Glossary.

A policy in which uncertainty, missing data or failed verification blocks the action rather than allowing it.

Volcano Observatory Notice for Aviation: an official volcano-observatory communication for aviation context.

Transport of a property or ideal parcel by a velocity field; here used as a simplified wind trajectory.

A formula for great-circle distance between two geographic coordinates on a sphere.

An explicit node-and-edge network representing transmission hypotheses; it does not prove causality by itself.

A probabilistic model representing data as a mixture of Gaussian distributions; here used for latent regimes.

A method that builds intervals from empirical out-of-sample residuals under stated assumptions.

Model choice at each step using only evaluation outcomes already observed in the past.

Delay between the period a datum refers to and the time it actually becomes available.

Median Absolute Deviation: the median absolute distance from the median, a robust dispersion measure.

Standardisation based on robust statistics such as the median and MAD, reducing sensitivity to outliers.

A known way in which a system may fail, stop or produce a misleading result.

A test that removes a source or variable to measure its contribution.

Agreement between stated probabilities and observed frequencies over time.

Storage of results, metadata and execution state.

The stage that determines whether to proceed, stop or publish an output.

The stage in which calculations, rules or models are executed.

The stage in which the system checks that data and conditions can be used.

The stage in which the system retrieves data or content from sources.

An event or schedule that starts a workflow.

A check performed with controlled data without relying on a live external source.

A chronological record of system events, checks and errors.

A component that transforms a specific source into a common format.

The result a model seeks to estimate or classify.

Unintended use of information that would not be available at forecast time.

A layered model of numerical transformations used for complex relationships.

A sequence of observations ordered in time.

A method that groups similar cases without predefined categories.

A method that estimates a numerical value from available variables.

Models that produce new text, images or other content from instructions and data.

Human control and responsibility where interpretation or decision is required.

A check of presence, freshness, format, consistency and missing values.

A value or behaviour different from what the system considers usual.

Assignment of a case to a predefined category.

An estimate relating to a future time or an event not yet observed.

A mathematical or logical relationship used to produce an estimate or classification.

Methods that learn relationships from examples or previous data.

Methods enabling a system to classify, estimate, recognise or generate content.

A precise finite procedure that transforms inputs into a result.

Automatic execution of defined activities or calculations, including controls and exceptions where needed.

A sequence of activities that software performs in a defined order to transform data into a result.

A calculation that updates a point position over time by following a wind field.

A simple reference used to assess the model’s added value.

An algorithm that groups nearby points and leaves isolated ones apart.

An estimate run only after a defined source event occurs.

A variable supplied to a model.

NASA near-real-time service for thermal anomalies.

A check that recognises identical or near-identical content to prevent duplicates.

A feed-forward neural network used to classify input configurations.

An automated check that allows, blocks or skips an output.

Difference in wind direction or speed between atmospheric levels.

Distance from the mean measured in standard deviations.

Decentralised protocol used by Bluesky to create records, upload blobs and manage identities.

Controlled reconstruction of missing or previously unprocessed historical data.

Mean squared error of a probabilistic forecast; lower values indicate better overall calibration.

Trace of sources and transformations that produced an output.

Use of a declared fallback or reduced output when a modality is unavailable.

A workflow triggered or made publishable by a defined event.

Property by which a retry does not create duplicate effects.

Practices for versioning, testing, deploying and monitoring models and pipelines.

A dataset using only information actually available at the simulated date.

Metadata on sources, timestamps, version, configuration and artifacts.

Automation of repetitive activity coordinating sources, rules, files, reports and external systems.

A valid stable identifier used to create or update an AT Protocol record.

Temporal evaluation in which training and test windows move forward without mixing future data.

A change in the input distribution relative to the development period.

A self-exciting temporal model: an event temporarily raises the probability of subsequent events; used here as an event-history baseline.

A selection rule that compares modalities and baselines across folds and chooses the best-calibrated probability source for a horizon.

A temporal evaluation that advances through time while removing observations near boundaries to avoid train-test contamination.

An excluded interval between training and test that reduces leakage from overlapping episodes or features.

Fire Radiative Power: a satellite measure of thermal-anomaly radiative energy, used as a proxy with observation-coverage controls.

Parameter of the Gutenberg–Richter relationship between earthquake count and magnitude, estimated only with a sufficient sample.

Exponentially weighted moving average that gives more weight to recent observations and smooths an index over time.

A transformation that limits outlier influence, often using quantiles, median or interquartile range.

Linear regression with an L2 penalty, useful for stabilising coefficients when features are correlated.

Explicit states and transitions governing system behaviour from thresholds, direction and confirming signals.

A short publishable fact derived from an admitted source and expressed with original controlled wording.

A set of independent sources aligned enough to be treated as descriptions of the same signal.

A comparison of sources on the same domain without claiming they describe the same event.

A label that prevents the final post from using two sources considered too close or dependent.

A domain registry containing reuse permissions, risk tier and the behaviour allowed in live publication.

A generation method in which the public text is built from controlled templates and communicable facts rather than free reuse of source prose.

A measure of shared five-word sequences between public text and sources, used as an anti-reuse gate.

AT Protocol metadata that makes a textual reference clickable inside the post, for example a source link.

The count of user-perceived characters, used to keep a Bluesky post within length limits.

The PDS is the server that stores the AT Protocol record; AppView is the public view that exposes it in Bluesky interfaces.