July 2026 · AisleLens v3.2
Correcting my own error bound
I published 0.83%. A defect with no quote to question changed the bound to 7.80%.
When this was measured
Measured during v3.2 CP3; recorded 27 July 2026. The exact run timestamp is not given in the source section.
Historical result, not current accuracy. Note written 1 November 2026.
I had published a cluster-adjusted 95% error bound of 0.83%. The general sample had 507 passing rows, and the earlier audit had confirmed zero false positives. That number sounded reassuring. A later check of one defect class found eighteen false positives in that same sample. The bound became 7.80%. I needed to correct the published claim, not explain the discrepancy away.
The blind spot was an identifier row. It said that a product's structured data published a manufacturer part number, or MPN. Unlike other rows, it did not show a quote that a reader could inspect. It looked much the same whether the value was a useful product identifier or a number the storefront had assigned to its own database record.
The engine rejected obvious placeholders and checked barcode check digits. But it accepted any non-placeholder MPN string. The row's promise was about an identifier useful outside that one store. A store-local object ID could satisfy the implementation without satisfying that promise. Reading the rendered conclusion again would not reveal the missing distinction.
A mechanical check went back to the captured bytes. It could compare the reported identifier with values used elsewhere in the page, including the storefront's internal product ID. That found eighteen defects in the general sample. The audit had not become more pessimistic. It had gained a way to see a class of error its earlier method missed.
The corrected figure needs care too. Eighteen was a floor from one mechanically checked class, not proof that every other class was clean. And 7.80% was a cluster-adjusted upper bound under the recorded method, not an observed error rate and not a current accuracy claim. Replacing one impressive-looking percentage with another would have repeated the mistake.
I now take a missing explanation as something to examine in its own right. If a result offers no underlying value or quote to question, the review method needs another route to the evidence. More rows cannot repair a blind spot shared by every row. This July 2026 correction stays in the record because the useful lesson is how the earlier number became possible, and what the correction still could not establish.