Research
What we can show, sourced. Where a figure comes from someone else it is cited; where it comes from our own systems, it is measured rather than asserted.
Northland’s job ads are up. Its employment is flat. Its unemployment rate went from 5.4 to 8.8 percent.
Four official series over the year to June 2026 that do not tell one story: ads up 4%, filled jobs down 2.2%, and the unemployment rate from 5.4% to 8.8% - the largest rise in the country. The easiest explanation ruled out, no cause claimed.
5 official sources · agency releases only · no cause claimed
Read the piece →Security · July 2026Prompt injection has come for hiring. So we attacked our own system.
Hiding instructions in a CV to fool an AI screener is a measured, spreading attack. What the primary evidence really shows, why every AI-in-hiring provider is exposed, and what an authorised adversarial test of our own pipeline found - including the one edge case it did.
9 sources · primary-checked · first-party security test
Read the piece →Psychology · July 2026Why an honest application is easier to carry
Peer-reviewed research on authenticity, deception and stress suggests truthful self-presentation lowers anxiety and cognitive load, and softens rejection. What the evidence says, what it doesn’t, and what to do with it before an interview.
9 sources · peer-reviewed · every citation checked
Read the piece →Method · July 2026How we test that the ATS screening tells you the truth
The ATS screening only helps if it is honest about the gaps. How we test that it is, why a second model does the grading, and where code had to back the model up.
3 cited sources · first-party evidence
Read the piece →Market research · July 2026The NZ Job Market 2024–2026: Past the Low Point, Still Under Pressure
Applications per advertised position are up 261% since 2022, agency data averaged 43 candidates per job, and job creation has fallen. Where the labour market stands and why applications have become harder to read.
22 sources · source-checked
Read the piece →Why we publish this
Applying for a job in 2026 means being read by software before a person sees you, and being told by other software what to send. Neither side of that shows its working. You cannot tell whether a tool is strengthening your application or quietly inventing things on your behalf, and you find out which at the interview.
We are building one of those tools, which makes us the last people who should be taken on trust. So we publish instead: what the numbers actually say, where they came from, and what our own system does when we test it honestly - including the parts that fail.
There is also plainly a gap worth filling. The data on the New Zealand labour market exists - Stats NZ, MBIE, the job boards’ own reporting, international research on AI in hiring - but it is scattered across releases months apart, and mostly written for economists. Reading fifteen sources and working out what they jointly mean for someone applying for a job this week is work almost nobody is doing for the public. It is work we are good at, so we do it and show every source.
How we hold ourselves to it
Claims are triangulated across sources rather than repeated from whichever was published most recently, and where sources disagree the report says so. Of 102 falsifiable claims extracted for the market report, one was refuted and removed - reporting what gets thrown out is the only way to know the rest was checked. The same standard applied to what the model writes.
Every claim we publish traces to a named source, and we say what a claim doesn’t cover rather than imply more than we’ve shown. Where a claim carries real weight we go further and check it against a second, independent source - those are different jobs, and each report’s method section says how many claims got the second treatment and what happened to the ones that failed it. Neither rules out every mistake. When we find one, we fix it promptly and date the change - and where a correction affects the substance, we say what changed, at the top of the piece.