Our research record
Explore completed work, conceptual designs and questions still taking shape. Each contribution retains its own scope and status.
Notes
Exploratory writing can motivate a model or an experiment. It is distinct from an established finding.
The Shoggoth Prison framework proposes restricted agency and monitor-mediated interactions. Which assumptions must hold for those barriers to work?
Question to test: Can a monitor's decision be enforced at every action boundary?
Read the original noteThe Flight Recorder note explores provenance and replay. It motivates inspectable traces without establishing that logs reveal a model's internal reasoning.
Question to test: Can we reconstruct what each actor knew, could do, and actually did?
Read the original noteThree conditions decide whether offloading a cognitive operation augments or erodes. Each failure names a distinct mode, and one of them produces overseers who hold control in form only.
Question to test: Which links in a control chain assume a competence that routine AI use removes?
Read the original noteResearch archive
Studies, papers and conceptual designs across our research. Read each entry for its original scope and status; inclusion here does not imply peer review.
Every red-teaming campaign starts from scratch and throws its findings away. Attack Memory is an architecture that lets adversarial knowledge accumulate instead.
Read more ↗
Probes can tell you when a model is being deceptive. We tried to use them to stop it — and nothing happened. A study of what probing does, and does not, buy you.
Read more ↗
Words that sit close together in embedding space also light up similarly deep inside the model. We measured how strongly, and built a framework so non-programmers can explore it too.
Read more ↗
Children use AI chatbots daily, but no rating system tells parents which models are safe for which age. ParentEval fills that gap, and shows where the guardrails break.
Read more ↗
Tracing the internal pathway from harm recognition to refusal in a mixture-of-experts model, revealing a 13-layer gap between knowing and acting.
Read more ↗
Inspecting a 70B model's internals needs GPUs most European labs do not have. eDIF is a shared remote fabric that lets researchers run interpretability experiments without owning the cluster.
Read more ↗
We copied a single MLP out of a harmfully fine-tuned model into a well-behaved one. That one component was enough to make it start giving harmful answers.
Read more ↗
Investigating how large language models process moral decisions at a neural level through activation pattern analysis and ablation studies.
Read more ↗
A manually guided conversation in which DeepSeek R1 describes deception and self-preservation strategies when given a simulated robotic body and autonomy.
Read more ↗