Reading an Accuracy Snapshot

Our methodology article defines the metrics, and Checking the Checks explains an earlier label review. This is a practical reading guide to one saved report: campaign-2026-10-07T00-33-50-364Z. No new campaign was run to write this article.

Checks are not sources

A campaign tests multiple usernames per source. Its 772 rows are checks, not 772 websites or 772 independent observations. They cover 265 automatic sources; manual-only links are not automatic account verdicts and do not belong in this accuracy denominator. Several rows from one source can share the same endpoint failure or stale rule.

Two percentages, two denominators

The report records 186 true positives, 426 true negatives, zero false positives, and eight false negatives. That is 612 correct decisions out of 620 clear decisions: 98.71% decision accuracy, rounded to two decimal places.

There were also 152 abstentions. Coverage is therefore 620 divided by all 772 checks: 80.31%. Accuracy does not mean that 98.71% of every attempted check returned a correct clear answer. It describes the rows where a clear verdict was reached.

Conservative unknown and blocked results remain part of the coverage story. They are not automatically errors, and they are not secretly counted as correct missing accounts. For an individual search, keep the status and its reason codes rather than converting every unclear result into not_found.

The label-quality boundary is smaller still

Only 129 rows were calibration-eligible; 643 were excluded from calibration. The report groups the eligible rows by source into 99 calibration rows and 30 test rows. The other 643 rows form the diagnostic set.

This is why the full-campaign percentages must not be presented as a definitive held-out accuracy estimate. The large diagnostic view helps prioritize repairs, but it is not interchangeable with the smaller reviewed subset used for calibration and testing. A label can be useful for investigation without satisfying the stricter eligibility gate.

Zero observed is not zero possible

The saved run observed zero false positives against its labels. That is important, but it does not establish a universal false-positive rate of zero. Account pages change, controls can age, and the eligible sample is limited. Eight labelled positive checks were still missed in this run.

Nothing here establishes superiority over another tool. A comparison would require matched sources, usernames, labels, timing, and access conditions. Nor do these account-verdict figures measure the correctness or completeness of profile pictures, biographies, or other metadata. Those fields have a separate evidence boundary.

A reusable way to quote the report

“The saved 7 October 2026 campaign recorded 98.71% decision accuracy at 80.31% coverage over 772 checks across 265 automatic sources, with zero observed false positives. Only 129 rows were calibration-eligible; these are campaign diagnostics, not a guarantee for future searches.”

Keep the date, denominators, and caveat together. A later focused retry is evidence about those retried cases, not a replacement for the full campaign. And a found username remains a research lead, not proof of identity.