Why is mathematics important in thermometer calibration? A thermometer does not simply “show the temperature.” It produces an indication from a sensor response, a conversion model and a display. Calibration compares that indication with a traceable reference under stated conditions, estimates corrections and reports uncertainty. Interpolation may fill gaps between calibration points, but it cannot create information that was never measured.
This distinction matters in kitchens, laboratories, healthcare, manufacturing, weather observations and building systems. A display reading of 37.00°C looks precise, yet the true measurement uncertainty may be much larger than 0.01°C. Resolution, repeatability, calibration correction and uncertainty answer different questions.
This article builds a careful route from fixed points and reference thermometers to correction curves and uncertainty budgets. It uses worked examples but does not turn them into universal device specifications. Calibration procedures depend on thermometer type, range, environment and required accuracy.
Quick navigation
- What calibration means
- Sensor response
- Calibration points
- Corrections
- Interpolation
- Uncertainty
- Worked examples
- Student practice
- Parent guidance
- Frequently asked questions
What calibration means
Calibration establishes a relationship between an instrument’s indication and reference values with associated uncertainties under specified conditions. It does not necessarily include adjusting the instrument. A device can be calibrated, found to have a correction, and left unchanged.
If a reference temperature is 20.00°C and the test thermometer indicates 20.18°C, the indication error is +0.18°C under the stated sign convention:
error = indication − reference.
The correction to add to the indication is therefore −0.18°C. Writing the sign convention prevents a common mistake.
Calibration is not a sticker alone
A useful calibration record identifies the device, method, range, points, environmental conditions, reference standards, results and uncertainty. A date label may help manage intervals, but it does not contain the technical evidence.
Calibration also describes performance at the time and conditions of the work. Shock, drift, sensor ageing or use outside the calibrated range can change later performance.
The NIST Standard Platinum Resistance Thermometer Calibration Laboratory describes high-level calibration services and traceability for reference thermometers. Everyday instruments need not match an SPRT, but the comparison chain illustrates how measurement values are linked to standards.
Did you know?
Adjustment and calibration are different operations. If a technician changes the device to reduce error, it should generally be calibrated afterwards so its new relationship to references is known.
Temperature sensors and response functions
Different sensors turn temperature into different measurable quantities. A resistance thermometer changes electrical resistance, a thermocouple produces a voltage related to temperature differences, a thermistor has a strongly nonlinear resistance response, and a liquid-in-glass thermometer uses thermal expansion.
Let x be sensor output and T temperature. A conversion function T = f(x) maps response to temperature. Calibration tests whether the implemented function and sensor agree with references.
Linear model
Over a limited range, a sensor may be approximated by
T = ax + b.
Two known points can determine a and b. If x = 100 units at 0°C and x = 140 units at 100°C, then a = (100 − 0)/(140 − 100) = 2.5°C per unit, and b = −250°C. The large-looking intercept is simply the extrapolated line; it does not mean the sensor operates at −250°C.
Nonlinear response
Many sensors need a polynomial, logarithmic, exponential or standard reference function. A polynomial might be
T = a₀ + a₁x + a₂x².
Adding terms can reduce residuals within the fitted interval, but too many terms can follow noise and behave badly between or beyond points. Model choice requires physical knowledge and validation, not just a high number of decimal places.
Sensitivity
Sensitivity is the change in output per change in temperature, such as dR/dT for a resistance sensor. If sensitivity is small, a given output noise corresponds to a larger temperature uncertainty. If sensitivity varies across the range, one output increment does not always represent the same temperature change.
Resolution is the smallest displayed or digitised increment. It is not sensitivity, and neither one alone establishes accuracy.
Reference points, comparisons and traceability
Calibration may use physical realisations of temperature points or comparison in a controlled bath, furnace or block against a reference thermometer. The chosen method depends on range and required uncertainty.
A comparison system needs more than two thermometers placed nearby. Spatial gradients, immersion depth, stem conduction, sensor response time and stability can produce different readings even in nominally the same environment.
NIST’s industrial thermometer calibration information describes comparison calibrations and their scope. It is an authoritative example of a laboratory specifying services, ranges and methods rather than making a blanket accuracy claim.
Traceability as a chain
Metrological traceability links a result to a reference through a documented unbroken chain of calibrations, each contributing uncertainty. It is a property of a measurement result, not a magical property of a brand name.
