The trace of a correction in Evidrel

Correcting a record changes more than its result. We also need to leave an understandable account of why we relied on the earlier result and why we no longer do.

There is a number in a research note. Its source is known, the conditions under which it was obtained are documented, and it has been reviewed. Some time later, another document arrives with a different value for the same measurement.

Writing the new number over the old one is easy. But doing so loses something that looks small: the transition between the two records.

Where did the earlier value come from? Does the new record describe the same experiment? Was a calculation error corrected, or did the conditions change? How should we read a decision based on the old value today?

In the previous essay, I considered the distinction between keeping a result and continuing to use it. Here I want to look at a smaller part of that distinction: the trace a correction leaves in the record.

Two numbers do not yet make a correction

Consider a simple example. An experiment report records 1.03 as the value of a particular metric. A calculation note received later gives 0.97 for the same metric.

These are hypothetical values chosen to explain the issue. They are enough, however, to show why the decision is difficult.

Two records being different does not establish that the second corrects the first. Perhaps a different dataset was used. Perhaps the scope of the experiment changed. Perhaps two different measurements share a name.

Before calling it a correction, we need to understand what stayed the same.

In Evidrel’s local research notebook prototype, this decision relies on human review. Records linked by a correction must have matching titles and scopes; their sources must be current, and the records must have been reviewed and approved. A record that merely mentions a claim cannot serve as its correction.

These are the boundaries of the operation. Matching titles do not prove that two statements mean the same thing. Nor does the system discover the scope by itself; the reviewer specifies it.

The central question therefore remains beside the sources: Do these two records really refer to the same thing?

What changes is the link between records

Suppose we have checked the sources and established that the new calculation concerns the same experiment and metric. We reviewed both records and decided that the new one corrects the old one.

At this point, instead of deleting the old record, we create a correction link between them.

Before saving, the effect of the operation is shown: when the new record becomes active, the old one remains in the history but leaves the default set of approved results.

We can therefore return to what the earlier report said. At the same time, a search does not present the earlier value as one of the currently accepted results.

This is one of the things I need from a research notebook: to see why a record was used in the past, and to follow why it has now given way to another.

A correction becomes more than placing two values side by side. There is a relation between them, established through a review decision and open to later questioning.

The decision has a history too

A correction can itself be wrong.

We may have matched the wrong records, defined the scope too broadly, or misread a statement in the new document.

The review decision should therefore not remain an invisible operation on the text.

For record reviews, the current prototype keeps the before and after fields, the decision, the entered reviewer name and the time. The correction link is recorded as a separate operation. The source passage remains alongside the record.

This trace can help us reconsider the grounds of a decision. But entering a reviewer’s name does not authenticate an identity. Approval does not establish scientific truth either; it is a review decision about how the source has been recorded within the stated scope.

What the notebook can show, in this limited sense, is what was consulted and what operation was performed.

When the correcting source itself changes

Take the example one step further. The document carrying 0.97 changes, and its new version is added to the notebook.

We need to check again whether the new version still supports the earlier record. In this situation, the prototype sends records attached to the old source version back for review.

What happens to the first record, the one carrying 1.03?

It does not silently return simply because the grounds of its correction changed. The old record dependent on that correction can also require review.

The thought behind this behavior is simple: losing the grounds for a new acceptance does not automatically justify the earlier one.

If a return is needed, the historical link must be explicitly revoked, the sources reconsidered, and, where necessary, a record created against the new version. Revoking the link leaves a trace in the history too.

Rather than automatic certainty, this opens a space that needs another look.

More traces, more responsibility

There is a cost to this approach. When we do not reduce everything to the latest value, we keep more records, links and review decisions. Treating source changes cautiously can also increase the work of reviewing again.

Evidrel’s current research notebook is not a general semantic system that removes this burden. It is a local, single-user prototype. Search relies on text matching, and a person specifies the scope. Performance on large archives, its effects on human review and time savings have not yet been measured.

Still, here is the question I find worth working on: can we record a correction clearly enough for a later reader to reconsider it?

When I return to the same note after some time, I want to see more than the latest value. I want to read why we accepted the earlier one, which source led us to change it, and under what conditions we can use that change today.

The trace of a correction should make that return possible.

---

Source note — 18 October 2026: This essay draws on the usage documents and the record, review, correction and source-version rules of Evidrel’s local Research Notebook prototype, previously named Örgü. The values 1.03 and 0.97 are hypothetical. No new experiment, test or human evaluation was conducted for this essay.

Connections: Evidrel research page · Previous essay: When should we stop using a result?

Atlas concepts: Knowledge · Relation · Context · History · Evidence · Uncertainty