CogniPrep writes about 77 assessment publishers, which means the site asserts a
few thousand facts that belong to somebody else. Those facts rot. Vendors
rename products, change clocks, launch a new range, and quietly drop a section.

Last weekend we re-opened one publisher's live site, product page by product
page, and fixed what had moved. The publisher is Test Partnership, a UK
company whose tests show up constantly in law and professional services hiring.
The sweep touched 22 files over five commits. Three kinds of change came out of
it, and the first one is the one I had wrong in my head before I started.

A rename is not a replace

Test Partnership has been moving its critical thinking test from "Concepts" to
"Insights". Our copy, written in September, hedged: it said the rebrand was
half finished, because at the time one page's body copy said Insights while the
factsheet it linked to said Concepts.

By 21 November the product page had settled on Insights. So the honest copy is
no longer "mid-rebrand", it is "renamed, and your invitation may be older than
the rename".

Here is the part I expected to be a sed command and was not. Counting the
string Concepts across the seventeen source files that mention it:

before the sweep: 69 after the sweep: 55 

Fourteen mentions died. Fifty-five are doing work, and they sort into five
kinds:

  1. Disambiguation prose. "Insights Critical Thinking (formerly Concepts Critical Thinking)". The whole value of the sentence is that it contains the dead name.
  2. FAQ questions phrased in the dead name. q: 'Is my test called Concepts Critical Thinking or Insights Critical Thinking?' A candidate holding an old invitation types the old name, and a page that has scrubbed it cannot answer them.
  3. Search aliases. Each test page carries a list of phrases people actually type, and concepts critical thinking and concepts critical reasoning both stay in it.
  4. A product that still carries the family name. Concepts Data Analysis is a real, current Test Partnership test. A blanket replace would have invented a rename the vendor never performed, which is the worst possible outcome of a tidy-up: a confident claim about somebody else's catalogue that nobody made.
  5. The vendor's own URLs. The factsheet we cite as a source is literally Product-Overview-Concepts.pdf. You cannot rename somebody else's file.

The rule that fell out of it, and that I would now apply to any catalogue of
third-party names: a dead name stays as an input and never as a label.
It is allowed in search aliases, in question text, in a "formerly" clause and in
a cited URL. It is not allowed as the name of the thing in a heading, a card
title, a page title or a test's own name field.

The claim we had to narrow

The second kind of change is the uncomfortable one. Several pages said this, in
various wordings:

Every assessment it sells is computer adaptive.

We could not verify that. The ability tests are adaptive and the game-based ones
are adaptive, both on the publisher's own description. The situational judgement
test, we could not confirm either way. So the sentence became:

Its tests are computer adaptive, from the Insights reasoning and critical
thinking tests and the TPAQ questionnaire to the MindmetriQ games.

and the thing we could not confirm went into the backlog as a question, not as a hedge in the copy:

Test Partnership (verified 21 Nov 2026): SJT adaptivity, its scenario count
and whether blank ratings are allowed.

That is a worse sentence to write and a better sentence to publish. "Every" is
the word that does the damage: it is the kind of universal claim you get for
free by writing confidently, and the only reason it survived two months is that
nobody re-read it against a product list.

The same pass downgraded a smaller one. Our copy said the personality
questionnaire is untimed; it is untimed by default, and an employer can attach a
time limit. One word, same failure mode.

The number the vendor gets wrong about itself

The third kind is my favourite, because it is the one you cannot fix by reading
more carefully.

Our pages said the publisher's game-based tests run per-item clocks of 15 to 30
seconds. That figure came straight from the vendor's own anti-cheat page, which
describes "short, game-based questions with time limits of 15 to 30 seconds".
Good provenance: first-party, current, and on the record.

It is also wrong. Opening the publisher's own public demos and timing them gives
clocks from 6 seconds to 20, depending on the game. The marketing page is
describing a range that no longer matches its product.

So the code comment that cites the source now cites both, in order of weight:
the demos as the figure, the marketing page as the claim that disagrees with
them. Three places on the public site that repeated "15 to 30 seconds" now give
the per-game numbers instead.

The lesson I keep relearning: first-party does not mean current. A vendor's
marketing page is a primary source for what the vendor says, and a secondary
source at best for what the vendor's software does.

See it

Everything above is on public pages, and the dead name is visible exactly where
it should be.

  1. Open /games/test-partnership. There is a block headed with the names a candidate might be holding, which states plainly that "Concepts is the old name". Further down, in the list of this publisher's tests we deliberately do not simulate, is Concepts Data Analysis, which is why the rename could not be automated.
  2. Open /games/test-partnership/critical-thinking. The page title and every heading use Insights. One FAQ is phrased in the old name: "What is the difference between Insights and Concepts Critical Thinking?" That asymmetry is the rule in action.
  3. Open /cheating/test-partnership and find the adaptivity paragraph. It now begins "Its tests are computer adaptive" rather than "Every assessment it sells". The same page carries the per-game clocks, "6 seconds per Link Swipe pair", where it used to carry the vendor's 15 to 30.

If you want to count the surviving mentions yourself:

curl -s https://cogniprep.app/games/test-partnership | grep -o Concepts | wc -l 

Expect a bigger number than the page shows, because a Next.js page ships its
text twice, once as HTML and once in the streamed payload the client hydrates
from. Count with grep -o | wc -l rather than grep -c either way: this site's
HTML is two lines long, so a line count tells you nothing.