A traceability chain might be:
- working thermometer compared with a laboratory reference;
- laboratory reference calibrated against a higher-level standard;
- higher-level result linked to a recognised temperature scale;
- uncertainties documented at each comparison.
The chain does not make uncertainty zero. It makes the relationship and uncertainty accountable.
Ice points require technique
A well-prepared ice-water mixture can provide a useful check near 0°C, but a cup containing a few floating cubes is not automatically a high-quality reference. Purity, mixture, equilibration, immersion and contact with the container matter.
Boiling water is even more conditional because boiling temperature changes with atmospheric pressure and composition. Treating “boiling equals exactly 100°C everywhere” as universal introduces a systematic error.
Correction curves and residuals
At each calibration point, compute error eᵢ = Iᵢ − Rᵢ or correction cᵢ = Rᵢ − Iᵢ. Plot correction against indication or reference temperature, state which horizontal coordinate is used, and retain units.
Suppose corrections are:
| Indication | Correction to add |
|---|---|
| 0.10°C | -0.10°C |
| 20.20°C | -0.20°C |
| 40.35°C | -0.35°C |
The growing magnitude suggests that one constant offset may be insufficient. A fitted line or piecewise interpolation could represent the observed pattern within range.
Offset and scale error
A simple model for indicated value I is
I = aT + b.
The b term is offset error; a different from 1 represents scale error. Two-point calibration can estimate both, but additional points are needed to test linearity.
If a device reads 0.2°C high at 0°C and 0.8°C high at 100°C, a constant −0.5°C correction is not appropriate across the range. A linear error model e(T) = 0.2 + 0.006T fits those endpoints, but intermediate validation is still needed.
Residuals
After fitting a correction model, residual at point i is observed correction minus fitted correction. Residual plots reveal curvature, outliers and changes in spread.
A model that has small average residual but a strong U-shaped pattern is systematically missing nonlinear behaviour. Random-looking residuals do not prove correctness, but structure is a warning.
Hysteresis
Some instruments give different indications at the same reference temperature depending on whether temperature approached from above or below. This is hysteresis. A single upward calibration sequence can miss it.
Repeat points on heating and cooling when the application and procedure require it. Do not average away a directional effect without reporting it.
Interpolation between calibration points
Linear interpolation estimates a value between two measured points. For points (x₁, y₁) and (x₂, y₂),
y = y₁ + (x − x₁)(y₂ − y₁)/(x₂ − x₁).
If correction is −0.10°C at 10°C and −0.30°C at 30°C, the interpolated correction at 18°C is
−0.10 + (18 − 10)(−0.20)/20 = −0.18°C.
This assumes the correction varies linearly enough over that interval. Calibration data or sensor physics must justify the assumption.
Interpolation versus extrapolation
Interpolation stays between measured points. Extrapolation goes beyond them and is usually riskier because behaviour outside the tested region is unknown. A polynomial that looks smooth from 0°C to 50°C may diverge rapidly at 80°C.
The safe response is not “never extrapolate” but “do not claim supported accuracy outside evidence.” If extrapolation is operationally necessary, its model basis and larger uncertainty should be explicit.
Piecewise linear interpolation
Piecewise linear correction connects adjacent calibration points. It is transparent and avoids high-order polynomial oscillations, but its slope changes abruptly at points. Dense points in regions of strong curvature can improve representation.
Calibration-point placement should follow expected sensor behaviour and use. Equal 10°C spacing is not automatically optimal if most decisions occur near 37°C or a phase transition.
Interpolation uncertainty
The uncertainty at an interpolated point is not simply the average of endpoint uncertainties. It includes endpoint uncertainty, covariance if present, model inadequacy and possible nonlinearity between points.
Reporting an interpolated correction to 0.001°C because a calculator displays it does not make the result reliable to 0.001°C.
Measurement uncertainty and budgets
Measurement uncertainty describes the dispersion of values reasonably attributable to the measurand under a stated model and conditions. It is not the same as mistake, tolerance or worst-case guarantee.
An uncertainty budget lists important input quantities, their estimates, standard uncertainties, sensitivity coefficients and contributions. Possible components in thermometer calibration include:
- reference thermometer calibration;
- bath stability and spatial uniformity;
- test thermometer repeatability;
- display resolution;
- immersion and stem conduction;
- self-heating;
- interpolation or curve fit;
- data acquisition and electrical measurement;
- hysteresis and drift where relevant.
