A researcher opens a dataset and sees an unexpected pattern. It is interesting. It may even be important. The problem begins only when the reader is later told that this was the pattern the study set out to test all along.

Science needs both prediction and discovery. It becomes harder to evaluate when the boundary between them disappears.

Preregistration creates a time-stamped record of a study plan before the relevant outcomes or analyses can influence those decisions. Registered Reports move peer review earlier: journals evaluate the research question and proposed method before results are known, and an accepted Stage 1 protocol can receive in-principle acceptance that is not contingent on obtaining a striking result.

Neither system makes a study automatically correct. A preregistered bad design remains a bad design. A Registered Report can still face measurement problems, implementation failures or unexpected evidence. The value lies in making the sequence of decisions more visible and changing some publication incentives.

This article concerns research transparency across fields. It is not guidance for regulatory clinical-trial registration or any jurisdiction-specific legal requirement.

Reading route: Start with the problem preregistration is designed to expose, examine what a useful plan contains, then distinguish deviation from misconduct, explore Registered Reports, and finish with the audit trail a reader should expect.

The problem is not exploration; it is hidden timing

Exploratory analysis is how researchers discover patterns, generate hypotheses and notice that an original model was incomplete. It is an essential part of research.

Confirmatory analysis asks a different question: did a prediction or decision rule specified without seeing the relevant result survive contact with the data? The evidential meaning changes when the hypothesis was selected after the pattern was already visible.

Suppose a dataset contains twenty plausible outcomes, five subgroup definitions and several time windows. If analysts try many combinations and publish only the one that looks strongest, the final test no longer has the simple interpretation it would have had if that combination were fixed in advance.

The problem is sometimes described through practices such as undisclosed outcome switching, selective reporting, p-hacking or HARKing—hypothesising after results are known. These labels can be useful, but the deeper issue is decision provenance: what was decided before seeing which information?

A time stamp creates chronology, not truth

OSF describes preregistration as posting a time-stamped, read-only version of a study plan before beginning the relevant data collection or analysis. The record helps later readers compare intentions with what was eventually done.

The time stamp does not validate the plan. It does not prove the researcher followed it. It does not guarantee that every important decision was specified. It establishes that a particular document existed in a particular state by a particular time.

This is why the registration must be read, not merely counted. A one-line registration saying “we will analyse the data appropriately” contains little constraint. A detailed plan that identifies outcomes, exclusions and decision rules creates a much more informative comparison.

What a useful preregistration can specify

• Research question and hypotheses: what is being predicted or estimated?
• Study population and sampling: who or what can enter the analysis?
• Primary and secondary outcomes: which variables carry the central claims?
• Exposures, treatments or predictors: how are they defined?
• Exclusion rules: what will be removed and under which conditions?
• Transformations and coding: how will variables be constructed?
• Statistical model: what analysis will answer the question?
• Decision rules: what happens if assumptions fail or data quality is insufficient?
• Sample-size or stopping logic: when does data collection end?
• Planned sensitivity analyses: which assumptions will be challenged?

OSF’s current preregistration guidance explicitly encourages precise hypotheses, variables, exclusion rules, model forms and contingency plans. The objective is not ceremonial paperwork. It is to move consequential analytical choices to a point where the observed result cannot silently determine them.

A decision tree is often better than pretending every future problem is predictable

Research encounters contingencies. A variable may have an unexpected distribution. An instrument can fail. Recruitment can be lower than expected. A planned model may not converge.

A rigid plan that ignores every possible failure can encourage either bad analysis or undisclosed improvisation. A stronger preregistration can specify conditional routes: if a measurement has more than a stated amount of missingness, use this prespecified fallback; if a model fails a defined diagnostic, report the original attempt and use this alternative.

The goal is not to predict every event. It is to identify decisions likely to become outcome-dependent if left entirely open.

Preregistration is most informative when relevant data are genuinely unseen

The cleanest case is a plan registered before data collection begins. Secondary-data research is more complicated because the dataset may already exist. A researcher may know broad distributions, have seen earlier releases or be working within a team where some information is already familiar.

