Constraining

Type: types/definition.md · Tags: learning-theory, constraining

Constraining is committing choices in a retained artifact that narrow which readings or behaviors its consumer should treat as valid, settle disputed boundaries of validity, or rank which valid reading it should choose. This KB calls the set of valid readings under an applicable interpretation the artifact's interpretation space. Where membership is settled, constraining narrows it or ranks within it; where membership is disputed, a retained choice can first settle the relevant boundary. When a retained constraint improves the system's capacity to act, constraining is learning in Simon's sense: one of the main ways a deployed system learns without weight updates. A change that only makes an artifact easier to find, shorter, or more useful is not constraining unless it changes what is permitted or preferred, or settles a disputed boundary.

Interpretation space is relative to the applicable reference or interpretation under a named consumption path. Keep separate what the contract permits, what a consumer takes it to permit, and what the consumer produces. A weaker model's additional readings are not automatically valid, and a stronger model's consistent choice does not remove alternatives the contract permits. Under English instruction execution, the boundary itself can be graded, disputed or unresolved. The set model is useful where the relevant membership judgments are sufficiently definite; it does not supply those judgments. A constraining step is a program refinement step when the narrowed set lies inside the old one and the narrowing is checked. Reliability, speed, cost control, and reviewability may follow from constraining, but they are consequences rather than the definition.

Scope

Use the term when an artifact, convention, type, validation rule, or implementation narrows a named consumer's interpretation space. The test is semantic: the retained choice excludes or ranks lower a reading previously permitted under the applicable interpretation, or settles a previously disputed boundary. Settling a boundary need not establish that a definite old permitted set has been narrowed.

Constraining can narrow three things, and one artifact can narrow more than one:

  • content — what the artifact means: which reading of a term, rule, or instruction is valid;
  • activation — when a requirement applies: which events or situations invoke it;
  • response — what counts as satisfying it: which outputs or actions are accepted.

Methodology enforcement treats activation and response as independent axes. The narrowing need not happen inside the original artifact: when a spec's common interpretation is extracted into a function, the spec admits the same readings as before, but the function has replaced it for that operation.

Common KB instances:

  • definition notes that replace vague terms with operational meanings;
  • title conventions that require a note to state a claim rather than name a topic;
  • structured sections that assign meaning to positions in a document, such as scope, exclusions, and misuse cases;
  • schemas, validators, and scripts that give one reading formal consequences;
  • hooks and required gates that fix when a rule applies.

Two regions: interpreted and codified

The representational form of the committed choice determines what a constraint leaves open. Two questions need to be kept apart: what the artifact permits, and how reliably its consumer complies. A reading can be invalid and still be produced; that is interpreter failure, not a weaker constraint.

Natural-language constraining — definitions, title conventions, structured sections, worked examples, review criteria — can make readings invalid, not only rank them lower. Where the applicable contract requires JSON-only successful responses, a Markdown wrapper violates it even if every model run adds one. By contrast, “escalate serious incidents” can leave disagreement about whether a ten-minute outage qualifies. Recording that outages longer than five minutes require escalation settles that boundary; merely observing one handler's escalation habits does not. Several outcomes can remain permitted even after the boundary is settled, and compliance remains a separate question.

Codification — schemas, validators, route tables, code, adopted model packages — settles what is permitted within its scope by holding consumers to a reference. Fixed rules determine permitted behavior, including declared variation, external selection, and clarification steps. A consumer that departs from the reference fails to conform; behavior that conforms can still fail the original requirement. Codification is constraining that crosses from natural-language into symbolic form, not a separate mechanism.

Many constraints never need to cross. A definition note narrows a term's meaning and stays natural language.

Exclusions

Constraining is not lineage. Whether an artifact was adapted from recorded source material is carried by the lineage labels (adapted-from, operationalized-from, derived-from, abstracted-from). The two vary independently: a route table can be frozen with no source behind it, and a skill can be derived from methodology notes while staying as underspecified as they were.

Constraining is not retrieval or navigation. A description field, index entry, or backlink may help an agent find the right artifact, but it is constraining only when it also narrows or ranks an interpretation space: what the artifact means, when it applies, or what satisfies it.

Constraining is not always enforcement. A convention, example, or inline gloss can constrain interpretation without any runtime enforcing it.

Constraining is not automatically an improvement. A constraint can freeze the wrong proxy theory, overfit a narrow case, or block useful generality.

Constraining is not result selection. Retaining one LLM output fixes which result exists, not how later consumers read it.

Misuse Cases

  • Calling every improvement constraining. Added information or easier operation is not enough: the change must affect permission, preference or a disputed boundary.
  • Treating constraining as synonymous with codification. Codification is only the region where the artifact's language has a formal semantics or is close to one, as symbolic form defines it.
  • Treating a violated natural-language constraint as a weak one. "Output JSON only" makes Markdown invalid even when a model produces it; the violation is noncompliance, not evidence that the rule only ranked readings.
  • Stating how constrained an artifact is without naming the consumer.

Trade-off and relaxing

Constraining can trade generality for reliability when the committed reading fits the task: a committed reading is more predictable, but it handles poorly the cases the commitment did not anticipate. That trade makes the direction of change a design decision.

Relaxing is the reverse: replacing a constrained component with a more general one so interpretation space reopens. It is warranted when the committed reading turns out to be a wrong proxy, or when a more capable consumer handles the general case well enough. See codification and relaxing navigate the bitter-lesson boundary and operational signals that a component is a relaxing candidate.

Prior analogues

Gradual typing, formal specification, and ontology engineering each narrow interpretation in their own medium. Carnap's explication narrows a concept's meaning; the definition type follows it, so every definition note is an instance of natural-language constraining.

Relevant Notes: