A trial site records a visit on a paper worksheet and types it into your system a week later. What has to be true for that data to count?

The worksheet is now the source record

The moment the measurement was first written on paper, that paper became source data, and everything else follows from that fact. Source data is where an observation is first recorded, and the electronic case report form has become a transcription of it rather than the original. So the paper cannot be discarded once it is typed in. It has to be retained, filed against the subject and visit, legible, attributable to the person who wrote it, and available to a monitor, an auditor and an inspector for as long as the trial's retention obligations run.

That single reframing answers most of the question. The value in the database is only trustworthy to the extent that it can be traced back to something, and a transcribed value whose origin has been thrown away is unverifiable no matter how carefully it was typed. If the site's intention is to work on paper as a matter of routine, then the worksheet is part of the trial's documentation and should be a controlled form with the subject identifier, the visit, the date and time of the observation, the units, and the initials of the person recording it — not a printout of the screen, and not a sheet of scrap.

The corollary is worth stating explicitly because candidates miss it: the medical record is often also source, and where a measurement was taken in routine care the medical record entry is the source rather than the worksheet. Knowing which document is source for which variable is a decision made when the trial is set up, not discovered during an audit.

Contemporaneous is about the record, not the typing

Regulators describe data integrity expectations in a familiar set of attributes — data should be attributable, legible, contemporaneous, original and accurate — and the awkward one here is contemporaneous. A week's delay does not, in itself, invalidate the data, because the paper record was contemporaneous with the observation. What matters is that the electronic record does not misrepresent when the observation happened.

The design consequence is that the visit or observation date and time is a data field in its own right, entered by the site and describing the clinical event, while the entry timestamp is system metadata applied automatically. Conflating the two is the classic defect: a system that stamps the observation with the time the form was saved has quietly produced a false record, and once a study report is built on those timestamps it becomes very difficult to unpick which values reflect the schedule and which reflect a backlog.

That separation also lets you compute the thing a data manager actually wants, which is the interval between the visit and the entry. That interval is a quality metric per site, not a footnote, and it is the input to the monitoring decisions discussed below.

The audit trail is what makes the value defensible

The reason electronic data capture is acceptable at all under the regulatory expectations for electronic records — 21 CFR Part 11 in the United States and equivalent expectations elsewhere, alongside the Good Clinical Practice guideline — is that the system can prove what happened to a value. Retrospective entry raises the load on that mechanism rather than reducing it.

The audit trail has to record who entered the value, when, what the previous value was, and, for any change after initial entry, why. That last field is where the interesting distinction lives, and a strong answer names it: correcting a transcription error is a different act from changing a clinical observation. The first is a data-entry correction with the paper as evidence, and the reason is simply that. The second means the observation itself was wrong or has been reinterpreted, which is a clinical judgement belonging to the investigator, and it should be visible as such. A system that offers one generic "corrected" reason code makes those two indistinguishable in the record and therefore forces an auditor to open every case.

Two further mechanisms carry weight. The audit trail must not be editable or disableable by the users whose actions it records, which is why permissions on it are usually raised in an inspection. And the investigator's confirmation of the data — the point at which a responsible person attests that the entries for a subject are accurate — has to be an act performed after the entries exist, not a checkbox at the start.

Alongside all of this, the query mechanism is part of the evidence: a value flagged as out of range, questioned by a monitor or a central review, and then either confirmed or corrected, leaves a trail that tells the story of the data more convincingly than the value ever could.

What the delay tells you about the site

Treating this purely as a data question misses the operational half. A week's lag is a signal, and the interviewer usually wants to hear that you read it.

Late entry degrades every control that depends on seeing data while the subject is still in front of you. Safety events reported late delay the sponsor's assessment and any reporting obligations that flow from it. Eligibility and stratification decisions that depend on entered values cannot be made. Central statistical monitoring, which is how a sponsor detects a site with implausible data, is blind for as long as the backlog lasts, so the site that is most likely to have a problem is also the one you can see least of. Missed visits and protocol deviations get discovered at the point where nothing can be done about them.

So the response has two parts: a data-integrity part, which is the source documentation and audit-trail discipline described above, and a monitoring part, which is to treat the entry lag as a risk indicator, escalate the site, ask why paper is being used at all — often a genuine access or workflow constraint worth fixing rather than a discipline problem — and consider whether the affected data needs additional verification against source at the next monitoring visit.

Why the value in the field is the least of it

The instinct is to focus on getting the number right. In a trial, whether the number is right is not something the system can determine, and the system's job is instead to preserve the chain that lets someone else determine it later: the paper it came from, who wrote it, when the observation occurred, who transcribed it, what it was before any change, why it changed, and who attested to it.

This is why the answer to "what has to be true for that data to count" is not a validation rule. A locked study database whose entries cannot be traced to source is worth nothing when a regulator asks, and the point at which that becomes apparent — an inspection, or worse, a submission — is long after the site has moved on and the worksheet has gone.

The paper is now part of the trial, and the database is a transcription of it. Keep the observation date separate from the entry date, make the audit trail say why a value changed as well as that it did, and read the lag itself as a signal about the site.