DEV Community

Daniel Pertu
Daniel Pertu

Posted on

We deleted 551 words of "we could not find a source" from 42 pages, and a grep found the same habit on 20 more

CogniPrep publishes 211 guides to employers' hiring assessments. Each one answers a search: which test does this company use, what is in it, how long is it, what should I practise. Everything on those pages is sourced, and for a long time we were proud enough of that to say so in the prose.

Here is what that looked like, on the page about a Dutch supermarket chain:

Neither listing says whose cognitive or personality test it is. We could not find a source that does, so we do not claim one. If your invitation names a publisher, that is the one to practise for.

Three sentences, of which the middle one is about us. A candidate who is sitting this test next week does not need to know about our afternoon. It now reads:

Albert Heijn's vacancies do not say whose cognitive or personality test it is. If your invitation names a publisher, that is the one to practise for.

Same facts, same honesty, one fewer sentence, and the thing it opens with is the thing the reader asked.

The measurement

One commit, 42 of the 211 employer pages, 106 insertions against 146 deletions. Counting words rather than lines: 1,875 out, 1,324 in, so 551 words of net deletion, and not one fact changed.

The constructions that went, counted across the diff:

phrase removed added
we found 11 0
we have not 5 0
no source 5 0
as checked (in November 2026) 5 0
we cannot 3 0
nothing we 3 0
according to 3 0
we could not find 1 0
we do not claim 1 0

Two pages were bad enough to get a full rewrite rather than edits, at 1,185 words out and 896 in between them.

A second example, this time from the footer disclaimer on a US telecoms page, because footers are where this kind of thing accretes:

This page rests on AT&T's study-guides page and the test guides linked from it, dated between 2017 and 2024, as checked in November 2026.

became

Details are correct as of November 2026.

The part that is not a find and replace

It would be nice if the rule were "delete the first person". It is not, because some of those sentences are carrying a qualification that the fact needs, and some are carrying nothing.

The test I ended up using: is this sentence about the world, or about our research? "The ADF has not said who supplies it" is about the world and stays. "Which is as far as the evidence goes" is about our research and goes.

That second phrase came off a page where the three paragraphs immediately above it list the three specific ways the test matches a known product. So the deleted sentence was telling the reader to be cautious about evidence they had just finished reading. The replacement states the match and leaves the evidence where it already was.

Not every case was that clean. One page said:

That makes Cappfinity the likely supplier, but nothing we found confirms the live test is theirs.

and now says:

That makes Cappfinity the likely supplier of the live test too, though the provider is not named.

That is a genuinely stronger claim, and I want to be straight about it: the hedge did not disappear, it moved into "likely" and "not named", which are doing the same work in half the words. But a sweep like this is exactly where a page quietly gains confidence it has not earned, so the diff is worth reading line by line rather than trusting the phrase list.

The sweep was a review, so a grep beats it

That commit came out of reading pages, not out of grepping for a phrase list, which means it covered the 42 pages somebody opened. Afterwards I ran the grep I should have run first, across every public page file:

we found | we could not find | no source we could | no source we found |
we could not establish | we cannot confirm | we do not claim | we have not found
Enter fullscreen mode Exit fullscreen mode

25 hits still in rendered prose, in 20 content files: 8 on employer pages, 13 on provider pages, 4 in blog articles. Five of those employer pages are ones the sweep never touched, so the habit is not gone, it is just gone from the pages that were read.

Which is the useful bit of this whole exercise. A prose convention enforced by review is enforced on the files in that review's working set. If you want it to hold across 211 pages, the phrase list has to run over all of them.

Reading all 25 changed what I think the rule is

I expected to find a backlog of the same filler. Almost none of it is. Grouped by what the sentence is actually doing:

Rebutting a claim the reader has already seen somewhere else. This is most of them, and all of them follow the same shape: the internet is confident, no first-party document agrees, say so.

Preparation sites place ACER behind the Australian Government Graduate Program, the APSC, the ABS and Services Australia, but we found no ACER statement, no employer statement and no procurement record supporting it ... We do not make claims we cannot source, so we do not claim it.

Nobody publishes this, including HireVue. Its CodeVue FAQ discusses due dates measured in days and never mentions a challenge clock, and no source we could find states a problem count either.

Delete the first person from either of those and you get a bare negative we have no standing to assert, or silence, which leaves the reader with the prep sites. The narration is the content.

Declining to claim something about our own practice material. Two of the 25, and they are the sentences I would least like to lose:

We do not claim any particular Arctic Shores task was the HMPPS one.

Four [of the eleven question types] are named by a Dutch preparation vendor, and those four are the ones we mix. We do not claim to know the other seven.

So the rule that came out of this is not "never write 'we'". It is:

  • A sentence about the world: keep. "Albert Heijn's vacancies do not name a publisher."
  • A sentence about our research, where someone else's claim is what you are answering: keep. That is a rebuttal and it needs a narrator.
  • A sentence about our research, where nothing is in dispute: delete. The reader gets the fact and keeps the 13 words.

The 38 deletions in that commit are all the third kind. I could not have written that distinction down before reading the diff, which is probably why the first version of this rule was just "cut the caveats".

What the pages hedge with instead

Across all 211 employer pages, the uncertainty now rides on the individual fact rather than the paragraph:

  • may 132
  • reported 125
  • likely 72
  • Not published 29
  • suggests 28
  • according to 15
  • as of November 2026 4

"Reported" attaches to a figure a third party published. "Not published" is a literal statement about a vendor's documentation, and it is the standing phrase for a number everyone else on the internet is happy to invent. Neither is a sentence about us.

None of the sources left the codebase either. They stay in the page registry and the research files, which is where the honesty record belongs: a reviewer can audit every claim's provenance, and the candidate reading on their phone the night before an assessment does not have to.

See it

Open cogniprep.app/employers/albert-heijn and read the two boxed callouts and the first FAQ answer. Every sentence is either a fact about the employer or advice to the reader.

Then open cogniprep.app/employers/bae-systems and search the page text for no source. That is one of the 25 survivors, and it is the kind that should survive: it answers a confident claim other prep sites make, which a bare negative could not do.

If you publish research-backed content, the grep worth running on your own pages is any sentence whose subject is "we" and whose verb is about looking. Then read every hit rather than replacing them, because a meaningful minority of them are the only honest answer you have.

Top comments (1)

Collapse
 
docsalive profile image
DocsAlive Team •

The distinction between a world-facing statement and narration about research is much more robust than a blind “delete we” sweep. I’d run the phrase scan against rendered public pages in CI as well as source files, so templates or generated pages can’t reintroduce a sentence the source grep missed.