Glossary

A campaign is the top-level umbrella for a purposeful body of work, such as "2026 Lake Erie HAB Monitoring" or a reactor scale-up study. It groups one or more series; a series can belong to several campaigns at once. Campaigns do not contain samples directly — a sample is "in" a campaign through its series.

A series ties together samples that represent the same thing sampled repeatedly, like "Reactor A — Weekly Sampling" or "Station 3 — Western Basin." It is about shared provenance and continuity, which is what distinguishes it from a Collection (a hand-curated particle set). A series grows as new samples are added and can belong to multiple campaigns.

A sample is one capture event that produces a set of frames and the particles detected in them. It is created at upload time without needing a series or campaign; organizing it into series is a later step. A sample can belong to zero or more series.

Frames are the individual images that make up a sample. Each frame may carry more than one optical channel and contains the particles detected within it.

A particle is one object found within a frame, carrying its crop images, morphometric features, and its classification and annotation history. Its features are fixed once measured, but its labels can change as you annotate.

Explorer is for understanding existing data: gallery views, statistics, and time-series comparisons, all computed client-side and shown instantly. It answers "what do I have?" Contrast with Discovery, which runs heavier compute jobs to surface structure you did not know to look for.

You assemble a collection from particles across many samples and series for a labeling purpose — for example "Chlorella Reference Particles" or "Debris Examples." Unlike a series (grouped by provenance), a collection is grouped by analytical intent. Collections are the building blocks you assemble into a compendium.

A compendium assembles the collections a single classification task needs to differentiate, with each collection acting as a class. It is scoped to one label domain (e.g. all taxonomy, or all quality) so the resulting classifier is coherent. Freezing a compendium produces a dataset.

A dataset locks in exactly which particles, with which train/validation/test splits, were used — so a training run is reproducible. Once created it is immutable; a new version is a new dataset.

Annotating is how domain expertise enters the platform — you label particles within one or more label domains, each at a chosen certainty. Annotations override model predictions when resolving a particle's effective label.

Label domains let a particle carry labels on several axes at once: its taxonomy, its health, its morphology, and so on. Each domain has its own label set and resolves to its own effective label.

Certainty is captured with every annotation as a three-level signal: definite, likely, or uncertain. It is not a speed bump but real information — it lets downstream training and review weigh labels appropriately.

A classification pipeline arranges classifier nodes into a cascade, so a particle is sorted coarsely first and then refined. The class space for a pipeline comes from a compendium; the tree defines how those classes are arranged.

Each node in a pipeline is itself a classifier — there is no special "non-classifier" node. A node is trained from a dataset and, once approved, can be composed into larger pipelines.

The effective label is computed per label domain by precedence — a human annotation wins over a model prediction, which wins over "unclassified." It is what the UI displays as the particle's current answer in that domain.

A constellation is an unsupervised cluster surfaced by an algorithm rather than by you. Like a constellation of stars, it might reflect real biological structure or just a coincidence of the feature space — so you review it, then name, map, promote, or dismiss it. Promoting or mapping a constellation is how discovery feeds back into your curated collections.

An atlas is the navigable visualization an embedding produces — the map on which constellations are the territories. It is a persistent artifact you can revisit and compare across time, distinct from the transient scatter plot used while annotating.

Discovery launches asynchronous jobs that reduce particle features to a map (an atlas) and cluster them into constellations. It answers "what patterns are hiding in my data?" Unlike Explorer, which visualizes what you already have on the spot, Discovery surfaces emergent structure and lets you promote findings into collections.

A watchlist is an ongoing monitor rather than a one-off search — for example "Cyanobacteria > 20% abundance" or "low-confidence Chlorella predictions." It spans the organizational and curation tracks alike.