Every Parallel Artifact Needs a Merge-Back

The technique worked. That is what makes it worth writing down.

I re-analysed a story I had already analysed, deliberately without opening my first pass. A blind second read, on purpose, because the first analysis anchors you: once you have written down what a document says, you re-read your own summary instead of the document.

The second pass surfaced three real questions the original had missed. Three. Not stylistic differences — genuine gaps in the first analysis.

Then the register printed the bill.

The Symptom

The story showed 9 open questions.

Only 6 of them are distinct.

What the Register Was Actually Indexing

Both analysis files now exist. Both are indexed against the same story. Each one carries its own numbering scheme, generated independently, so the same question appears twice under two different IDs.

Three questions are double-counted. Six distinct questions, nine rows.

Nothing malfunctioned. The tooling did the only sensible thing available to it: it found two analysis artefacts for one story and indexed both, because it has no way to know that one is a second opinion about the other rather than a second piece of scope. Files are the unit it can see. Intentions are not.

I had kept the first file "for reference." That phrase is doing all the damage in this story.

Why "For Reference" Is Not a Neutral Act

A review technique gets judged on what it finds. Almost never on what it leaves behind.

I evaluated the blind second pass entirely on its yield — three new questions, clearly worth the hour. That evaluation was correct and it was incomplete, because the technique has a second output that nobody counts: an additional artefact describing the same story, with the same authority, in the same folder.

Kept for reference, it becomes a source of truth. Not because anyone decided it should be, but because there is nothing about it that says it is not one. A file that looks like an analysis, sits where analyses sit and is named like an analysis will be read as an analysis by the next person and indexed as one by the next tool.

What It Was Not

It was not a defect in the register. The count of 9 is arithmetically correct given its inputs.

It was not an argument against blind second passes. The three found questions are real and would still be missing.

And it was not a numbering collision I could have avoided with better IDs. Two independent schemes were the honest outcome of writing the second analysis without reading the first — sharing an ID scheme would have meant looking at the original, which is precisely the thing the technique forbids.

The defect is not in any of the two artefacts. It is in the absence of a step after them.

The Metric That Would Have Caught It

Here is the sharp bit.

The way you would notice traceability degrading is by looking at the question counts per story. That is exactly the number the second artefact corrupted. The instrument and the damage are the same object.

So the technique that improved my coverage silently degraded my traceability, and it did it in the one place where the degradation is invisible — inside a number that got bigger, which in a question register reads like diligence.

Nine open questions on a story looks like thorough analysis. It looks better than six.

The Rule

Any review that produces a parallel artefact must end with a merge-back into the canonical one.

Not "keep both and remember." Not "archive the old one somewhere." Merge the findings into the single artefact that downstream tooling indexes, then delete or clearly de-index the parallel copy — including moving it out of the path the tooling scans, because a file marked superseded in its own header is still a file to a directory walk.

The merge-back is part of the technique. Not admin afterwards. A blind second pass that ends with two files is half-executed, in the same way that a refactor that ends with the old class still compiling is half-executed.

Budget for it when you plan the review. If the hour for the second pass is available and the twenty minutes for the merge is not, do not run the second pass — you will get the questions and the double count together, and only one of them announces itself.

How many of your stories are described by more than one document right now, and which one does your tooling believe?