Skip to the page
ResearchGuidesHelpJoin the waitlistJoin

Honest before it speaks

Whenever Eunosa says something about itself (what it remembers, what it knows about you, how sure it is), it’s designed to check that sentence against its own records before you see it.

By Meagan Victoria Angelucci, Eunosa’s founder. Last updated 3 November 2026.

  1. Idea and promises time-stamped (done)3 Nov 2026
  2. Test plan registered (current step)By 13 Nov 2026
  3. Tests runAfter registration
  4. Results publishedWhatever they show

What the promise means

The aim is that a claim the records can’t back up is blocked or corrected before it reaches you. When something isn’t on record, Eunosa is designed to say so: I don’t have a record of that. Here’s what I do have.

What’s new is the timing. Some assistants already show where a memory came from after they’ve answered. The promise here is that the check happens before the sentence is shown at all.

How we’ll test it

This part involves no people. We’ll run scripted conversations with Eunosa, and with a version that has the check switched off, then compare the two. We’ll measure:

  • how many of its claims about itself are correct against its records;
  • how many unsupported claims are caught before you would see them;
  • how often true information is wrongly blocked;
  • whether answers stay complete when something isn’t on record.1

Checking claims while a sentence is being written tends to make them more precise but can leave answers less complete, so completeness is measured, not assumed. The exact measures, numbers and analysis will be fixed in the test plan before any testing.

What would count as failing

The pass and fail lines are set in the test plan before any testing, and shown here once the plan is registered, by 13 November 2026. They won’t move after that.

Results

No results yet.

When the tests have run, the results will appear here, whichever way they go.

Sources

  1. Saxena, Bommireddy, Padia and Gaur, 2025. Generation-time vs. post-hoc citation: a holistic evaluation of LLM attribution. doi.org/10.48550/arXiv.2509.21557 Back to the text

Change log

  1. 1 Nov 2026Page published. The test plan is being finalised and will be registered by 13 November 2026.
  2. 3 Nov 2026The practice runs from building the check are published as an earlier check. They aren’t this test, and the test won’t reuse their conversations.
  3. 3 Nov 2026Correction: this page said the idea and promises were time-stamped on 30 October 2026. They were written down then, but the time-stamp hasn’t been made yet. It will be in early November, before the test plan is registered.
  4. 3 Nov 2026The idea and both promises are time-stamped privately on the Open Science Framework. The record opens when a draft paper is posted.

All research

ResearchGuidesHelpPrivacyTerms

© 2026 Eunosa