One Word Was Doing Two Jobs

Our methodology page explains how a gift earns its place here. It scores every product on four dimensions, and one of them is called TRUST. Further down the same page sat a second, separate system with a confusingly similar name: it audited the evidence behind a product and produced a Confidence Score out of one hundred.

Two different things, one word, and a reader met both of them inside a single scroll. A name that has to be explained twice is a name that is doing two jobs.

So the evidence audit is now called the Confidence Audit, which is simply what it already produced. The four scoring dimensions did not move. Anyone who has ever linked to the old section still lands in the right place: the previous address stays alive on purpose, because breaking every link that ever cited us would be a strange way to become clearer.

Renaming something published is not a search and replace, and treating it as one is how a rename goes wrong. Before a word was changed, an instrument measured every place that name lived: the theme, the machine files we publish for AI systems, the plugins, the specifications, the live articles, and the label on every one of our product pages. The first list, built by hand, had covered about a sixth of it. The instrument measured where the name lived. It could not measure what the page said each name meant, and those five descriptions were wrong the whole time.

Reading the page that closely turned up two things worth admitting.

The first was a citation. Three of the four scoring dimensions describe an effect that real published research has measured, and each names its paper. The fourth, TRUST, is not a finding: it is our own verification procedure. It had a paper attached to it anyway, and that paper had already been removed from our research database. The link under it pointed at a page where the study no longer appears. We had borrowed academic authority for something we do ourselves. That card now points at our Skeptic Protocol, which is the thing a reader can actually hold us to.

The second was a worked example. To show the audit in action, the page walks through one real product bar by bar. Three of those five bars had drifted away from what the product record actually held, while the total and the label still matched. That is exactly why nobody caught it: every quick check landed on agreement. The example no longer carries typed numbers. It reads them from the product, so it cannot drift again.

Every change here passed the same route as any other: an automated pass proposes, the checks run and their verdicts are quoted rather than summarised, and a human approves before anything ships.