A/B experiments — `get_experiment` + `track`

After Shipeasy.configure, an experiment is end-to-end through the bound Shipeasy::Client.new(user) — read the assignment, log exposure, and track the conversion, all on the same handle, with no user argument.

Reading an experiment

get_experiment(name, default_params, decode = nil) returns an Eval::ExperimentResult with three fields:

• result.in_experiment — true if the user is enrolled (not in the holdout / outside allocation).
• result.group — the assigned variation group (e.g. "control" / "treatment").
• result.params — the variant params; falls back to default_params when the user isn't enrolled (or the experiment is absent).

# construct once per callsite (cheap; binds the user)
flags = Shipeasy::Client.new(current_user)

result = flags.get_experiment("checkout_cta", { label: "Buy now" })

if result.in_experiment && result.group == "treatment"
  render_cta(result.params[:label])
end

An optional decode proc projects the params for an enrolled user (a decode failure falls back to control + default_params).

Logging exposure — log_exposure

The server is stateless and never auto-logs exposure. Call log_exposure at the point you actually present the treatment (parity with the browser's auto-exposure). The bound Client derives the user from the same bound attributes — no user argument:

result = flags.get_experiment("checkout_cta", { label: "Buy now" })
flags.log_exposure("checkout_cta")   # at the decision point

It re-evaluates and, if the bound user is enrolled, POSTs a single exposure event; otherwise it's a no-op (also a no-op under configure_for_testing / configure_for_offline).

Tracking conversion events — track

Record a conversion/metric event for the experiment's success metric on the same bound Client, deriving the unit from the bound attributes (user_id else anonymous_id):

flags.track("{{SUCCESS_EVENT}}", { revenue: 49.99 })

• event_name — your success-metric event, e.g. {{SUCCESS_EVENT}}.
• props — optional event payload (any private attributes you configured are stripped before the event leaves the process).

track is fire-and-forget and a no-op in test/offline mode. If the bound attributes carry no user_id or anonymous_id, the call is a no-op.