Glossary

Each term below has one meaning across the docs. An entry gives the definition and links the page that covers it.

Abstention

A declined verdict, returned where the evidence falls short of the minimum accuracy the request sets. The abstention’s kind names the check the evidence failed. See Accuracy and abstention.

Alias

A model identifier that resolves to another. a8 resolves to your organization’s latest ready horizon on the latest version, and stays on the earlier version until your teaching is refitted on a new one. See Model identifiers.

Assistant

The AI assistant you talk to over MCP, such as Claude in Claude Code. It calls A8’s tools for you. See MCP.

Content filter

The outcome of a request A8 filtered for harmful content: an outcome apart from a verdict and an abstention, reported in each interface’s own content-filter slot, such as finish_reason: "content_filter" in Chat Completions. See Known limits.

Coverage

The share of requests A8 answers. Coverage is always reported beside accuracy, which is measured over the answered requests. See Testing.

Criterion

What is being judged, stated in plain words. The plural is criteria. The system message, or the instructions field of the interface, carries it. See Writing criteria.

Earned level

An accuracy level a criterion’s evidence has shown, also called a level. A request’s minimum accuracy lands on the nearest earned level at or above it. See Accuracy and abstention.

Fit

Producing a new version of your organization’s model from the examples you taught. Each fit mints a horizon. See Teaching.

Follow-up

A user message sent after A8’s response in the same conversation. A8 answers it with the whole conversation in view. See API overview.

Horizon

The timestamp that names a fitted version, such as 20260715t140322z. See Horizons.

Interface

One of the standard APIs A8 answers: Chat Completions, Responses, Anthropic Messages and MCP. See API overview.

MCP host

The application that runs MCP servers, such as Claude, ChatGPT or Cursor. You add A8’s server to it once. See MCP.

Minimum accuracy

The accuracy a request asks its verdict to meet, from 0 to 99, 90 by default. A request sets it as the model’s tag, a8:95, and the response reports the level it was served at as min_accuracy and in its model. See Accuracy and abstention.

Model identifier

The name of the model that answers a request, such as a8-1-1:95@20260715t140322z: a family, a version, the minimum accuracy as its tag, and a horizon. See Model identifiers.

Pin

Request a specific version by its full identifier, so its verdicts hold still. See Horizons.

Precedent

A taught example cited behind a verdict. See Precedents.

Provisional horizon

The horizon the next fit mints from what you taught so far. The models list shows it with status: "provisional" until the fit completes. See Horizons.

Reasoning

How A8 reads a request, switched by each interface’s own reasoning field. On, the default, A8 reads a request however it is laid out, a grader prompt included, and usage reports the reading as reasoning tokens. Off, A8 reads the layout as written: the system message is the criterion and the latest user message the subject. See API overview.

Recall

A8’s answer to “What have I taught you about refunds?”: the taught examples most relevant to the question, word for word. See Precedents.

Reference model

The model every organization starts from, as we publish it. Call it directly as a8-1-1@reference. See Horizons.

Retract

Remove a taught example, so the next fit uses the examples that remain. See Teaching.

Subject

The thing under judgment. The user messages carry it, and a follow-up after A8’s response can change it. See API overview.

System fingerprint

The system_fingerprint token on every response. It changes whenever anything behind a verdict moves. See Reproducibility.

Teach

As a verb, send a request with the answer it should get. As a noun, one such example: “a teach”. The activity is teaching. See Teaching.

Verdict

What A8 returns for a subject: a label, a number or a JSON object in the shape the request asks for. See API overview.

Version

A fitted model, named by its horizon. See Model identifiers.

Your organization’s model

The model your teaching trains: the reference model plus everything your organization taught. It stays within your organization. See Teaching.