Custom probes

Probes trained for your model and your policy

Enterprise activation probe training for the open-weight or proprietary models you host, against your own risk categories, with consulting from the team behind the research.

Your model
Open-weight or proprietary, with access to activations
Your policy
Probes trained on your own risk categories
<0.2%
Typical probe size relative to the base model

Enterprise training

Probes trained on your model’s activations

Your taxonomy

Categories defined by your policy, not a generic list

Consulting

Direct access to the researchers who build the probes

Measured results

Evaluated on your model before you deploy

1. Define the policy

We turn the behaviors you need to catch into clear, testable probe targets.

2. Build the data

We assemble labeled examples together, scoped around your data-handling constraints.

3. Train and evaluate

We train probes on your model’s activations and report how they perform.

4. Deploy and refine

We help you integrate probes into serving and adjust them as your policy changes.

Deployment options we build with you

Available through an enterprise engagement, scoped to your stack.

Real-time moderation API

Probe scores integrated into your generation service.

CI / PR gate

Probe-based checks on AI-generated diffs before merge.

IDE inline flag

Risk signals surfaced inside your coding assistant.

In-house training SDK

Probe training set up for your own team and models.

On-prem / air-gapped

Deployment scoped to your own infrastructure.

Train a probe for your hardest policy

Tell us what you need to detect and which model you run.

Talk to us