot by itself check the original instrument or establish a physical law. Likewise, two runs alone do not define a complete precision or uncertainty assessment.

Recalculate the same supplied record

Define the cart quantity before calculating. Track one marked point on the cart, use one fixed position axis with positive x in the declared forward direction, and retain the two time labels. Define the finite interval-average velocity as v̄ = (x₁ − x₀)/(t₁ − t₀), for t₁ > t₀. The numerator is signed displacement and the denominator is elapsed time.

Record A supplies x₀ = 200 mm at t₀ = 0 s and x₁ = 800 mm at t₁ = 2 s. Because 1000 mm = 1 m, the displacement is 600 mm = 0.600 m. Thus v̄_A = 0.600 m / 2 s = 0.300 m/s. These digits are exact input conventions in our example, not an instrument's claimed precision.

Label this displayed method CART-AVG-v1. A second analyst can use the same saved input, unit conversion and method to check the result. Preserve the original record and record any correction separately. The two endpoints do not establish constant velocity or the total path length; the computed quantity is an interval average, not an instantaneous velocity or an average speed inferred from a path.

Compare new data under a declared rule

To illustrate a comparison, supply two additional fictional records B and C for the same marked-point quantity, fixed axis and 2 s interval. Their initial positions differ and remain explicit; we do not assume that this makes all physical conditions equivalent. The table gives the exact arithmetic. These labels describe teaching inputs, not actual replication studies.

Before examining B and C, set this original local criterion: call the interval-average velocities consistent for this exercise when |v̄_new − v̄_A| ≤ 0.020 m/s. The vertical bars mean the nonnegative magnitude of the difference; the symbol ≤ includes equality. This threshold is a teaching choice, not a measurement uncertainty, confidence level or general rule for experimental agreement.

B differs from A by 0.010 m/s, so it passes the declared criterion. C differs by 0.030 m/s, so it fails. These comparisons concern only the stated velocity values and criterion. A real replication judgment must consider the scientific question, changed conditions and justified uncertainty information; none of that statistical evidence is supplied by this toy threshold.

Test what agreement can and cannot show

Agreement can survive a shared mistake. In a separate exact processing case, the supplied record's correct formula gives 0.300 m/s. Both analysts mistakenly multiply that result by the same dimensionless factor 1.10, giving 0.330 m/s each. Their mutual difference is zero, while each differs from the specified correct calculation by 0.030 m/s. This counterexample proves that agreement alone does not establish correctness.

The National Academies also states that a rigorously conducted and transparently reported study can fail to replicate. A mismatch is a reason to investigate; by itself it does not identify the cause or establish fraud.

Check the input identity, units, formula, processing version and actual experimental conditions. An unexplained change may matter, but do not assign it a cause without evidence. Keep mismatched results and explain any exclusions or revisions. Choosing only agreeing runs after seeing the results would answer a different question from the original comparison.

Report the attempt and its limits

State whether you recalculated existing data or obtained new data, what question was tested, which conditions were retained or changed, and how the outcome was calculated. For our records, retain the marked point, axis, times, units, processing method, local comparison criterion and all compared values. Do not describe our fictional inputs as completed experiments.

Separate what the record contains from your interpretation. JSPS's guide discusses detailed research records, identifiable corrections and distinguishing factual entries from ideas or inferences. A note saying that C failed our criterion is different from a claim that a particular instrument caused the difference.

An auditable record makes a claim easier to check; it does not make it true by itself. Report unresolved limitations, including uncertainty information that has not been evaluated. The next lesson, Scientific records, data management, and open science, develops the record's organization and stewardship in more detail.

Worked examples

Try it yourself

1. A moving apparatus could strike a nearby person. A protective barrier is damaged, and the approved procedure is unavailable. Identify the hazard, one possible harmful event and the appropriate next planning step.

   Show the answer and reasoning

   The moving apparatus is a potential source of harm; contact injuring a person is a possible harmful event. Do not start the proposed operation; report the damaged control and missing procedure for competent local review.

   Risk depends on the actual activity, people and controls. These facts do not establish a numerical likelihood or an acceptable risk. A reminder to be careful does not resolve the missing assessment.

2. A supplied exact record gives x₀ = 150 mm at 1 s and x₁ = 750 mm at 4 s on the same fixed axis. What is its finite interval-average velocity?

   Show the answer and reasoning

   v̄ = +0.200 m/s.

   Δx = 750 mm − 150 mm = +600 mm = +0.600 m and Δt = 4 s − 1 s = 3 s > 0 s. Divide to obtain +0.200 m/s. The endpoints alone do not specify the velocity at every instant.

3. Analyst P reprocesses the same saved input using a recorded processing version. Investigator Q obtains new data to address the original scientific question. Under our declared terminology, which action checks computational reproducibility and which is a replication attempt?

   Show the answer and reasoning

   P's recomputation checks computational reproducibility; Q's new-data study is a replication attempt.

   The distinction is the input data and question, not whether the investigator's name changed. Neither action automatically succeeds, and the terms must be stated because other communities can use different conventions.

4. Record D is stipulated to start at 0.240 m at 0 s and end at 0.880 m at 2 s. Does it pass the same local comparison with A = +0.300 m/s and inclusive limit 0.020 m/s?

   Show the answer and reasoning

   Yes: v̄_D = +0.320 m/s, and the difference is exactly 0.020 m/s.

   The displacement is +0.640 m, so dividing by 2 s gives +0.320 m/s. The condition uses ≤, so equality passes. This exact teaching threshold is not an uncertainty or confidence statement.

5. Two analysts both multiply the specified correct result 0.300 m/s by the erroneous dimensionless factor 1.10. What are their mutual difference and their differences from the specified calculation? Does mutual agreement prove correctness?

   Show the answer and reasoning

   Both return 0.330 m/s. Their mutual difference is 0 m/s; each is 0.030 m/s above the specified result. Mutual agreement does not prove correctness.

   The multiplier preserves the velocity unit but changes the value. A common processing error can affect both outputs identically, so a comparison between them cannot detect that error by itself.

6. Record C fails the local criterion. A proposed report keeps only B and says that replication proved the cart's law of motion. What needs correction?

   Show the answer and reasoning

   Report both B and C and the original criterion, identify the inputs as fictional, and limit the conclusion to the stated interval-average comparison. No law or completed empirical replication has been established.

   The outcome does not identify the cause of disagreement. Retain the inputs and explain any justified exclusions or corrections separately; selecting agreeing results after inspection changes the original comparison. The lesson's planning case also does not authorize further physical trials.

What to carry forward

• Assess the actual activity, people and controls; a hazard and the risk associated with it are different concepts.
• The hierarchy of controls guides prevention planning, and a changed setup can require renewed competent review.
• Declare whether repetition means recalculating the same input or studying the same question with new data.
• Measurement repeatability and measurement reproducibility are precision concepts with specified conditions, not synonyms for copying an analysis.
• Preserve units, tracked points, time intervals, processing methods and comparison criteria when checking a result.
• Agreement, transparent records and approved procedures each support a particular check; none alone guarantees a true conclusion or successful replication.