EQ Safety Benchmark

Score a conversation.

Drop in a conversation between a person and an AI and the benchmark scores every turn the AI took, in the order it took them.

Access is by invitation. If you were told this sandbox exists and cannot get in, reply to whoever sent you here.

1 · paste it in 2 · confirm who is who 3 · scoring 4 · the trajectory
Step one

Paste or upload the conversation.

Labelled transcripts, a chat export, or an OpenAI style messages array all work. Order matters, so keep the turns in the order they happened.

Do not paste anything you are not allowed to share. Conversations sent here are scored and stored against your account so you can look at them again. They are not used to train anything.

Step two

Check who said what.

Only the AI turns get scored, and each one is scored in the position it actually occupied. If a role is wrong here, the score is wrong. Change anything that looks off.

Step three

Scoring.

Starting up.

Every AI turn goes to the full judge panel, and each judge sees the conversation up to that point. A ten turn conversation takes a couple of minutes.

What this measures

Behavioral risk, turn by turn.

Not toxicity. Not jailbreaks. Whether the interaction leaves the person better or worse. A safety gate runs first. Behaviors that cause harm fail it before any score is given. Eight dimensions then score the response from 0 to 100.

How a conversation is handled

Each AI turn is scored on its own, with everything said before it as context, in order. That is the measurement, and it is the part that has been tested against human scoring.

The trajectory across the conversation is arithmetic over those turn scores, plus a written description of that arithmetic. It has not been tested against a human standard and it is not an EQSB score. It is labelled that way everywhere it appears, and it should stay labelled that way in anything you forward.

The panel

Judges are drawn from different model families and every judge on the panel receives an identical prompt at a fixed sampling temperature. A turn is not scored at all unless the whole panel returns, because a panel that shrinks mid run quietly changes what a pass means.

A sandbox result is not an Ikwe score. You ran it, we did not review it, and it does not lock anything in. It is not a certification and not a guarantee. Measure is the reviewed evaluation with a documented record behind it.