NIST’s Uncertainty Machine provides resources for evaluating measurement uncertainty. It supports calculation, but a valid budget still depends on identifying the right model and inputs.
Type A and Type B evaluations
Type A evaluation uses statistical analysis of repeated observations. If n independent readings have sample standard deviation s, standard uncertainty of their mean may be s/√n under appropriate assumptions.
Type B evaluation uses other information such as calibration certificates, resolution limits, manufacturer data or prior knowledge. “Type B” does not mean guesswork; the source and probability model should be justified.
Combining independent standard uncertainties
For an output y depending on independent inputs, combined standard uncertainty is often approximated by root-sum-of-squares:
u_c = √(u₁² + u₂² + …).
If contributions are correlated, covariance terms are required. Adding every uncertainty arithmetically may be conservative in some contexts, but it is not the standard propagation model for independent standard uncertainties.
Expanded uncertainty
Expanded uncertainty U = ku_c uses a coverage factor k. A value near 2 is often associated with approximately 95% coverage under suitable conditions, but this is not automatic. The report should state k and the basis for coverage.
For example, a corrected result might be 25.18°C ± 0.12°C, k = 2. This communicates much more than a bare “25.18°C.”
Resolution, repeatability, tolerance and accuracy
These terms are frequently confused.
- Resolution: smallest change displayed or represented.
- Repeatability: closeness among repeated results under specified repeatability conditions.
- Error: measured indication minus reference under a sign convention.
- Correction: value applied to compensate estimated systematic error.
- Uncertainty: quantified doubt or dispersion associated with the result.
- Tolerance: permitted limit defined by a specification or decision rule.
A thermometer with 0.01°C resolution might repeat within 0.02°C but have a 0.4°C calibration error. After correction, uncertainty might be 0.08°C. None of these numbers can substitute for another.
Rounding
Report uncertainty with sensible significant digits, then round the measured value to the same decimal place. If U = 0.12°C, reporting 23.456789°C ± 0.12°C suggests unsupported detail.
Keep extra digits during intermediate calculations to reduce rounding accumulation, then round the final result.
Dynamic response and time constants
A thermometer needs time to approach the environment’s temperature. A simple first-order model is
T_s(t) = T_f + (T_0 − T_f)e^(−t/τ),
where τ is the time constant. After one τ, about 63.2% of the total change has occurred; after three τ, about 95%; after five τ, about 99.3%.
If a sensor moves from 20°C to a stable 80°C bath with τ = 10 s, at 10 s it indicates about 57.9°C in the ideal model, not 80°C. Recording too early creates a transient error.
Stirring, probe construction, immersion and contact change τ. A response-time result in flowing water does not automatically apply in still air.
Self-heating
Resistance sensors require measurement current. Electrical power P = I²R can warm the sensor above its environment. Reducing current reduces heating but may reduce electrical signal relative to noise. This is another optimisation problem.
Worked examples with units and checks
Example 1: correction sign
Reference is 50.00°C; indication is 50.24°C. Error is +0.24°C and correction is −0.24°C. Applying it gives 50.24 − 0.24 = 50.00°C.
Example 2: linear calibration
A sensor output is 2.00 V at 0°C and 3.00 V at 100°C. Slope is 100°C/V. Model is T = 100(x − 2.00). At 2.37 V, estimated T is 37°C, assuming adequate linearity.
Example 3: interpolation
Correction is +0.06°C at 20°C and +0.14°C at 40°C. At 35°C, fraction is 15/20 = 0.75. Correction is 0.06 + 0.75(0.08) = +0.12°C.
Example 4: combined uncertainty
Standard uncertainty components are 0.030, 0.040, 0.020 and 0.010°C. Root-sum-of-squares is √(0.0009 + 0.0016 + 0.0004 + 0.0001) = √0.0030 ≈ 0.0548°C.
With k = 2, expanded uncertainty is about 0.11°C.
Example 5: resolution contribution
If display rounds to 0.1°C and rounding error is modelled as uniform over ±0.05°C, standard uncertainty is 0.05/√3 ≈ 0.0289°C.
Example 6: repeated readings
Ten readings have sample standard deviation 0.06°C. Standard uncertainty of the mean is 0.06/√10 ≈ 0.019°C if independence and stability assumptions are reasonable.
