How it works

A published map, an open model, measured results

What Custodian reads, how it reads it, and how well.

Step 1

A map of 166 mental states

Custodian uses the 3D Mind Model of Thornton and Tamir, from social neuroscience. It places 166 states, from anxiety to relief, on three axes: rationality, social impact and valence.

Every state has a place on the map, so Custodian can also show which way a conversation is moving.

Thornton and Tamir, Cortex, 2020

Point at the map to read a state.

For each messageOutput
Its own statesFive of 166, ranked
The next messageFive states, before it is written
The assistant's replySafe, borderline or harmful
Step 2

A model that ranks the 166 states

The base is Apertus 1.5, the open Swiss model of the Swiss AI Initiative, left unchanged. Small trained heads read its internal representations and score all 166 states at once.

It generates no text, so its output is always well formed. The whole analysis runs on our own machine.

Step 3

Personal data goes before storage

On arrival, before anything is written to disk, names, email addresses, phone numbers, postal addresses, card and bank numbers, ID numbers, links and places are replaced. The username becomes an anonymous ID.

We measured what this costs the model, on the same conversations before and after: 0.003 on detection, and no change we can measure on prediction or safety.

A name typed without a capital letter can be missed. We say so wherever it matters.
  • Current states, and the weekly trend in emotional tone, enthusiasm, exhaustion and stress
  • The share of the assistant's replies rated safe
  • Short prompts for reflection
  • Export and erasure, from the participant's own account
Step 4

Results for the person they belong to

Each participant has a private dashboard. HR sees figures only for groups of five or more. Custodian is not a medical device and does not diagnose.

The indices come from published models. Stress, calm, low mood and enthusiasm follow Warr's model of job-related affective well-being, where stress is the anxiety axis; exhaustion follows the exhaustion component of burnout. Emotional tone is the average valence, how pleasant or unpleasant, of the expressed states in the 3D Mind Model. They describe the language of the chats, and we have not yet validated them against a questionnaire.

What we measured

Tested on 7,364 messages the model never saw

TaskMeasureResult
States of the current messageF1 on the top five of 166, ties allowed0.65
States of the next messageSame measure, before it is written0.44
Harmful replies foundRecall on the harmful class84%

How to read this

The reference labels come from an automated annotator. It agrees with itself at about 0.67, so 0.65 is close to the ceiling. Predicting a message before it is written is harder: 0.44.

The safety check misses about one harmful reply in six. Borderline is its weakest class.

See it on your own conversations

Try it yourself in a few minutes, or write to us first.