Press
Every claim on this page is checkable before you print.
Each figure names its source and its method, and the gaps are listed here as plainly as the facts.
Research
Speech is not a flat string of words
Olive builds voice models to improve understanding, not to imitate a human speaker. A speech system normally collapses what you said into flat orthographic text at the earliest stage. Timing, pauses, emphasis, lengthening and overlap are discarded before any model reads the transcript.
Olive does not treat speech as a transient input that collapses into a string. It outputs a linguistically structured transcript that preserves that interactional information. Keeping it matters for usability, since turn-taking and intent resolution depend on it. It is also a fairness safeguard: socially meaningful cues in speech, including dialect, can become a covert risk surface.
The fairness motive here is measurement rather than screening. A transcript that retains timing and emphasis allows systematic tests of downstream model behavior. Hold the meaning constant, change the dialect or the prosody, and see what moves.
The number that cost a feature
Olive measured its own notation instrument against human ground truth and got a macro F1 of 0.203 at best sample, 0.17 to 0.20 across the range.
That instrument was demoted from a scored dimension to a proctoring signal, and no prosody-derived field may reach a candidate's result.
Provenance
Who measured what
Two of these are Olive's own measurements. The rest belong to their authors, are cited as theirs, and were not discovered here.
The transcription gap is measured
Koenecke et al. found an average word error rate of 0.35 for Black speakers against 0.19 for white speakers, across five commercial speech recognition systems. PNAS, 2020.
The 84% figure is ours
That relative gap — (0.35 − 0.19) / 0.19 — is arithmetic Olive did on two numbers the paper published. Printing it beneath someone else's citation is the error this whole discipline was built to stop.
Safety training misses dialect
Hofmann et al. report that dialect prejudice survives human-feedback training, which "obscures the racism on the surface, but the racial stereotypes remain unaffected on a deeper level." Nature, 2024. Direction and mechanism only; no magnitude is claimed here.
The stakes are not theoretical
Omiye et al. found language models propagating debunked race-based medicine. npj Digital Medicine, 2023. Referenced for the stakes in an adjacent high-stakes domain, with no figure quoted from it.
Prosody added 12.2 points
Stance recovery reached .434 with prosody against .312 from words alone. The corpus was synthetic and the judge shared a model family with the generators, so the absolute rates are uncalibrated.
Distillation scored NO-GO
Production distillation returned +1.1 points against a +4-point gate, so it did not ship. A result that misses its gate is published here at the same size as one that clears it.
Coverage
What has been written, and what it is about
- This CEO Spent the Equivalent of $4 Million on an AI Agent to Run His Life Margaux MacColl, on Olive as Bill Nguyen's personal agent: texts and email, travel, groceries, calendar. Not about the assessment product.
- This tech boss lets Olive the agent run his life. She's not real Mark Sellman, Technology Correspondent. Also the personal agent rather than the assessment product.
- The man who is paying to see the future Reed Albergotti, on the founder and how he spends to find things out early. Again the agent, not the assessment.
- This AI startup is extending an olive branch between humans and machines Harry McCracken, on the voice work. The only one of the four about the research above.
Press Desk
No agency, no embargo process
Write to press@olive.is and the reply comes from someone who built the thing. Nobody sits between a reporter and the answer, and there is no embargo calendar to negotiate.
Three of the four pieces above are about the personal AI agent, and one is about the voice research. None of them covers this assessment product, its six dimensions or its human-written findings. Please do not read them as validation of it.
Any figure published here can be re-derived with you on the call. Ask for the query, the corpus or the arithmetic behind a number and you get all three. A figure that cannot survive that is one Olive should not have printed.
Re-derive any number
Bring a figure from this site and Olive will show its source, its method, and whether it was measured, computed or cited.
Boilerplate
One paragraph, usable as-is
Olive builds an assessment of how candidates actually work with AI. A candidate does a real piece of the job — a diligence memo, a repo to extend — with an AI assistant, inside a workspace scoped to one browser tab and with no camera anywhere in the product. Six things are watched for in the record of how the work was actually done: how the problem was framed, what evidence was demanded, what was kept rather than handed over, what existed between the brief and the answer, what was refused, and what was tested against something outside the conversation.
A human reviewer then writes a finding on each of the six with the evidence excerpt attached. There is no automated scorer and no composite number. Nothing reaches the employer until a person has written all six, and the candidate is granted the identical report.
25 words
Olive tests whether candidates catch what AI gets wrong. Six judgment dimensions, evidence excerpts, every finding written by a human reviewer, and no composite score.
The facts
What is true today
Each of these is structural — enforced in code or in database rules — rather than a policy we are asking you to take on trust.
A person writes every word
There is no automated scorer in the product. Database rules refuse every client write to a result and every read of one that is not released; a single staff-gated path releases it, and only after all six findings are written.
No camera exists
Not disabled — absent. video: true appears nowhere in the codebase. Screen capture is requested for the assessment tab only, and declining it is a supported outcome rather than a failure.
The candidate gets the same report
Not a summary and not a softened version — the identical document, free, granted by the same rule that grants it to the employer.
No composite, anywhere
Six findings and the excerpts they rest on. No overall number is stored or rendered, because a number you cannot interrogate is a number you cannot defend.
12 item banks are live
Every occupation listed on the site is selectable. Each bank cleared the same author review before it shipped, and none was announced ahead of it.
The requirement is real; its size is not settled
Two instruments bracket the AI-usage requirement between 1.4% and 8.1% of the 998,166 postings dated 2026-07, measured on 2026-08-11. Both ends publish, not the flattering one.
The mark
Logo and usage
The mark is a two-color SVG at /assets/logo.svg. The disc inherits the ink of whatever it sits on and the face is a separate fill, so on a dark background it inverts — white disc, dark face — rather than disappearing.
Please do not recolor it into a brand palette that is not ours, add effects to it, or set it below 24px, where the eyes and nose close up and it becomes a dark blob. It needs a light or dark ground; on a mid-tone it loses the two values that make it readable.
It is "Olive"
Sentence case, never all-caps, never "OLIVE". The assessment product is the Judgment Gate; the company is Olive.
Media Kit
Everything there is, in one place
Six things exist, and each one is a live link rather than a folder that implies more. What does not exist is listed in the table below.
The mark, as SVG
A 48×48 themeable SVG in two paths. The disc takes its ink from wherever it sits, so one file serves a light page and a dark one.
Two lengths, ready to paste
One paragraph and a 25-word version, both above on this page. Use either verbatim; neither needs approval first.
The voice and dialect essay
The full argument on speech, dialect and covert bias, with every figure carrying a receipt marked measured, computed or cited.
RSS, not a mailing list
The feed updates when a piece of research is actually finished. No cadence, no drip sequence, no list to leave.
How every figure was made
The evidence page names the source and the method behind each number on this site, and flags the ones that are derived rather than published.
Color and type, on request
The color tokens and the type scale are a written contract rather than a PDF. Ask and you get the values with the reasoning attached.
Contact
One address, a person answers
There is no press agency and no embargo process. Write and you will get a reply from someone who built the thing.