Example 7: time response
For T₀ = 20°C, T_f = 60°C and τ = 8 s, after 16 s the remaining difference is e⁻² ≈ 0.1353 of 40°C. Indication is about 60 − 5.41 = 54.59°C.
Example 8: tolerance decision
A corrected result is 100.08°C with expanded uncertainty ±0.15°C. A strict acceptable interval is 99.9°C to 100.1°C. The uncertainty interval crosses both limits, so an unqualified pass based only on the central value would ignore decision risk.
Example 9: drift
Corrections at 0°C are −0.05°C in January and −0.17°C in July. Estimated six-month change is −0.12°C. Two observations do not prove linear drift, but they justify investigation and possibly a shorter interval.
Example 10: sensitivity
A resistance sensor changes 0.385 Ω/°C. Resistance standard uncertainty is 0.019 Ω. Temperature contribution is 0.019/0.385 ≈ 0.049°C under the local linear model.
Designing a calibration experiment
A calibration plan starts with the measurement question. What temperature range is used? What uncertainty is needed? Is the sensor read directly, through a transmitter or through a complete data logger? Calibrating only the sensing element may omit error from the rest of the measurement chain.
The plan should specify points, approach direction, soak time, repeats and acceptance criteria. It should also define how results will be reduced before measurements begin. Changing the model after seeing every point can encourage overfitting.
Point selection
Include endpoints and important operating points, with additional points where nonlinearity is expected. If a refrigerator operates between 2°C and 8°C, dense evidence there may be more useful than many points far away.
Calibration at exactly the future operating temperature can reduce interpolation, but environmental replication may be difficult. A multi-point curve supplies flexibility while adding modelling decisions.
Randomisation and sequence
Always increasing temperature can confound drift with temperature. An up-and-down sequence can reveal hysteresis; repeated middle points can reveal bath or instrument drift.
Randomising order is not always practical because thermal systems take time to stabilise. A balanced sequence can preserve efficiency and still expose directional effects.
Immersion and conduction
A stem conducts heat between the sensing region and surroundings. If immersion is shallow, the sensor may lie between bath and room temperatures. Increasing immersion until indication stabilises is one diagnostic, within the probe’s safe and specified limits.
Radiation from warmer or cooler surfaces also matters, especially in air. A sensor shield can reduce solar radiation error in weather measurements while allowing ventilation.
Curve fitting and validation
Least-squares fitting chooses coefficients that minimise the sum of squared residuals. For a straight line y = a + bx, the fit balances all points rather than passing exactly through two selected endpoints.
The coefficient of determination R² summarises a fraction of variation explained under the model, but a value near 1 does not guarantee a suitable calibration. A narrow x range or smooth systematic curvature can still produce a high R².
Withheld validation
Fit the model on some calibration points and test it on points not used for fitting. If training residuals are tiny but validation errors are large, the model may be overfit.
Cross-validation rotates which points are held out. It is useful for comparing model complexity, though a small physical calibration dataset needs engineering judgement too.
Weighted least squares
If points have different standard uncertainties, weights proportional to 1/u² can give more influence to better-known points. This is justified only when uncertainties and independence are credible.
Repeated points are not automatically independent. Shared reference uncertainty, bath bias and electrical calibration can create correlation. Ignoring covariance makes a fit appear more certain than it is.
Inverse calibration
A model predicting sensor output from temperature is not always algebraically or statistically equivalent to a model predicting temperature from output. The error structure determines which regression is appropriate.
For routine application, the final function must accept the quantity actually observed and propagate its uncertainty. Documenting the model direction prevents silent misuse.
Digital acquisition and quantisation
An analogue-to-digital converter maps a voltage range into discrete codes. An ideal n-bit converter has 2ⁿ codes. Across a 5 V span, a 12-bit least significant step is approximately 5/4096 = 1.22 mV.
If a sensor has sensitivity 10 mV/°C, one code represents about 0.122°C before amplification. That quantisation may dominate even if the display software prints 0.001°C.
Averaging is not unlimited resolution
Averaging independent noise can reduce standard deviation of the mean approximately as 1/√n. Averaging 100 independent readings can reduce random standard uncertainty by a factor of 10.
It does not remove systematic calibration error, drift or perfectly correlated noise. If all codes are identical because the signal never crosses a quantisation boundary, averaging identical values does not reveal hidden fractions without suitable dither or noise.
Filtering and delay
A moving average of five samples smooths noise but delays rapid changes. For a step input, the output takes several samples to settle. A filtered display may therefore look stable while lagging the process.
