Glossary

Methods that use or learn a causal model of a system to answer interventional and counterfactual questions, not only to predict from association.

Algorithms that search data for graph structure consistent with it, under assumptions such as causal sufficiency and faithfulness. Domain knowledge, such as the order of events in time, fixes what the data alone cannot settle.

A directed acyclic graph where nodes are variables and directed edges represent direct causal influence.

Estimating causal effects from data under explicit assumptions, typically encoded in a causal graph or structural model.

Correlation is statistical association. Causation means intervening on one variable changes the distribution of the other.

A variable with two incoming edges on a path. Conditioning on it opens the path and can induce spurious association.

A variable that causally affects both the treatment and the outcome, creating a non-causal association between them.

A query about a specific unit or event under a hypothetical intervention, holding its background conditions fixed.

Quantifying the causal effect of a treatment on an outcome, for example as an average treatment effect, under the identification assumptions of the causal model.

Whether a causal quantity can be uniquely computed from the observed data distribution under the stated assumptions.

Setting a variable to a fixed value, written do(X = x), which removes the influence of its usual causes.

Pearl's hierarchy of association, intervention and counterfactual queries. Each rung requires stronger assumptions than the one below.

A variable on a causal path X → M → Y that transmits part of X's effect on Y.

Observational data is collected without manipulation. Interventional data comes from controlled changes such as experiments or replays.

A robustness test of an estimated effect, for example adding a placebo cause or a random common cause, to see whether the estimate changes when it should not.

Often implies a single cause. In agent systems outcomes usually have several contributing factors.

An association that does not come from a causal path between the variables, for example one created by a confounder.

A set of structural equations, one per variable, defining it as a function of its direct causes and an independent noise term. Supports interventional and counterfactual queries.

A continuous-optimization approach to structure learning that enforces acyclicity with a log-determinant constraint.

An open-source Python library that structures a causal analysis as four steps: model, identify, estimate and refute. Auga's causal engine uses it for effect sizes.

An open-source Python library of estimators for heterogeneous treatment effects, such as double machine learning and causal forests. Auga's causal engine uses it for effect sizes.

Fast Causal Inference: an extension of the PC algorithm that tolerates latent confounders and selection bias. It returns a partial ancestral graph rather than a DAG.

Linear Non-Gaussian Acyclic Model: it assumes linear relations and non-Gaussian noise, which makes edge directions identifiable from observational data.

A constraint-based causal discovery algorithm. It removes edges using conditional independence tests, then orients the rest from collider structures and orientation rules. It assumes no unmeasured common causes.

A causal discovery algorithm for time series. It combines a PC-style selection of conditions with momentary conditional independence tests, and also finds links within the same time step.

An information-theoretic measure of directed information flow between time series. It shows predictive flow, not by itself a causal effect.

A system where an LLM selects actions, such as tool calls, in a loop, based on a goal, its context and the results of earlier actions.

The input to the model at a given step: instructions, conversation history, retrieved documents and tool results.

The degree to which a model's output is supported by the retrieved context it was given, as opposed to being fabricated.

A check applied to inputs, outputs or actions that blocks, modifies or escalates them based on a policy.

Output not supported by the model's context or by fact.

The task goal as specified to the agent, and the final observed result of execution.

A neural network trained to predict tokens, used as the reasoning and generation component of an agent.

A model prompted to score outputs against a criterion, such as hallucination or groundedness. It is an evaluation method with its own error rate.

An open protocol for connecting LLM applications to external tools and data sources. In Auga's architecture, MCP clients are external agents that use the API.

A system of agents that exchange messages or shared state to complete a task. Failures can propagate between agents.

The input given to a model and the output it returned for one call.

Retrieval-augmented generation. Relevant content is fetched from an external source and inserted into the model's context.

A structured function invocation emitted by the model, executed by the runtime, with the result returned to the model's context.

Systematic measurement of model or agent output against criteria, datasets or reference answers.

The ability to understand a system's internal state from its external outputs, mainly telemetry.

Observation is recorded evidence. Inference is a conclusion derived from evidence under a model. Hypothesis is a candidate explanation awaiting testing.

A vendor-neutral standard and toolkit for generating and exporting traces, metrics and logs. Auga uses OpenTelemetry traces as input.

Data such as a name, an email address or an ID number that identifies a person directly or in combination with other data. Regulations such as the GDPR restrict how it may be processed.

A timed unit of work within a trace, with attributes, events and a parent relationship.

Logs, metrics and traces emitted by a running system.

A tree of spans describing one end-to-end execution, such as a single agent task.

The service that reads the extracted variables from the database, runs discovery and estimation algorithms, and grades every edge before it is trusted.

Observability extended with causal modeling, so that traces can answer interventional and counterfactual questions, not only descriptive ones.

A typed value derived from trace data under an allowlisted schema. It is used as a node in the causal graph.

The ten criteria Auga uses to tag laws by what they ask for people affected by automated decisions.

A view of the application organized around one question. The current lenses cover the EU AI Act, the Oversight Inbox, cost, operations and agent health.

Tamper-evident, time-ordered records of system actions and decisions, sufficient to reconstruct and review them.

Regulation (EU) 2024/1689. It sets obligations by risk category and applies in phases.

Explainability gives human-understandable reasons for specific outputs. Interpretability concerns the internal mechanisms of a model. Auga works mainly at the system level: explaining agent behavior from evidence, not model internals.

A category under the EU AI Act subject to requirements on risk management, record-keeping, transparency and human oversight.

Measures that let humans supervise an AI system, interpret its output and intervene or stop it, as required for high-risk systems.

The ability to link an output to its inputs, intermediate steps and responsible components.