classify

A typed classification as one graph decision, after policy admission.

Each operation version fixes its pure wire and receipt bounds. The request callback supplies current HTTP Gun and classification settings after policy admission, including approval and recovery. Receipts retain native answers, distributions, models, usage and protocol evidence; replay needs no network or credentials. Confidence is concentration evidence, not the probability that a decision is correct.

Types

Call

opaque </>

pub opaque type Call

Values

call

</>

pub fn call(
  client: http_gun.Client,
  config: classify.Config,
  model: String,
  state: value.Value,
) -> Call

Live request context; it is never written to the graph record.

decision

</>

pub fn decision(
  identity: run.DefinitionId,
  input: codec.Codec(input),
  questions: question.Batch(answer),
  wire: classify.Wire,
  request: fn(context, input) -> Call,
) -> @internal Operation(
  context,
  input,
  classify.Outcome(answer),
  operation.Kind,
  operation.Recovery,
  operation.Invocation,
  operation.Error,
)

request runs only after policy admission and should be pure. Derive live settings from its fresh context; the wire and questions retain their meaning for this operation version. A different protocol needs a new version. Live byte limits above the wire’s receipt bounds fail before credential access or network I/O. Lower live limits do not affect stored receipts.