Who builds this

If a system is going to return a verdict about whether someone really said something, you should be able to see who built it and what they know. Anonymous detection is a contradiction.

A short page, for a reason

Most team pages in this category are a grid of four to eight faces with titles like Head of Machine Learning and a sentence about a passion for trustworthy AI. Some of those grids are real. A fair number are stock photography and job titles that describe an ambition rather than a payroll.

Since the entire argument of this product is that you should be able to check what it tells you, the team page is the last place to start rounding up. So: Lacewing Technologies is small, Truthring is pre-launch, and the roster below is one person because one person is what has been published. If that grows, the page grows, and the growth will be datable.

What a small team changes, and what it does not

It changes some things honestly. Fewer people means slower coverage of new generation systems, a narrower support window and less capacity to answer a hard question quickly. Where that affects you, it belongs on the limitations page rather than in a footnote here.

It does not change the standard for evidence. An error rate measured badly by a large team and an error rate measured badly by one person are equally useless. Sample sizes, held-out clips, audio conditions and dates are published on the accuracy page and in the evaluation dataset so the work can be judged on how it was done rather than on how many people did it.

It also does not license the usual small-company shortcut, which is to borrow credibility from somewhere adjacent. Nobody outside Lacewing has audited this method. No academic group has reviewed it. No standards body has certified it. Those are three separate absences and each one is stated in full on the pages where it matters.

Advisers and outside review

There are none to report. No external adviser, academic collaborator or reviewer has been engaged, and no third party has assessed the detection method, the evaluation data or the security posture.

If that changes, what gets published is the reviewer’s name, what they actually looked at, when, and what they concluded, including anything unflattering. A named person attached to a defined scope is a credential. A phrase such as reviewed by industry experts is not, and you should treat it as an absence wherever you see it, here or anywhere else. [VERIFY: external reviewers or collaborators, if any are engaged before launch]

Who to reach when you disagree with a result

Verdicts get challenged. That is the correct response to a probability, and there is a person on the other end of it rather than a form.

Send the reference code from your result to review@aivoicedetctor.com [VERIFY: mailbox not yet created]. You will be told which model version produced the verdict, which analysis passes drove the score, and what the published error rate is for that audio condition. If the answer was wrong, the clip goes into the next evaluation set and the correction appears in the method changelog with a date on it.

Because the team is small, a response time is not being promised on this page until one can be held. [VERIFY: realistic response time for challenged verdicts] Everything else that has its own route, including press, security and corrections, is listed on contact.

Who writes what you read here

Team and authorship are different questions, and this site keeps them apart. The authors page sets out who writes the articles, who reviews them, and how a page is marked when nobody has put their name to it. Expert review is recorded only where a named person genuinely reviewed the material, which at present is nowhere.

Questions