§1Definition
A brief is free-form natural language organized around six elements. The training run interprets the brief and uses it to retrieve, synthesize, or elicit the samples it needs. The quality of the resulting model depends on how completely and specifically the brief describes the standard along each element.
§2The six elements
Intent
What the model is for — the decision it supports or the judgment it produces. "Score README files for their ability to earn developer trust" is more specific than "evaluate READMEs."
Domain
The content type, the context, and the audience. A narrower domain produces a more selective model. "Landing pages for enterprise technical buyers" is narrower than "marketing copy."
Exemplars of good
Qualities that the model should treat as positive evidence, described rather than enumerated. "Specific over general, earned claims over asserted ones, plain language over jargon." These implicitly define candidate traits.
Failure modes
Common ways content does not meet the standard. "Buzzword density, unsupported superlatives, buried value proposition." Failure modes define the negative side of the contrast training learns from.
Edge cases
Specific calls that the model should handle a particular way. "Technical jargon is acceptable when the target reader is a specialist." Stating edge cases in the brief prevents them from being decided implicitly at scoring time.
Constraints
Attributes the model should not optimize for. "Do not reward length; do not penalize brevity." Constraints guard against incidental signals becoming the trait axis.
Figure 1. The six elements of a brief. Each element addresses a distinct input the training run needs.
§3Mechanism
A brief is parsed into its six elements and translated to a sample set before training runs. Exemplars and failure modes define the contrast training learns from; the intent, domain, and edge cases scope the retrieval or synthesis; constraints are applied as filters on the generated sample set. Training then consumes the samples identically to any other training input.
§4Mixing a brief with samples
A brief and a small set of direct samples can be combined in the same training run. The brief carries the articulated standard; the samples anchor the contrast at concrete points. The combination is often more specific than either input alone, particularly when the brief is short and the samples are exemplary rather than exhaustive.
§5Related concepts
- Supervision — briefs are one of four forms.
- Samples — the direct form, and the common representation a brief resolves to.
- Trait discovery — uses the brief's exemplars and failure modes when proposing traits.
- Traits — the named axes a brief implicitly defines through its exemplars and failure modes.