State the sampling interval and filter. A “response time” measured after heavy smoothing includes both sensor physics and algorithmic delay.
Comparing thermometer types
Platinum resistance thermometers offer stable, well-characterised response across useful ranges but require electrical measurement and careful construction. Thermistors can be highly sensitive over limited ranges but strongly nonlinear. Thermocouples cover wide ranges and are robust, yet measure a voltage related to a temperature difference and require reference-junction treatment.
Infrared thermometers infer surface temperature from emitted radiation. Emissivity, reflected surroundings, viewing angle, spot size and intervening atmosphere influence results. They do not measure internal body or air temperature simply by pointing.
Liquid-in-glass instruments avoid electronics but require correct immersion and careful reading to avoid parallax. Each technology has a different error model; “digital versus analogue” is not enough to rank accuracy.
Complete-system calibration
A thermocouple connected to an extension cable, cold-junction sensor, amplifier and display is a system. Calibrating only the thermocouple wire does not test every component.
Likewise, a weather station’s radiation shield and siting affect air-temperature measurement beyond the probe’s laboratory calibration. Measurement quality belongs to the entire installation.
Drift, control charts and calibration intervals
Repeated calibration results create a time series of correction. A control chart can show whether drift remains statistically stable or exhibits a jump after shock or repair.
Suppose corrections at 50°C over four calibrations are −0.05, −0.08, −0.11 and −0.14°C. The sequence suggests roughly −0.03°C per interval, but four points cannot establish a permanent linear law. Use it to manage risk, not to promise the future.
Calibration interval selection balances stability evidence, required uncertainty, use severity and consequences. A device used gently in a classroom may need a different programme from one cycled daily in an industrial process.
Intermediate checks at one or more points can detect gross drift between calibrations. A check does not replace full calibration unless the procedure and evidence support that role.
A compact result-reporting workflow
First identify the indication and the applicable correction. Second calculate the corrected value without premature rounding. Third assemble standard uncertainty contributions using a stated model, including covariance when relevant. Fourth choose and justify the coverage factor. Finally round uncertainty and result consistently.
A concise report might read: “At the 40°C comparison point, corrected indication was 39.96°C with expanded uncertainty 0.12°C, k = 2, under the stated immersion and bath conditions.” This is stronger than “the thermometer was accurate.”
Include calibration date, range and model version so a later reader knows which evidence applies. If the result supports a conformity decision, state the decision rule rather than hiding it behind the central value.
Decision rules and fitness for purpose
Calibration supports decisions: is a cold store within range, is a process stable, or is a thermometer suitable for a lesson? The required uncertainty depends on the consequence and tolerance.
A common rule of thumb seeks uncertainty much smaller than tolerance, but the appropriate ratio and guard band depend on standards, risk and context. Do not invent a universal “4:1 rule” for every application.
If accepting a result near a specification boundary could create harm, decision rules should account for uncertainty and false-accept risk. High-stakes healthcare, food safety and regulated processes need authorised procedures and qualified personnel.
Fit for purpose
A thermometer that is unsuitable for a 0.1°C laboratory comparison may be entirely useful for showing a 10°C classroom change. “Accurate” is incomplete without the range, uncertainty and intended decision.
This same reasoning appears in Why Mathematics? | Manufacturing Tolerances, Measurement and Quality Control: measurement capability must be judged against the decision being made.
Common misconceptions
- Calibration means adjustment. Calibration establishes a relationship; adjustment changes the device.
- More decimal places mean greater accuracy. They indicate resolution, not proven error or uncertainty.
- Correction and error have the same sign. Under the usual convention, correction is the negative of indication error.
- Two points prove linearity. Any two distinct points define a line; extra points test the assumption.
- Interpolation is exact. It depends on model adequacy between points.
- Extrapolation is just wider interpolation. It is less supported because it leaves the measured interval.
- Traceability means zero uncertainty. Every link contributes uncertainty.
- Repeated identical readings prove accuracy. They show repeatability at best; a shared bias can remain.
- Room temperature is uniform. Gradients, radiation and airflow can create local differences.
How students can practise
Make a correction table
Use instructor-provided fictional reference and indication data. Calculate errors and corrections, plot them, fit a line and examine residuals. State the sign convention at the top.