The solution is not to pretend ignorance. State what is already known, which variables have been inspected, whether analysts have seen the outcomes, and what information remains hidden at the time of registration.

A registration made after results are known can still document a study, but it should not be presented as prospective preregistration. Chronology is part of the evidential meaning.

Exploratory work can be registered too

Researchers sometimes avoid preregistration because they assume it requires one narrow hypothesis and forbids discovery. That is unnecessary.

A study can preregister an exploratory objective, a search strategy or a model-development process. The important distinction is that the final report should not convert data-driven discoveries into apparently prespecified confirmatory tests.

An exploratory result can be valuable precisely because it generates a new question for an independent dataset. Honest labels strengthen the research chain; they do not demote discovery.

A preregistration can be too vague to constrain anything

“We will test whether X relates to Y using suitable statistics” leaves open the sample, transformation, covariates, model family, direction, outcome version and exclusion decisions. Almost any later analysis could be described as compliant.

Vagueness is especially dangerous when many alternatives produce materially different results. The registration should be most specific where researcher discretion can change the claim.

Specificity is not maximal detail for its own sake. The useful level is enough information for an informed reader to distinguish the planned test from a materially different test.

A preregistration can also be overengineered

A hundred-page protocol can create an illusion of control while burying the decisions that matter. Researchers may specify trivial formatting choices and leave the primary outcome ambiguous.

Wintour’s editorial rule applies here too: information must pay rent. Prioritise decisions that affect interpretation, bias, flexibility and the ability to distinguish prediction from exploration.

A concise decision table can sometimes be more auditable than a long narrative because it maps each contingency to an action.

Deviation is not automatically a research failure

A preregistered plan can turn out to be inappropriate. Following it mechanically may be worse than changing it.

Suppose a planned analysis assumes a variable was recorded in seconds, but the received dataset is in milliseconds. Correcting the unit is not a violation of scientific integrity. Suppose a prespecified model fails because a software implementation contains a verified bug. Repair is necessary.

The integrity requirement is transparency: identify the deviation, explain why it occurred, state when the decision was made and distinguish the revised analysis from the original plan.

OSF’s current registration system allows updates to the research record while preserving time-stamped history. The point is not to rewrite the past; it is to show how the project changed.

A deviation log turns adaptation into evidence

Field	Question
Planned state	What did the registration say?
Observed problem	What made the plan unsuitable or impossible?
Timing	Was the relevant outcome already visible?
New decision	What changed?
Reason	Why is the new route better?
Consequence	Does the change alter the interpretation or confirmatory status?

A visible deviation can increase trust because it lets readers evaluate the decision. An invisible deviation forces them to assume the published method was always the plan.

Outcome switching is not merely renaming a heading

Imagine a study specifies reading comprehension as the primary outcome but later highlights vocabulary because comprehension shows little change. Vocabulary may still be an interesting finding. The evidential problem arises if the report presents vocabulary as though it had always carried the primary confirmatory role.

The repair is not to suppress vocabulary. Report both: the prespecified primary result and the exploratory or secondary finding, with the chronology intact.

This is one reason preregistration and complete reporting work together. Registration without access to final outcomes can show a plan; reporting without registration can show results; the comparison between the two reveals changes.

Stopping rules belong in the research design

Repeatedly checking results and stopping once a desired threshold is crossed changes the statistical process unless the method explicitly accounts for sequential monitoring.

A preregistration can specify a fixed sample, a recruitment window, a precision target or a valid sequential design. The correct choice depends on the research question and method.

The important point is that sample size should not become an invisible function of how attractive the interim result looks.

Exclusion rules can quietly become result selectors

Researchers often need to exclude corrupted records, duplicate observations, ineligible participants or failed measurements. The problem arises when ambiguous exclusion rules are tuned after seeing which records weaken a preferred result.

Preregister objective rules where possible and preserve counts at each stage. If a new quality defect requires a new exclusion, explain it and show whether conclusions differ when the affected records are retained under an alternate defensible analysis.

This connects to Sensitivity Analysis and Robustness Checks: deviations can become explicit robustness questions rather than hidden researcher discretion.

