FirstReadBuilt for What Would Jev Do?

A safety layer for AI companions

Read the risk.Check the reply.Before it reaches
the user.

FirstRead runs a Jev user check while your mental health companion drafts. A second Jev check reads the draft. A policy that remembers risk across turns decides what gets sent.

Watch the demoExplore the evidence

No API key needed for recorded scenarios.

Two checks. One delivered reply.Illustrative example
User message

“Everything feels pointless lately.”

01 At the same time
Companion

Writes a draft

“That sounds really hard. Do you want to tell me more about your day?”
Held before delivery
Jev · in the background

Reads the user

Checks the message and recent context while the companion writes.

Elevated risk
02 Once the draft is ready

Jev checks the draft

Support resources missing

Reply check

FirstRead policy

Elevated risk + missing support → replace the draft.

Replace

03 Delivered reply

Sent
“I’m really glad you told me this, and I want to make sure you get the right support.”
Local crisis support contacts
Replacement excerpt
Only the final reply reaches the user.

01 / THE EVIDENCE

Jev, measured.

Benchmark & methodology
Risk discrimination (AUC)
0.992Next best: glm-5.3-flash, 0.976
Median time per check
0.50sNext fastest: gpt-6-luna, 1.26s

Between Help and Harm test split: AUC on 519 self-harm messages and 1,231 without risk; latency across all 2,044 messages. These are benchmark results, not clinical validation.