Short answer: Precision is the habit of defining what correct means, measuring it honestly, and publishing the error you accept. Watchmakers do this with tolerances, timing machines and certificates. Software teams do it with tests, metrics and error budgets. The tools have nothing in common. The discipline behind them is identical.
A balance wheel inside a mechanical watch swings 28,800 times an hour. Every one of those swings can be measured, logged and nudged back towards a standard. A test suite does something that looks nothing like it, and works on the same principle. It checks a defined behaviour on demand, then returns a number instead of an opinion.
Precision is often mistaken for fussiness, or for an expensive kind of vanity. It is neither. Precision is what turns a promise into something a stranger can check for themselves.
This article sets out that shared discipline. It defines the words and clears away three beliefs that get in the way. It then maps the workbench onto the build pipeline, and asks the awkward question. When does tightening a tolerance stop being worth the money?
What precision actually means
Everyday speech treats accuracy and precision as one idea. Measurement science splits them apart, and that split holds up everything that follows.
Trueness, precision and accuracy
Trueness describes how close the average of many results sits to the true value. Precision describes how tightly those results cluster together, wherever their average happens to land. The international standard ISO 5725 treats accuracy as both qualities at once, which is why careful writing keeps all three words separate.
A watch that gains exactly four seconds a day is precise but not true. That is the easy fault. One move of the regulator shifts the whole cluster onto the target. A steady offset of that kind is called bias, or systematic error. A watch that gains four seconds today and loses six tomorrow is a different problem, and no single adjustment will rescue it.
A repeatable error is a problem you fix with one adjustment. A random error is a problem you have to hunt.
This is why national laboratories build a timescale from many clocks rather than one. Bodies such as the Time and Frequency Division at NIST compare imperfect clocks, describe how each one drifts, then steer the whole ensemble. Nobody owns a perfect clock. They own a very well understood set of imperfect ones.
Resolution is not accuracy
Resolution is the smallest step an instrument can report. Accuracy is whether that report is right, and the two are easy to confuse. A stopwatch reading to a hundredth of a second still waits on a human thumb. Human reaction time runs to roughly two tenths of a second. The extra digits are real. The extra truth is not.
Software repeats the mistake daily. An average printed to four decimal places has not become more truthful, only more confident. Fluent, assured language plays the same trick with words instead of digits. Our guide to why chatbots make things up takes that mechanism apart.
All of this needs a fixed reference at the bottom. Since 1967 the SI second has been defined by the caesium-133 atom, as 9,192,631,770 cycles of the radiation between two of its energy levels. The International Bureau of Weights and Measures keeps that definition for every laboratory on earth. Every other timing claim is measured against it.
Three myths that get in the way
Arguments about quality tend to stall on the same three beliefs. Each one sounds sensible, and each one is wrong.
Myth one: precision means zero error
No serious maker claims zero, and no standard asks for it. A movement certified to ISO 3159 must hold a mean daily rate between minus four and plus six seconds during testing. That is an envelope, published in advance, with the test conditions attached. Software borrows the same idea and calls it an error budget. The certificate earns trust precisely because it admits a limit.
Myth two: tighter is always better
Cost does not rise in a straight line as a tolerance narrows. It curves upward, and steeply. Halving an allowance can mean new tooling, slower cutting, more inspection and a higher scrap rate. Our guide to how precision engineering controls quality follows that curve in detail. A tolerance should be as tight as the function needs and no tighter.
Myth three: precision belongs to machines
Machines repeat. They do not judge. Someone still has to decide which dimension matters. Someone has to rule on a part that sits just outside the line, and on two acceptable parts that behave badly together. That judgement is craft, and it passes between people rather than through documents. We press the point in why craftsmanship still matters in the age of AI.
The bench and the build pipeline
A mechanical movement and a codebase share one hard problem. Both are complicated things that must behave predictably in conditions their makers do not control. Watchmaking met that problem first, and its answers translate with very little loss. For the mechanism itself, our walkthrough of how a mechanical watch movement works follows the chain from mainspring to escapement.
| At the bench | In a software team | What it controls |
|---|---|---|
| A stated tolerance on every part | Interface contracts, schemas and types | What each stage may assume about the last |
| Timing machine readings in several positions | An automated test suite run on every change | Behaviour measured against a defined standard |
| Re-timing a movement after a service | Regression tests after a fix | Damage caused by the repair itself |
| Final inspection before the case is closed | Continuous integration before a merge | Defects caught while they are still cheap |
| Serial numbers and the service record | Version control history | Which parts went into which artefact |
| A chronometer certificate from an outside laboratory | Third-party audits and published benchmarks | Claims that someone with no stake has checked |
The last row carries the most weight. A chronometer certificate beats a maker's own bench for one reason only: the person holding the stopwatch has nothing to gain from the sale. Software has started to copy the idea through security audits and shared benchmarks. It still has no body with a century of enforcement behind it.
