Supervision

Supervision is any expression of the standard a training run accepts as input. Four forms are defined, each distinguished by the class of signal it carries. All four resolve to the same common representation before training consumes them.

§1Definition

Supervision is the input that tells a model what distinguishes content that meets the standard from content that does not. Each form of supervision differs in the class of signal it carries — whether the standard is stated as labels, embedded in artifacts, articulated as intent, or inferred from behavior.

§2The four forms

Samples Direct Signal lives on labels attached to content. Labeled content items with positive or negative verdicts; the standard is expressed by which items are positive and which are negative.
References Embedded Signal lives inside existing artifacts. Policy documents, exemplar content, style guides, videos; the standard is expressed by the artifact itself and extracted on ingestion.
Briefs Articulated Signal lives in written intent. A structured description of the standard in natural language, covering intent, domain, exemplars, failure modes, edge cases, and constraints.
Feedback Behavioral Signal lives in observed behavior. Accumulated during scoring — content that was accepted, rejected, or corrected downstream; the standard is expressed through decisions on live content.
Figure 1. The four supervision forms. The class column names where each form's signal lives.

§3Mechanism

Each supervision form resolves to the same common representation — labeled samples on the content-trait axes — before training runs. The resolution path differs per form:

FormResolution to samples
SamplesConsumed directly; no translation.
ReferencesThe artifact is parsed and samples are extracted from it. Positive and negative items are identified based on the artifact's own structure and content.
BriefsThe brief's elements are interpreted to retrieve, generate, or elicit samples that reflect the articulated intent.
FeedbackScored content paired with downstream signal is captured during live operation and appended as new samples.

Table 1. How each form becomes samples. Downstream steps (calibration, break placement, scoring) are identical across forms.

§4Mixing forms

A training run can combine multiple forms. A brief paired with a small set of direct samples is common when articulated intent is clear but concrete anchor points are useful. Feedback supplements any base training by continuing to refine the model during operation. Training behavior is unchanged when forms are mixed — all supervision resolves to samples, which training then processes uniformly.

§5Related concepts

  • Samples — the direct supervision form and the common representation all other forms resolve to.
  • Briefs — the articulated supervision form.
  • Training — the process that consumes supervision and produces a model.
  • Roadmap — planned extensions including online learning (continuous feedback ingestion) and cold-start evaluation.
Next
Samples The direct supervision form → Briefs The articulated supervision form →
Scores are approximate — not a substitute for human judgment.
built by @onebit0fme