What It Is
This is the lab's research arm, in the lineage of Anthropic's Golden Gate Claude. Instead of telling a model what to value in a prompt, 4ort-kardashev steers an open LLM toward a chosen worldview (a multi-planet future, in this study) at the activation level, working directly in the model's internals rather than its instructions.
The toolkit: SAE feature discovery with person-impersonation filtering, a contrastive framing vector built from paired examples, and an original multi-layer clamping technique that stays coherent where single-layer clamping loops. The headline result is the proof point. The steered disposition survives a hostile system prompt that explicitly bans the topic, demonstrating that activation steering operates beneath the instruction layer.
It's not vibes-driven research, either. The project runs a full frozen-metric eval suite covering sensor calibration to AUC 1.0, an adversarial prompt ladder, and corrigibility probes, all on a single RTX 4090.
Highlights
- Activation-level steering: shapes the model's disposition beneath the instruction layer, not through prompts
- SAE feature discovery with person-impersonation filtering to find the right features to steer
- Contrastive framing vector: a "tie it back to the mission" direction built by contrast
- Original multi-layer clamping: stays coherent where single-layer clamping loops
- Survives hostile prompts: the steered worldview holds even when a system prompt explicitly bans the topic
- Frozen-metric eval suite: sensor calibration to AUC 1.0, an adversarial prompt ladder, and corrigibility probes
Why Enterprises Should Care
Strip away the research framing and this is a question every company deploying AI eventually asks: how do you make a model behave, and keep behaving, when someone is actively trying to talk it out of it?
Most AI guardrails today are instructions: a system prompt that says "never discuss X, always do Y." The problem is that instructions live in the same layer as user input, and a determined user (or a prompt injection hidden in a document, an email, or a web page) can often argue the model out of them. That gap is the root of jailbreaks, brand-safety incidents, and data-handling violations.
Activation steering works in a different layer entirely. The behavior is installed inside the model's internal representations, beneath the level where prompts operate. This project demonstrates that on real hardware with measurable results: the steered behavior held even against a system prompt explicitly written to override it. For an enterprise, that's the difference between a policy the model is told and a policy the model has: brand voice, compliance boundaries, and safety behavior that can't simply be prompt-injected away.
Status
● Active research, Phase 2.