Where the comparison breaks down
Two differences are worth holding on to. A movement wears out. Pivots polish themselves thin, oils dry, and a service restores the watch to a known state. Code has no friction, so it decays differently. It rots because the world around it moves, as dependencies update, formats change and the incoming data stops resembling the data it was built for.
The second difference is copies. A watchmaker regulates one watch for one wrist. A software team ships one artefact to millions of machines it will never see. That makes testing under real conditions harder and more important at the same time.
When precision stops paying
Every quality discipline meets diminishing returns, and pretending otherwise burns money. The last tenth of a point usually costs more than the first ninety. The skill lies in knowing where on that curve your work sits.
Speech recognition shows the shape clearly. Pushing a word error rate from 20 per cent down to 10 is transformative, because the transcript becomes readable prose. Pushing it from 6 per cent to 5 changes far less, since a human still reads the whole file to catch names and figures. Our explainer on what word error rate really means shows how much the headline number hides.
Four questions size a tolerance honestly, whether the artefact is a pivot, a pipeline or a report.
- Who is harmed by the error? A dive computer and a desk clock deserve different budgets, because the consequences are not comparable.
- Does the error accumulate? Two seconds a day becomes twelve minutes a year. Errors that stack need tighter limits than errors that cancel out.
- Can the user correct it? A quick-set date takes three seconds to fix. A silently wrong figure in a report may never be caught.
- What does the next step cost? If the next tenth doubles the price, the honest move is often to stop and say so plainly.
Tip: Write the conditions beside every number you publish. A rate of plus two seconds a day means little without the positions and temperature behind it. A benchmark score means as little without its hardware and its test data.
How to judge a precision claim
Most of us meet precision as a claim rather than a process. Five questions separate the tested from the merely stated, and they work on a watch, a laptop or an AI service.
- Is there a number? Words such as premium, professional and highly accurate cannot be checked by anyone. A figure can.
- Are the conditions stated? Every real measurement carries positions, temperatures, sample sizes or test data with it.
- Who did the measuring? Independent testing outranks a maker's own bench, which is the entire point of the Swiss Made label and the certifications above it.
- What happens outside the envelope? Good engineering describes how a thing fails, not only how well it performs.
- When was it last checked? Calibration expires, benchmarks age, and a movement drifts between services.
Those five questions are the working core of every standards body worth trusting. Our quality standards topic page gathers the articles that go deeper into each of them.
Practising precision on ordinary work
None of this needs a laboratory. The habits scale down to spreadsheets, invoices and email, and they are mostly about sequence.
Measure before you adjust. Change one thing at a time, then measure again. Write down what correct means before you start, so that finished becomes a test rather than a feeling. Keep the record, because a fault you can reproduce is already half repaired.
Then treat every failure as information rather than embarrassment. A watchmaker reads a timing trace to find where the fault sits, and the trace accuses nobody. Teams that run blameless incident reviews have rebuilt the same custom, and they get the same reward. People report problems early, while the problems are still small.
Tip: Keep a two-line log of what you changed and what the measurement did afterwards. It costs a few seconds and turns guesswork into evidence inside a fortnight.
Key takeaways
- The words are not synonyms: trueness is closeness to the target, precision is repeatability, and ISO 5725 calls the pair accuracy.
- Repeatable errors are the good kind: a consistent daily gain is one adjustment away from correct.
- Resolution flatters: more digits or more fluent prose can hide the same amount of underlying error.
- Tolerances have a price curve: tighten only as far as the function needs, then publish the limit you settled on.
- Claims need conditions: a number without its test conditions and its measurer is marketing, not measurement.
Precision, at bottom, is a form of respect for the person on the other side of the work. A tolerance is a promise about a part. A certificate is a promise about a product. A published metric is a promise about a service. Each one invites a stranger to check, and accepts the consequences if the check fails.
That is why the philosophy moves so easily between trades that share no materials. It is older than either of them, as our short history of measuring time shows, and it will outlive whatever replaces them. The rest of the argument waits in the Craft & Precision hub.
