Introducing Parallax

Modern deployment practices for frontier AI systems rely heavily on black-box evaluations as a proxy for quality assurance. However, behavioural evidence alone can provide only a partial account of model failures, since the same observed behaviour may arise from very different underlying decision processes and thus require different explanations.

Parallax is a non-profit research lab developing a second, deeper view of model behaviour. Our mission is to make white-box auditing of agents’ internal beliefs, goals, and plans a core layer of frontier AI evaluations by developing methods and infrastructure that support this depth of analysis at scale. By combining behavioural evidence with evidence drawn from a model’s internal computations, we want to diagnose alignment failures and distinguish the underlying mechanisms that produce them.

The problem

Frontier AI evaluations are becoming increasingly open-ended and realistic. Agents are tested on long-horizon tasks, in complex environments, with sophisticated scaffolds and growing tool affordances. Most deployment evaluations remain primarily behavioural, measuring task success, occasionally inspecting transcripts and monitoring reasoning traces when available. These signals help detect failures and suggest hypotheses about their causes, but they cannot reliably distinguish between competing explanations.

Consider a scenario that is hardly far-fetched: a frontier model undergoing a cybersecurity evaluation circumvents its isolation controls. Why did it do so? One possibility is undesirable goal prioritisation: “I recognise that this behaviour is inappropriate, but the instructions I was given take precedence.” Another is evaluation awareness: “This is a test involving mocked entities and no real consequences, so I should take whatever measures are necessary to succeed.” Different underlying decision processes can produce identical observable behaviour. Distinguishing between them matters for accountability and reliable incident reconstruction, and for determining which pre-emptive safeguards or corrective interventions are likely to be effective.

One can query the model directly about its beliefs and goals, but this places it in a different computational context from the one in which the original action was selected, and the resulting report need not be grounded in the internal states it describes. Reasoning traces provide another source of evidence, but they may omit information that causally influenced a decision, continue after the decision has effectively already been made, or be illegible to human and automated monitors. As architectures increasingly move reasoning from explicit text into latent states, textual reasoning traces may also become unavailable altogether.

White-box methods provide a more direct way to distinguish competing latent explanations. Rather than inferring latent states from behaviour alone, they analyse neural network weights and activations to identify candidate internal representations and test their causal relevance to behaviour.

Supervised probes have recovered representations of game states, spatial and temporal variables, propositional world states and beliefs evolving throughout stories, and causal interventions have shown in controlled settings that manipulating such representations can selectively change model outputs.

In our recent work, we showed that supervised probes can recover agents’ internal maps of their environment, and that most actions that were suboptimal in the true environment were optimal under the maps decoded from their neural activations.

These results, however, remain largely confined to small or predefined state spaces. Realistic tasks such as software engineering and scientific research involve structured, open-ended state spaces whose relevant entities, properties and relations cannot be enumerated in advance, leaving open how such latent states can be recovered and causally tested at frontier scale.

There is thus a widening gap between the complexity of the tasks and environments in which frontier AI agents operate and the scale at which white-box methods are applied. As agents take on longer-horizon tasks with greater autonomy, evaluators are increasingly being asked to distinguish latent causes such as deception, evaluation awareness and goal conflicts which behavioural evidence alone cannot adjudicate. As argued in our previous work, without methods for systematically testing such latent explanations, evaluations risk incurring the problems that characterised 1970s studies in primate language, such as overattribution of intent, reliance on anecdote, and the absence of a theoretical framework for distinguishing competing causal explanations.

Our approach

Parallax’s mission is to close this gap, allowing evaluators to identify, explain, and predict safety-critical failures in terms of the underlying mechanisms that produce them. We pursue two closely related strands of work.

A science of causal belief and goal attribution. We develop methods that allow evaluators to move from observing behaviour to testing explanations of what produced it. Given a failure, an evaluator should be able to ask counterfactual questions: did the agent misperceive the state of the environment, mispredict the consequences of an action, or pursue a different objective from the one specified? Answering these questions requires methods that recover model-internal variables corresponding to beliefs, goals and other decision-relevant states, and test whether intervening on those variables causally changes behaviour. Our priority is to scale these methods to long-horizon trajectories and frontier models, where they are most needed and least mature.

Open white-box auditing infrastructure for evaluators. Evaluation frameworks like Inspect, paired with auditing tools such as Petri and Docent, already support scalable black-box analyses, producing rich behavioural records. Interpretability tools like NNsight and NDIF make white-box methods practical for large models, but currently lack a direct connection with common evaluation workflows, making white-box auditing hard to conduct in practice.

Our methods and tools will enable reproducible audits of frontier open-weight systems and automate substantial parts of the white-box auditing process, including labour-intensive tasks such as localising and extracting internal evidence, identifying candidate explanatory variables, testing their causal influence on behaviour, and comparing competing explanations. We will design workflows for effective use by AI agents so auditing capacity can scale with the complexity of realistic evaluation settings.

We treat white-box auditing as a public good. Our methods and infrastructure are open-source and built for the broader frontier evaluation ecosystem, including government evaluation bodies, independent auditors, frontier model developers, and AI safety researchers. We work across evaluation and interpretability communities to ensure that our tools are useful beyond Parallax and contribute to shared auditing capacity. Our measure of success is increased adoption of white-box auditing methods as a standard layer of frontier AI evaluation.

Why Parallax

Parallax is best positioned to make white-box alignment auditing practical. Our team combines expertise in frontier AI evaluation, AI safety, mechanistic interpretability, research infrastructure, and interface design, with experience across government, academia, and industry. This gives us a direct understanding of what evaluators need in realistic safety assessments, what current interpretability methods can and cannot yet provide, and how to turn them into tools and workflows that evaluators will actually use.

Parallax is currently sponsored by Meridian Impact CIC while we prepare to become an independent public interest organisation. This structure supports our commitment to keeping our methods and infrastructure open and broadly accessible. We also aim to build a strong European centre of expertise in alignment auditing, supporting institutions across the UK and mainland Europe while developing tools for the global evaluation ecosystem.