Saltar al contenido principal

refine

Since v0.1.0 · Used inside a declaration

Grammar​

# Flow-level form (sibling of step):
refine <Target>
refine <Target> { strategy: <strategy_name> }   # optional body skipped structurally

refine iteratively improves a candidate output. Where validate halts on failure, refine loops: it asks the backend to critique its own output, applies the critique, and re-emits — until either the predicate clears or the iteration budget is hit.

refine is the canonical "propose → self-critique → revise" pattern surfaced as a typed primitive. The strategy can be backend-driven (LLM self-reflection), rule-driven (apply declared rewrites until fixed-point), or composite.

Surface​

refine is nested — it appears as a flow-step sibling of step, not as a top-level declaration.

flow DraftEmail(brief: EmailBrief) -> Email {
    step Compose {
        given: brief
        ask: "Draft the email."
        output: Email
    }
    # Iteratively improve the draft until the tone validator clears.
    refine Compose.output
    step Send {
        given: Compose.output
        ask: "Render the final email."
        output: Email
    }
}

refine <Target>​

A single identifier naming the value to refine. Typically a previous step's output (<Step>.output) — refine operates in-place on the flow scope: after refinement, <Step>.output carries the refined value.

refine <Target> { strategy: <name> } (optional body)​

A braced block may follow the target. The body is reserved for explicit strategy binding from the closed catalogue (self_reflection, propose_critique_refine, rewrite_until_fixed, grow_until_threshold). Today the parser skips the body; the runtime uses the default strategy registered for the target's declared type.

Iteration budget​

refine carries a max_iterations budget controlled by the surrounding run or agent:

  • Inside a run-level flow: run … effort: high raises the cap; effort: low lowers it. Defaults: 1 (low), 3 (medium), 6 (high), 10 (max).
  • Inside an agent: bounded by the agent's max_iterations: field.

Budget exhaustion does NOT halt the flow; the refine loop exits with the best candidate so far, audit row refine:budget_exhausted is emitted, and execution continues. This is the soft-fail discipline — refinement is a quality boost, not a hard predicate.

Runtime behaviour​

refine at flow level produces a typed RefineStep IR node. On execution:

  1. Fetch the target value from the flow scope.
  2. Loop: a. Run the strategy's critique step. b. If critique reports "good enough" → exit, write back. c. Else apply the critique → new candidate → goto (a).
  3. Audit row records (target, iterations, final_score, exit_reason).

What this primitive is NOT​

  • Not a halt. Refine is soft-fail by design. Use validate for hard predicates.
  • Not a free retry. Refine has an iteration budget; it consumes tokens + cost. Declare deliberately.
  • Not mutation of the source step. Refine operates on the flow-scope binding of <Target>, not on the step's declared body.
  • Not a substitute for human review. When the iteration budget is exhausted without convergence, the right next step is often escalate (human-in-the-loop), not "trust the best-so-far candidate".

See also​

  • axon://primitives/validate — hard-fail counterpart.
  • axon://primitives/probe — observe-only counterpart.
  • axon://primitives/anchor — run-level continuous predicate.
  • axon://primitives/reason — declares HOW the model thinks; refine declares HOW the model REVISES.