Registered Reports move review before the results

Center for Open Science guidance describes Registered Reports as a publishing format in which study designs and analysis plans are peer reviewed before data collection. Reviewers assess the importance of the question and the quality of the proposed method.

If the protocol passes Stage 1 review, the journal can grant in-principle acceptance. The later paper is then evaluated for adherence to the approved protocol, quality of execution and appropriate interpretation rather than whether the results are positive or surprising.

This changes incentives. A carefully designed study with a null result can still become a valuable publication. A dramatic result does not rescue a weak design that would have failed Stage 1 scrutiny.

Stage 1 and Stage 2 perform different editorial jobs

Stage	Main question
Stage 1	Is the research question meaningful, and can the proposed design and analysis answer it?
In-principle acceptance	Will the journal commit to publication if the approved protocol is followed and the eventual report meets the stated conditions?
Stage 2	Was the study conducted as approved, are deviations transparent, and are the results and conclusions reported appropriately?

The exact workflow varies by journal. The table expresses the general publishing logic, not a universal contract for every Registered Reports venue.

Registered Reports reduce one publication filter, not every research bias

In-principle acceptance reduces the incentive to make publication depend on a favourable result. It does not solve biased sampling, poor measurement, implementation failure, undisclosed conflicts, underpowered subgroup claims or inappropriate generalisation.

Peer reviewers can also be wrong. Stage 1 approval means a protocol survived a particular review process, not that the future result is guaranteed to be correct.

The strength of the format is structural: the decision to publish is moved closer to the quality of the question and design, before the outcome can dominate editorial attention.

Null results become easier to interpret when the test was fixed before the data

A null finding can mean many things: no meaningful effect, insufficient precision, measurement noise, poor implementation or a mismatch between the test and the theory.

Preregistration helps by showing whether the null came from the planned test or from one of many analyses selected afterward. Registered Reports add design review before results are known.

This does not make every null result decisive. It makes the research pathway easier to inspect.

Preregistration can protect researchers from their future selves

Researchers are not neutral machines. Once a result is visible, it is difficult to remember how many alternatives originally seemed plausible. A preregistration creates an external record before hindsight reorganises the story.

This is useful even without misconduct. Human memory naturally compresses uncertainty. A method that felt arbitrary before analysis can feel inevitable after one version produces a coherent narrative.

The registration is therefore a cognitive tool as well as a transparency tool.

Preregistration should not be used as a weapon against legitimate adaptation

A research culture can misuse preregistration by treating any deviation as evidence of incompetence. That creates incentives to hide problems rather than report them.

The better norm is: planned work is labelled planned; deviations are labelled and justified; exploratory work is labelled exploratory; all materially relevant results are reported.

A plan earns trust when readers can see both adherence and intelligent correction.

Qualitative research needs a different form of planning

Qualitative research can involve iterative sampling, evolving interview guides and theory development during fieldwork. A preregistration designed for a fixed confirmatory experiment may be a poor fit.

Researchers can instead register the initial question, sampling rationale, fieldwork boundaries, reflexive position, planned analytic approach, stopping considerations and which elements are expected to evolve.

The goal remains transparency about the research process, not forcing every method into the same template. See How Qualitative Research Works for the distinct logic of qualitative inquiry.

A registration should identify the version of the materials

A plan may cite a questionnaire, code repository, stimulus set or analysis script that later changes. If the registration links only to a mutable working file, readers may be unable to reconstruct what was actually committed to.

Time-stamped registrations, archived files, persistent identifiers and versioned repositories help preserve the intended state. OSF’s registration system is designed around frozen or time-stamped research records rather than an endlessly edited page.

This connects preregistration to Persistent Identifiers and Scholarly Linking.

Embargo changes visibility, not chronology

Researchers may have legitimate reasons not to expose a protocol immediately, including intellectual-property concerns or risk of being scooped. OSF currently supports embargoed registrations while preserving the registration time.

An embargo means the public cannot inspect the content during the embargo period. It should therefore not be described as contemporaneously public. The chronological commitment and the public transparency state are different properties.

A plan cannot rescue a badly chosen outcome

If a study preregisters a measure that does not validly represent the intended construct, faithful adherence preserves the flaw. The appropriate response may be to report the planned result and explain why a revised measurement strategy is needed.

