Supervision

The guidance you give a model about what good means — samples, references, briefs, and feedback.

Supervision is the guidance you give a model about what good means. It comes in four forms — samples, references, briefs, and feedback — and every form resolves to labelled samples that training consumes.

Definition

Supervision is any demonstration of the standard a model should learn. The four forms differ in how directly they express that standard — from explicit labelled examples to behavioral signals collected during scoring. Internally they converge: each form is translated into labelled samples before the engine runs. The choice of form is about how you prefer to express judgment, not about what a model can learn.

Mechanism

Form	What you provide	How it reaches the model
Samples	Labelled examples — content with a label per trait.	Sent directly; the most explicit form.
References	Artifacts that already embody the standard — documents, exemplar content.	Read and distilled into samples at build time.
Briefs	Articulated intent — a structured description of what you want.	Synthesized into samples at build time.
Feedback	Labels captured during scoring.	Appended as samples for continuous learning.

Each form is provided through a concrete mechanism:

• Samples — sent as JSON to the REST API or MCP add_samples, or supplied as JSONL files for a file-based build.
• References — supplied as sources in a file-based build: a local or remote file, a dataset, or a document fetched by URL, distilled into samples before training.
• Briefs — the minimal specification of a trait — its poles, a question, a short description — that the build hydrates into samples.
• Feedback — a feedback map on a score request records labels for the scored content; those labels accumulate as samples and fold into the next version. This is how a ready model keeps learning — online learning — without leaving service.

Interpretation

Reach for the form that matches what you have:

• Samples when you already have labelled data.
• References when the standard lives in documents or exemplar content.
• Briefs when you can describe the standard but have no data — the fastest path to a first model.
• Feedback to keep a live model improving from real traffic.

The forms compose. A brief can seed a model that feedback then refines, and supplied samples can sit alongside synthesized ones. Cold start works from any single form: a model becomes scorable from a brief alone, before any usage data exists.

Edge cases

• Every form reduces to samples, so score-card semantics are identical no matter how a model was supervised.
• Explicit samples and references shape a draft before training; feedback is the only form that adds to a model already serving scores.

Next

Samples

The direct form, and the one every other resolves to.

Briefs

Supervision by description, synthesized into samples.