2026-05-30 · Docs
Experiments — multivariate / statistical readout / winner auto-promote / `experiment.assigned` event retracted
- Audit of `feature-flags/service.ts` + `featureFlags` schema + `createFlagSchema` surfaced 4 over-claims. (1) "Flag types — boolean, string, JSON, multivariate" — Zod enum is `boolean / string / number / json`; no `multivariate` value. (2) "Statistical experiment readout via Analytics — funnels + retention auto-cohorted by variant" — no statistical engine; `flag_evaluations` table records variant assignments but auto-cohorting doesn't exist. (3) "Winner auto-promote to In-App Messages + Push A/B variants — same statistical engine" — fabrication. (4) "Workflow trigger — `experiment.assigned` becomes a first-class event" — `recordEvaluation` writes to `flag_evaluations` table only; nothing is written to the `events` table.
- Real shape: CRUD + toggle endpoints, evaluate + evaluate-all, rule engine (max 20 rules per flag, each `{ audienceId?, percentage?, value? }`), hash-bucket per-entityId sticky rollouts, per-environment toggles, audit log on toggle, evaluation log in `flag_evaluations` (fire-and-forget), SDK helpers on Web + RN.
- LaunchDarkly comparison rewritten honest. 4 new rows specifically mark statistical readout (❌), winner auto-promote (❌), scheduled flag changes (❌), `experiment.assigned` first-class event (❌). whyStay LEADS with the missing features. Pricing column adds LaunchDarkly Foundation's $12/seat + $10/1K MAU.
- Real wins kept + emphasised: audience-targeted rollouts in the same tenant as Customers, sticky-by-user hash bucketing, per-environment toggles, kill-switch + audit log, evaluation log for BYO experiment readout via SQL join on `entityId`.