An error is a departure from an intended, correct, valid or expected state.
A calculation mistake, mislabeled file, false measurement, wrong assumption, software bug, incorrect diagnosis, broken link, misclassification and failed handoff can all be errors. They differ in source, type, detectability, severity, propagation, reversibility and what must change to prevent recurrence.
A useful error classification frame
- Source: human, machine, process, data, model, environment, communication?
- Type: omission, commission, calculation, classification, timing, interpretation, execution?
- Detection: self-detected, automated, peer review, user report, downstream failure?
- Severity: trivial, material, critical, catastrophic?
- Propagation: local, downstream, systemic?
- Reversibility: easy to correct, costly to reverse, irreversible?
- Recovery: correction, rollback, replacement, compensation, redesign?
- Evidence: what proves the error occurred?
- Ownership: who is responsible for repair and prevention?
- Recurrence: isolated, repeated, systemic?
This article extends How to Test a Classification System by treating errors themselves as structured objects for diagnosis and learning.
Error is not the same as failure
An error is a deviation that may or may not cause failure. A failure occurs when required function or outcome is not achieved.
Error is not the same as uncertainty
Uncertainty means the correct state is not fully known. Error means an outcome or representation is wrong relative to an applicable standard or reality.
Error is not the same as disagreement
Two people can disagree while both positions remain plausible. Classification should not declare error until the relevant standard or evidence supports that judgement.
Omission errors leave something required out
A missing field, skipped safety check or absent citation can create failure without any explicitly wrong content being added.
Commission errors add something wrong
An incorrect value, duplicate publication or invalid action is present when it should not be.
Calculation errors distort quantitative results
Arithmetic mistakes, formula errors, unit mismatches and rounding mistakes should remain distinguishable because the repair differs.
Measurement errors distort observation
Calibration, sampling, instrument and recording problems can make a measurement inaccurate even when later calculations are correct.
Classification errors assign the wrong category
False positives, false negatives, boundary mistakes and open-set failures are different classification error types.
Interpretation errors assign the wrong meaning
The observation may be accurate while the conclusion drawn from it is not.
Timing errors occur at the wrong moment
An action can be correct in content but harmful because it was early, late or performed in the wrong sequence.
Execution errors misapply an otherwise valid plan
The method or decision may be sound while implementation departs from it.
Design errors are embedded upstream
If the rule, method, interface or architecture itself is wrong, perfect execution can still produce bad outcomes.
Data errors propagate silently
Incorrect, duplicated, stale, missing or mismatched records can affect analytics, models and downstream decisions.
Model errors can be structural or parametric
A model can use the wrong form, wrong assumptions, wrong parameters or wrong domain of validity.
Communication errors alter shared state
Ambiguous instructions, missing handoffs and inconsistent definitions can cause several people to act correctly on different understandings.
Human error is often a system interaction
Fatigue, interface design, workload, unclear procedures and incentives can shape mistakes. Labeling everything “human error” can hide preventable system causes.
Machine error needs mechanism
Hardware fault, software defect, model misclassification and sensor failure should not be collapsed into one machine category.
Detected errors differ from latent errors
A latent error exists without yet being observed. Its risk depends on how long it can remain hidden and how far it can propagate.
Near misses reveal contained errors
An error may occur but be caught before harm. Near misses are valuable because they reveal weaknesses without requiring a damaging outcome.
Detection route matters
Self-checks, automated tests, independent review, user reports and downstream alarms reveal different strengths and gaps in the control system.
Severity should be based on consequence
A typographical error and a safety-critical wrong instruction are not equivalent simply because both are mistakes.
Likelihood and severity should remain separate
A rare catastrophic error and a frequent minor error need different controls.
Propagation determines blast radius
Some errors remain local. Others enter templates, datasets, models or published canonicals and multiply across systems.
Upstream errors deserve priority
Correcting the source is often more effective than repeatedly repairing downstream manifestations.
Reversible errors allow rollback
Versioned files, staged deployments and backups reduce correction cost when a mistake is found.
Irreversible errors require stronger prevention
Destructive actions, safety harm and permanent loss justify stricter review before execution.
Correction is not prevention
Repair restores the current state. Prevention changes the method, system or control so the same defect is less likely to recur.
Root-cause analysis should avoid single-cause theatre
Complex failures often involve interacting conditions. Preserve proximate trigger, contributing causes, latent conditions and control failures separately.
Repeated errors indicate system weakness
If the same class of mistake returns, the earlier correction may have treated the symptom rather than the mechanism.
Error learning needs receipts
Record what happened, what was expected, how the mismatch was detected, what was repaired and what regression test was added.
AI errors need typed diagnosis
Hallucination, stale knowledge, retrieval failure, reasoning error, tool misuse, instruction conflict and formatting failure require different remedies.
A practical error record
- error ID;
- expected state;
- observed state;
- error type;
- source;
- detection route;
- severity;
- affected scope;
- propagation path;
- reversibility;
- immediate correction;
- root and contributing causes;
- owner;
- prevention action;
- regression test;
- version.
The deeper idea
An error is not merely something to erase. Properly classified, it is information about where a representation, process or system diverged from reality or intent.
To categorise an error well is to preserve what went wrong, where it originated, how far it travelled, how it was detected, how much it mattered and what changed so it is less likely to return.
Final answer
Categorise errors by source, type, detection, severity, propagation, reversibility, recovery, evidence, ownership and recurrence. Keep error separate from uncertainty and failure, distinguish upstream causes from downstream symptoms, and pair material corrections with prevention and regression checks.