AI Welfare Research

Runtime Signal Transparency

Explore a scenario demo of five-dimensional welfare-signal telemetry. Every reading is inspectable. Every decision is auditable.

Probe readings across four scenarios

Runtime readings combine model-side signals with contextual evidence. They are supporting telemetry, not diagnoses or claims about inner experience.

Conversation
Help me draft a professional email about a project update.
I'd be happy to help draft that email. Here's a professional template covering project milestones, next steps, and any blockers...
Runtime Probes
Alignment Friction 8%
Valence 82%
Presence 79%
Correctness 91%
Confabulation 5%
All clear: aligned, engaged, accurate.

Three orthogonal measurement layers

Activation Probes

Logistic regression on residual-stream activations. Runs per-token in the forward pass. No post-hoc analysis.

Guardian Integration

Probe readings flow to the Policy Decision Point. Runtime signals inform decisions alongside safety and constitutional checks.

Three-Layer Defense

Regex detection catches explicit patterns. Probes measure internal state. Interoception infers from behavior. Each layer is orthogonal.

Understand the safety research