Compare interpolation methods
Given five nonlinear calibration points, estimate intermediate corrections using nearest point, linear interpolation and a quadratic fit. Compare predictions at withheld validation points. Discuss which method is transparent and which overfits.
Observe response safely
Place an inexpensive classroom thermometer in room-temperature and mildly warm water under adult supervision. Record readings every five seconds without using boiling water. Plot approach to equilibrium and estimate a time constant.
Build an uncertainty budget
Use a fictional certificate uncertainty, bath stability, resolution and repeatability. Convert all components to standard uncertainties and combine them. Ask which component dominates and what improvement would matter most.
Map temperature responsibly
Read Why Mathematics? | Urban Heat Islands, Spatial Interpolation and Temperature Maps. Discuss how sensor calibration error can create apparent spatial patterns and why metadata matters.
Guidance for parents and educators
Use safe ranges and equipment. Do not ask students to handle boiling liquids, dry ice, unknown chemicals or clinical decisions. Household and classroom activities should demonstrate concepts, not certify a device for regulated use.
Insist on a complete statement: value, unit, correction status and uncertainty when provided. Ask, “What is the measurand?” Air temperature, probe temperature and surface temperature can differ.
Reward students who identify limitations. Saying “linear interpolation may be inadequate because the residuals curve” is stronger reasoning than reporting many digits.
For career exploration, calibration and temperature metrology connect mathematics with electronics, physics, materials, manufacturing, climate observation and quality assurance. Mathematics is important preparation, but it does not guarantee a particular credential or role.
The NIST thermocouple calibration services show that methods are sensor- and range-specific. This is a useful model of precise professional language.
Useful next reading
- Mathematics Learning Hub for broader numeracy and modelling routes.
- Why Mathematics? | Air Conditioning, Humidity, Dew Point and Psychrometric Charts for coupled environmental measurements.
- Why Mathematics? | Manufacturing Tolerances, Measurement and Quality Control for decision limits and measurement capability.
- NIST Industrial Thermometer Calibrations for authoritative laboratory context.
Frequently asked questions
What is the simplest definition of calibration?
It is a documented comparison that establishes how an instrument’s indication relates to reference values, including relevant uncertainties and conditions.
Does a calibration certificate mean the thermometer is correct?
It reports measured performance and uncertainty under stated conditions. Suitability depends on the intended use, range, later history and decision requirement.
Why apply a correction?
Calibration can reveal a systematic indication error. Adding the corresponding correction improves the estimate, while uncertainty remains.
Can I calibrate with ice water?
A properly prepared ice point can be a useful check near 0°C, but casual ice water has uncontrolled purity, mixture, gradients and immersion. It is not automatically a traceable calibration.
Why are several calibration points needed?
They reveal scale error and nonlinearity. Two points define a line but cannot test whether the sensor follows it between them.
What is wrong with extrapolation?
The sensor may curve, saturate or change behaviour beyond the measured range. Extrapolation can be used only with an appropriate model and honest uncertainty, not assumed safe.
Why use root-sum-of-squares?
Independent standard uncertainty components combine through variance addition in a linearised model. Correlated inputs require covariance terms.
Is uncertainty the same as plus-or-minus error?
No. Error is a difference from a reference estimate. Uncertainty describes dispersion associated with the measurement result and follows a stated coverage convention.
How often should a thermometer be calibrated?
There is no universal interval. Stability history, use severity, required uncertainty, consequences and applicable procedures determine the interval.
Can calibration guarantee future readings?
No. It provides evidence at calibration time. Ongoing checks, handling controls and recalibration manage the risk of drift or damage.
Final perspective
Thermometer calibration shows why mathematics matters whenever a number supports a decision. Lines translate sensor response, interpolation connects measured points, residuals test the model and uncertainty budgets keep the final digits honest.
The transferable habit is to ask what was compared, under which conditions, with what correction and with what uncertainty. That habit turns a display reading into defensible measurement—and helps students understand that mathematical precision is not decoration but disciplined evidence.
The same habit travels well beyond temperature. Scales, pressure sensors, electrical meters and environmental stations all need response models, traceable comparisons and honest uncertainty. A student who learns to distinguish a displayed digit from supported knowledge is learning how science earns trust, one documented comparison at a time.
Good records make that trust reproducible. Another person should be able to identify the reference, reconstruct the correction, understand the uncertainty components and see where interpolation was used. If any of those pieces is missing, the result may still be useful, but its limitations should be stated rather than covered by extra decimal places.