Skip to the page
ResearchGuidesHelpJoin the waitlistJoin

How we’ll know if Eunosa works

Eunosa makes two promises. Before we say either one is true, we test it with a plan written down in advance, and we publish what we find, whatever it shows.

By Meagan Victoria Angelucci, Eunosa’s founder. Last updated 5 November 2026.

The rules we hold ourselves to

The plan comes first.
We write down what we’ll measure, and what would count as failing, before any testing starts. Then we register it, so it can’t be quietly changed later.
We publish whatever we find.
A result that goes against us goes on this page too, in the same place a good one would.
Nothing is claimed until a study reports.
Until then, everything Eunosa says about itself is written as what it’s designed to do.
Some things we’ll never claim.
That Eunosa is conscious or has feelings, or that it treats mental health. Engineering results will be reported separately from any speculation.

The two studies

  • Honest before it speaksEvery claim Eunosa makes about itself is designed to be checked against its own records before you see it. We’ll test whether that holds.Plan being finalised. To be registered (published with its date, before any testing) by 13 November 2026.Read about the plan
  • More capable, and accomplished moreEunosa is designed to leave you both getting more done and able to do more on your own. A study with real people will test whether it does.Plan being finalised. To be registered (published with its date, before any testing) by 27 November 2026.Read about the plan

The work so far

  • The honesty check so far: a full reportEverything we’ve tried since 30 October: 183 practice conversations with the real voice, what the check caught and missed, what we changed, and how the honesty test is being made ready.Updated 5 November 2026. The honesty test itself hasn’t run yet.Read the report

Earlier checks

Smaller tests that shaped how Eunosa is built, published the same way: plan first, results whatever they showed, limits stated.

  • Can an ordinary laptop be Eunosa’s voice?A small AI model running on a laptop wrote good-enough replies, but far too slowly. That is why Eunosa now uses a cloud voice by default.Done. Plan locked 30 October 2026; tested 31 October 2026.See the check
  • Building the honesty check: twelve practice runsBefore testing the first promise, we built the check and tried it on practice conversations with the real voice. In the last of twelve runs no false sentence got through, and true ones held back were one sentence over the line.Done. Built and tried from 2 to 4 November 2026. This isn’t the honesty test, which comes next.See the check
  • Checking Eunosa’s memory against a published testWhat Eunosa suggests remembering was almost always true. It still gave a wrong answer about the person one time in seven, mostly by quoting a fact that had since changed. Two of five lines passed.Done. Plan and lines fixed 5 November 2026; tested the same day. A fix is being tested now.See the check

On the record

The idea and both promises were time-stamped privately with the Open Science Framework, a research registry that keeps a dated copy no one can change, on 3 November 2026. They were first written down on 30 October. Neither test plan had been registered yet. The record opens to the public when a draft paper is posted, planned for around April 2027.

Questions about the research: hello@eunosa.com.

ResearchGuidesHelpPrivacyTerms

© 2026 Eunosa