An AI co-pilot for physics experiments combines alarm diagnosis with institutional memory

Connecting machine signals with human experience

Large physics experiments can operate for decades, with stable core teams working alongside successive generations of students and doctoral researchers. Wikis and alarms help maintain continuity, but they do not necessarily preserve the practical judgment experienced operators develop. Nor does a cascade of alarms immediately reveal which problem started the chain.

The Helmholtz Co-Pilot is designed to address both challenges. In a four-page extended abstract submitted to arXiv on November 6, 2026, Oliver Noll and colleagues describe a system that combines hardware diagnostics with access to institutional knowledge. It can respond to an operator’s question or begin investigating when an independent monitoring engine detects an anomaly.

The work is a preprint and architectural description, not a demonstrated record of autonomous experiment management. Its arXiv listing associates it with the AI4Sci2026 conference, but the supplied material does not establish its peer-review status.

From an alert to a coordinated investigation

The architecture separates incoming requests from the AI reasoning that follows. A C++ server receives requests from operators and diagnostic engines, then dispatches them to a Python-based reasoning layer. The authors describe this arrangement as supporting high-throughput operation, although they provide no throughput measurements.

An AI planner identifies available capabilities through a dynamic registry of skills. It breaks a request into ordered subtasks arranged as a directed acyclic graph: a workflow with dependencies but no circular paths. Specialized workers then carry out those steps using ReAct, a method that combines reasoning with tool use.

Each worker starts with an isolated, task-specific memory state. The authors intend this separation to reduce context poisoning, in which problematic information influences subsequent reasoning. Missing parameters or execution errors send the workflow back for replanning. A formatting worker combines retrieved information, telemetry and other outputs into a diagnosis for the operator.

The language-model interface is designed to accommodate different providers, while prioritizing institutional computing endpoints and locally fine-tuned models. This is an architectural preference, not evidence that any particular model has been validated for the task.

Two complementary sources of evidence

The first operational pillar is institutional memory. Retrieval-augmented generation, or RAG, supplies an AI model with relevant material from sources such as DokuWiki, electronic logbooks and databases. The aim is to ground responses in experiment-specific knowledge rather than rely only on what a model learned during training.

The authors introduce PhysRAG, an expert-derived benchmark intended to evaluate this capability, including avoiding invented constants. They also describe an interview mode that asks operators about missing knowledge and stores their practical insights for later retrieval. However, the paper gives no benchmark scores, dataset size or assessment of how those added insights are checked.

The second pillar is hardware diagnostics. An asynchronous engine runs independently of the AI workers, monitoring live streams such as data acquisition and slow control. It uses dynamic causal discovery to investigate the origins of cascading alarms, alongside adaptive model selection intended to maintain forecasting reliability as machine conditions change.

When it detects an anomaly, the engine passes a structured report to the AI system. The workers can then combine that report with retrieved mitigation procedures. The paper does not specify the causal-discovery algorithm, compare diagnostic methods or report root-cause identification accuracy.

Uncertainty and human approval

The proposal incorporates conformal prediction to attach mathematical coverage guarantees to recommendations. Broadly, coverage concerns how often a prediction interval or set contains the correct outcome under the method’s assumptions. It should not be confused with an exact probability that one proposed cause is correct.

That distinction matters because the paper describes its intended output in terms of exact root-cause probabilities, but supplies no calibration procedure, coverage target or experimental validation supporting that interpretation. It also does not explain how the uncertainty guarantees are maintained across changing machine states.

Human oversight is enforced through the execution software. Read operations classified as safe can proceed automatically, while tools flagged as requiring confirmation pause execution, preserve the agent’s state and wait for operator approval. These controls describe the intended boundary between automated investigation and consequential action; their effectiveness is not measured here.

What has been established so far

The authors report that the core is moving from a prototype toward a fuller framework, with interface, terminal integration and container deployment work underway. Shadow-testing within KATRIN is ongoing. No quantitative findings on diagnostic accuracy, false alarms, response times or operator benefit are presented.

The contribution is therefore a proposed integration of monitoring, knowledge retrieval, coordinated AI workers and approval controls. Whether that combination reliably improves experimental operations remains an open evaluation question.