291 employer profiles, 301 citations, and the 44 that record finding nothing

CogniPrep has 180 pages of the form "what assessment does employer X set". 48 of them were added last week. Every one is generated from one 325 KB TypeScript file whose job is to make an unsourced claim hard to write.

The shape:

export interface ProviderEvidence {
  provider: AssessmentProvider;
  /** A URL that was actually opened and checked, not a plausible-looking one. */
  url: string;
  type: EvidenceType;
  /** Publication or retrieval date, ISO-ish. 'undated' when the source has none. */
  date: string;
  confidence: Confidence;
  endedOn?: string;
  /** What the source actually says. Shown to editors, not rendered. */
  note: string;
}

EvidenceType is a five-member union, ordered by how direct the evidence is: the provider's own case study or press release, the employer's own careers page or job posting, multiple independent candidate reports agreeing, press with an editorial standard and a named masthead, and last, a competitor preparation site's claim, documented in the type itself as "weakest; never sufficient alone".

291 profiles, 301 associations, 132 distinct source domains. The type histogram:

• first-party-provider: 174
• employer-official: 105
• prep-site: 12
• candidate-reports: 7
• press: 3

Eight citations are date: 'undated'. Nine point at web.archive.org, because a vendor case study from 2022 often exists nowhere else.

The 44 profiles with no associations at all

111 of the 291 profiles have no page. 44 have no associations either: an empty array, a hasGuide: false, and a note saying what was checked. They exist because a research result of "nothing usable" is a result, and the only way to stop four people rediscovering it over a year is to write it down where the next person will look.

The plan for last week's round says it in one line: reporting "found nothing" is a correct result, not a failure.

Ten of those 44 do have a page, through one optional field:

/**
 * The real vendor when it is one we do not simulate, or the employer's own
 * in-house instrument. Present so a page can say what a candidate will actually
 * sit even when we do not simulate it.
 */
unsupportedVendor?: string;

Sixteen profiles use it. A few of the notes behind it:

• Starbucks: its own "Applicant Insights" questionnaire. The careers FAQ gives the duration, the seven-day window and the 90-day carry-over, and names no vendor at all.
• The Home Depot: its own online assessment, with item counts and a 96-hour window on its own candidate page, and no vendor named.
• Panera Bread: "None found. No assessment vendor is reliably reported, and recent candidate reports do not mention a test at all."

Writing a page that tells a candidate "the thing you are about to sit is not something we sell practice for" is commercially stupid and editorially the only option. Those pages exist because the search traffic is real and the alternative is somebody else's page guessing.

Two gradings that drifted apart

Here is the finding I did not expect when I counted. confidence is documented as a function of evidence type and date: low is "a single prep-site or candidate claim". The data says otherwise.

• 28 of the 29 low-confidence citations are first-party-provider, downgraded because they are undated or old.
• 11 of the 12 prep-site citations are graded medium, usually two independent prep sources describing the same test in detail.

So the two axes are genuinely orthogonal in practice, and the comment that defines them has not caught up. Either the gradings or the definition is wrong, and it is a real thing to fix rather than a curiosity: the medium floor in the function that maps an employer to practice links means a grading decides what gets linked.

Where the rule does hold exactly: 23 profiles whose every citation is low confidence, and not one of them has a page.

Eight profiles rest on prep-site evidence alone and do have pages, which is the interesting edge of the rule. Those pages are about the employer's process, which is sourced from the employer, and the vendor line is labelled as a report. Live copy from one of them: "A Wonderlic association with Medline is reported by candidates and by assessment-preparation sites, but we have not found any first-party confirmation from Medline or Wonderlic. Treat it as a well-corroborated report, not a confirmed fact."

The notes that correct us, not the source

Several notes exist only to kill a claim the site used to make:

Medline is NOT on wonderlic.com/customers, contrary to what this guide used to claim.

Allstate is NOT on wonderlic.com/customers, contrary to what this guide used to claim.

A citation field that records a disproved claim earns its keep more than one that records a true one. Nobody re-adds a claim that has a dated note next to it explaining why it went.

Four more corrections landed during last week's round, all of them subtractions: a Predictive Index ownership claim deleted from a vendor hub because no source supported it, thirty lines of employer names deleted from four vendors' company lists because their sources showed a commercial relationship rather than hiring use, an airline claim dropped, and an ownership history fixed (Hay Group bought Talent Q in 2014, Korn Ferry bought Hay Group in 2015).

And one correction that was itself wrong. Mid-round, a commit decided a vendor does not publish whether its tests penalise wrong answers, and rewrote several pages to attribute that to prep sites instead. The last commit of the round reversed it: the vendor's own candidate preparation page does say it, in general terms, and the pages now quote the vendor. Two commits, fifty minutes apart, in opposite directions. Nothing in the process catches that except opening the source again, which is why every citation carries a date.

endedOn is the field for the opposite problem. Two associations have it, one added last week. The evidence stays in the file so a page can explain the history to a candidate holding an old invitation, and the function that maps an employer to practice links never returns an ended association, because that list is a promise about what somebody is about to sit.

See it

Open cogniprep.app/employers/medline and read the "What is sourced" block. It names the limit of the evidence on the page, not in a code comment.

Then open cogniprep.app/employers/starbucks and find the "Who builds it" section: Starbucks does not name a vendor, some prep sites do, and the page says we found nothing supporting it and will not repeat it. Compare that against what the first three search results for the same query will tell you.

cogniprep.app/employers is the index, if you want to see how many of these there now are.