Rules

Belay.configure { |config| config.rule_approval_threshold = 3 } # default

A gate approved consistently, and without edits, enough times learns a rule - a future occurrence with the exact same (input, outputs) and the exact same when: condition then auto-settles without ever pausing for a human. Nothing needs to be configured to opt in; it happens automatically from ordinary approve! calls, governed by Belay.config.rule_approval_threshold (default 3).

Learning

Only an UNEDITED approve! (no edited_payload:) counts toward learning - editing the output is a human saying the gate’s own answer wasn’t quite right. A single reject! invalidates the matching rule immediately and permanently: no tolerance for “mostly consistent,” and no re-learning afterward for that exact occurrence, no matter how many approvals arrive later.

Auto-settling

An auto-settled gate’s step row looks like an unedited human approval (status: "succeeded", output: nil), except rule_applied_id is set and there’s no Belay::Approval row - no human decided that one.

This is the one place in this gem that bypasses a human decision point entirely, so matching is scoped as tightly as possible: a rule learned under one when: condition cannot match a gate now running a different one, even with identical input/outputs, even if the gate kept the same name. See DESIGN_DECISIONS.md #18 for why that distinction needs its own mechanism, separate from workflow_version, and for the honestly-documented residual limits of that mechanism.

See gates.md for the rest of the approve!/reject! mechanics this builds on.