RESEARCH RECORD

Modular Pretraining Enables Access Control

Ethan Roland · Murat Cubuktepe · Erick Martinez · Stijn Servaes · Keenan Pepper · Mike Vaiana · Diogo Schwerz de Lucena · Judd Rosenblatt · Addie Foote · Cem Anil · Alex Cloud · AE Studio (lead); Anthropic (collaborator); independent researcher

Original publication ↗ · BibTeX · RIS

The authors report configurations approximating filtered-model performance.

Research question

Can a model selectively restrict sensitive knowledge?

Design & assumptions

Gradient-routed modules are compared with separately filtered models and alternative methods across model scales.

Findings

The authors report configurations approximating filtered-model performance.

Limits & counterevidence

Preliminary research, not a deployed safeguard; compute-normalized removal does not mean absolute capability falls with scale.

Source evidence

No structured numerical extraction is available in this reading package.

Reading scope & review

Nonoperational measurement and limitation summary only; no hazardous domain instructions extracted.

Model-assisted reading; independent evidence review pending.

Check the original source ↗