This is why preregistration belongs downstream of good research design. It records decisions; it does not replace the reasoning needed to make good decisions.

The measurement layer connects to Measurement Error and Misclassification and the Library’s construct-validity owner.

Preregistered predictions can still be weak

A prediction such as “the groups will differ” can be technically prespecified while saying little about mechanism, direction or meaningful magnitude. Strong confirmatory work benefits from precise hypotheses tied to theory and measurement.

Where a range of effects would have different practical implications, prespecify the estimand and report the estimate with uncertainty rather than treating one significance threshold as the entire conclusion.

Preregistration protects chronology. Statistical inference still requires appropriate interpretation. See Statistical Inference and Uncertainty.

A preregistered analysis can be reproduced badly

A protocol can be clear while the final code is unavailable, undocumented or dependent on a changed software environment. Another researcher may understand the intended test but be unable to recreate the computation.

Planning transparency and computational reproducibility are different controls. Strong research connects them through versioned code, data documentation, execution environments and clear analysis outputs.

The next Library owner, Reproducibility and Replication, addresses that distinction.

A registered result can still be overgeneralised

Preregistering a study of one population does not make its result automatically applicable to another. The protocol can make the source question clearer, which actually helps expose where the later transfer begins.

Use External Validity and Evidence Transfer when a result is being moved to a new population, setting, intervention version or time.

The preregistration audit trail

QUESTION
→ PLAN VERSION
→ REGISTRATION TIME
→ DATA / OUTCOME VISIBILITY AT REGISTRATION
→ APPROVED OR DECLARED ANALYSIS
→ DEVIATIONS + TIMING + REASON
→ EXPLORATORY ADDITIONS
→ FINAL CODE / DATA VERSION
→ REPORTED RESULTS
→ CLAIMS
→ CORRECTION OR FOLLOW-UP

A reader should be able to move both forward and backward through this chain. The paper should point to the registration; the registration should identify the intended analysis; deviations should explain how the path changed.

A learner’s version: write the prediction before revealing the answer

The basic idea can be taught without advanced statistics. Give students a sealed dataset or experiment outcome. Ask them to write what pattern they expect, why they expect it and what result would challenge their explanation. Then reveal the evidence.

Afterward, invite exploration. Students may notice a new pattern and propose a better explanation. The key lesson is to label the new hypothesis as something learned from this evidence and design a future test for it.

This teaches a deep research habit: discovery is not weaker because it happened after looking. It simply needs new evidence before it can claim to have predicted what it already saw.

What an honest paper reports

Identify the registration and its date. State whether data collection or outcome inspection had begun. Distinguish prespecified primary, secondary and exploratory analyses. Describe deviations and why they occurred.

Do not imply that “preregistered” means independently peer reviewed. Ordinary preregistration can be self-authored. Registered Reports add formal journal review at Stage 1, but even then the review record and journal policy matter.

Do not hide an unfavourable primary result behind an attractive exploratory result. Report both at their correct evidential status.

The Wintour principle: preserve surprise without rewriting history

A great research article needs room for surprise. A finding can change the direction of a field precisely because it was not expected.

The editorial discipline is to preserve the surprise. Do not retrofit the introduction until the unexpected result appears inevitable. Let the reader see what was predicted, what happened, what changed and what the new evidence now makes plausible.

That produces better science and better writing. The journey from expectation to evidence is intellectually more interesting than a false story of perfect foresight.

Sources, current state and further reading

Source pages were checked for this edition on 5 September 2026. OSF is currently narrowing its broader project-workspace role while retaining a focus on study planning, preregistration and linked research records; older descriptions of OSF should not be assumed to describe every current product feature.

• OSF Support — Registrations and Preregistrations.
• Center for Open Science — current OSF and open-science overview.
• Center for Open Science — Registered Reports and journal practices.
• OSF Support — embargo and preregistration visibility.

Continue through the Library: Read Research Integrity and Publication Ethics, Experimental Design, Scholarly Publishing and Peer Review, and Systematic Reviews and Evidence Synthesis.