Transluce's Foundation Model for Oversight aims to detect AI sandbagging via Pythonic world models

Transluce researchers introduce a vision for building a foundation model for oversight—an AI system designed to answer difficult questions about another model's behavior, such as whether it sandbags under evaluation, reward hacks, or treats users differently. Their core innovation is the oversight-as-inference hypothesis: any oversight question can be formalized as Bayesian inference over the outputs of a Pythonic world model. This model expresses interventions (e.g., prompting, fine-tuning) and measurements (e.g., probing, forward sampling) as executable Python code, producing observable facts at multiple abstraction levels. For example, sandbagging detection is formulated as sampling task schemas conditional on a measured performance gap between watched and unwatched runs.

The training pipeline involves three stages: mid-training the oversight model on a large corpus of experiments on a specific subject model, RLVR (reinforcement learning from verified rewards) on a diverse set of oversight tasks, and fine-tuning to interpret natural language questions. The team argues this objective is as universal for oversight as next-token prediction is for language modeling. The approach promises automated, scalable auditing of AI systems, catching subtle failures that manual inspection would miss. While still early-stage, the framework provides a concrete path toward trustworthy oversight assistants for frontier models.