Error discovered in your own work: what next?

Freeze the evidence before fixing anything

Directly editing the only file can destroy the clearest account of what happened. Make the affected state read-only. Preserve document history, analysis inputs, source records and any submission receipt. Open a separate working copy for investigation. Give both states distinct, dated identifiers.

Start with observations, not a diagnosis. Record when and how the issue appeared, where it occurs and which independent evidence contradicts the work. Label any suspected cause or downstream effect as provisional until tested. This avoids both minimising a serious problem and making a confession that goes beyond the established facts.

Before submission, examine the whole dependency chain

Freeze the flawed draft and correct a copy. One incorrect input may propagate through a table, chart, discussion and conclusion. A misidentified source may support several paraphrases. Search for downstream uses rather than repairing only the place where the problem first became visible.

1. Open the original evidence and confirm the discrepancy independently.
2. Inventory dependent files, formulas, references and prose.
3. Define the correction rule before recalculating.
4. Produce a new output and compare it with the frozen state.
5. Reassess central findings and limitations.
6. Link old and new releases on the correction card.
7. Run a second-person check or a reproducible automated test.

A typographical error with no change of meaning needs a different response from a reversed effect estimate. Both can still enter the version log once a draft has passed formal review. This explains why outwardly similar files differ and keeps the final package internally consistent.

After submission, follow the governing procedure

Do not assume that an online portal permits a replacement merely because an upload button remains visible. Keep the submitted file and receipt unchanged. Consult the assessment brief, regulations or portal for the named contact. Depending on context, this may be a course lead, supervisor, examinations office, editor or integrity contact.

Explain the verified problem and its currently known reach, then request instructions. Do not promise a particular outcome or impose your own deadline. Processes vary by institution, work type and stage of assessment. A published dataset or paper may also require a versioned public correction separate from an academic submission process.

Assess significance through dependencies

Distinguish presentation, input, method and inference. A decimal typo may be minor when the correct value is evident elsewhere. The same character can be decisive when it changes the direction of an effect. Classify impact using evidence rather than the visual size of the edit.

Create a chain from input data through transformation and analysis to figure, statement and abstract. Mark each node unchanged, corrected or unresolved. Rerun the complete analysis when a computational input changes. Replacing the final number by hand leaves the process broken.

An honest error is not automatically research misconduct. Mistake, negligence and deliberate conduct require different evidence. The careful misconduct decision tree provides orientation, while only the competent process can make a formal finding.

Send a factual, bounded correction notice

Identify the work, exact version, submission date and affected locations. State the verified discrepancy, known cause, impact checks, proposed scholarly correction and unresolved questions. Do not attach confidential raw data to an ordinary email without approval; ask for an appropriate transfer route.

A useful structure is: “In the version submitted on [date], I found an error on [date] at [location]. The work states [old state]; the original record supports [correct state]. My current dependency review identifies [affected items]. The central conclusion remains unchanged / requires reassessment. I have preserved the submission, error card and amended working version separately. Please advise which correction procedure applies.”

If the destination is uncertain, a research integrity ombudsperson can advise confidentially on routes. The research integrity hub gathers related evidence practices. A plagiarism-check foundation guide may help inspect source similarity, but it cannot evaluate computational reach or prescribe a correction.

Close the correction without erasing history

• Record the cause and the control that failed.
• Design a future check around that actual failure class.
• Keep the accepted correction and responsible decision together.
• Mark superseded versions while retaining them as evidence.
• Notify collaborators whose outputs depend on the error.
• Issue versioned corrections for public data or code where required.

The aim is not a flawless retrospective story. It is a transparent account of what was known, when it became known and how the scholarly statement changed. A bounded, prompt correction strengthens the inspectability of the work.

Seven-question completion test

1. Is the affected original preserved?
2. Can an independent reader locate the error?
3. Are fact, suspected cause and uncertainty separated?
4. Have all dependent results been checked?
5. Does the new version show its validation route?
6. Was the correct formal contact used?
7. Are the final decision and any public correction recorded?