Failure catalogue · The number is plausible, and wrong · 4 of 51

OBSERVED FAILURE MODE

The Total row that is also data.

A summary row sits in the same columns as the rows it summarizes, so a pipeline reading records reads it as one more record. Sum it alongside the rows it already sums and the answer doubles. Nested subtotals do worse, and a totals row printed above its detail is counted by anyone reading top to bottom.

What we saw

A grouped circulation report at Meridian Journal Group placed a subtotal after every title and closed with a grand total, all of them in the same columns as the detail rows. The naive sum came out a clean multiple of the truth, which is the tell: a figure that is exactly two or three times right is almost always aggregation rows counted beside detail rows. Listing the subtotal labels by hand would have worked that week and broken silently the first time a title was added, which is why the exclusion is written as a label pattern.

Why it passes a glance

A total row has the shape of data because it is printed in the data's own columns. Row counts look plausible, every value is real, and the arithmetic of the sum is faultless. A reviewer glancing at the top of the file never reaches the row at the bottom, and a check that recomputes a total from the same rows can never disagree with itself.

What addresses it

The anchor-discovery skill treats an embedded total as the anchor to tie against, never a row to sum: the anchor is selected by what the row says rather than where it sits, excluded from the measure by a negative pattern that survives new groups, and tested before it is proposed. Doctrine principle 1, an AI never authors a number, requires that tie against the file's own stated total.

Check your own file in two minutes

1. Search your key column for values beginning with Total, Subtotal, Sum, All, or the entity's own name, and list every row that matches.
2. Sum the money column with those rows in and again with them out; a clean two, three, or four times gap means the file summed itself.
3. Check the top of the file as well as the bottom, since some exports print their totals row above the detail.
4. Exclude the aggregation rows by label pattern, never by listing the group names you can see today.

What this does not catch

Pattern exclusion depends on the file labelling its aggregation rows. A subtotal with a blank label, or one carrying a real product name, is invisible to it, and a hierarchy with more than one level still needs a person to say which level the anchor is meant to sum.

Quick answers

Why not exclude the total row with an allowlist of the values I want?

Because an allowlist silently drops every future value that is not on it, and the total it produces still reconciles against itself.

Is the totals row always the last row?

No, some exports print their totals above the rows they summarize, and reading top to bottom counts that row before you know it exists.

What does the total row buy once it is excluded?

It becomes the anchor: the one figure the file asserts about itself that your computed sum has to match.

Nearby failures

Last updated 2026-09-02 · Dylan, founder · one of 51 observed failure modes, every one seen in a real build or in our own audits, none invented.