GRAM Is Now Awardable on the Tradewinds Solutions Marketplace

CDAO Tradewinds Solutions Marketplace Awardable badge

AE Studio’s GRAM solution has been assessed as Awardable on the Tradewinds Solutions Marketplace, run by the Department of War’s Chief Digital and Artificial Intelligence Office (CDAO).

Tradewinds is a marketplace of solutions that have already been through a competitive assessment process. Government customers can view Awardable solutions, speak directly with providers and use the marketplace as a route to faster acquisition.

GRAM, short for Gradient-Routed Auxiliary Modules, is a new approach to controlling what a language model knows.

Today, a model typically absorbs everything it is trained on into the same set of weights. GRAM changes that. During pretraining, updates from different datasets can be routed into separate modules. If a module containing sensitive or dangerous knowledge is removed before deployment, that capability leaves with it. Nothing recoverable remains in the model’s shared weights.

The result is a model that keeps different bodies of knowledge compartmentalized, each in its own module. That could be particularly useful in environments where different users are permitted access to different information or capabilities.

“GRAM was built around a problem that matters directly to government: how to use powerful models without collapsing different compartments of sensitive knowledge into the same system,” said James Bowler, Head of Research Partnerships at AE Studio. “Being Awardable on Tradewinds gives government teams a much simpler route to evaluate GRAM and, where it fits their needs, move toward deployment.”

GRAM comes out of the research paper “Modular Pretraining Enables Access Control,” an ICML 2026 Spotlight co-authored by AE Studio researchers with collaborators at Anthropic. The work tested GRAM on realistic dual-use data spanning virology, cybersecurity, nuclear physics and specialized code, across models from 50 million to 5 billion parameters.

The method was designed to approximate the strongest form of access control, training separate models on filtered datasets, without requiring a separate training run for every capability profile. In the paper’s five-profile setting, GRAM reduced training cost by 5× compared with data filtering.

The code is public.

• Paper: Modular Pretraining Enables Access Control
• Code: github.com/agencyenterprise/modular-pretraining

Government customers can watch AE Studio’s five-minute solution video, GRAM: Verifiable Capability Access Control and Compartmentalization for Frontier LLMs, on the Tradewinds Solutions Marketplace. Viewing requires a government account. Government customers can also contact AE Studio through the comments feature on the video page to request more information or a demonstration.

James Bowler
Head of Research Partnerships, AE Studio
[email protected]

AE Studio | Los Angeles, CA | UEI JAK7ABNY7N18 | CAGE 9EK79