Evaluation semantics

How Gates decides. Every answer comes with a reason telling you which step decided — use it for debugging, never for branching in product code.

Feature flags

Checked in order — first hit wins:

1. Environment switch — flag off in this environment? Serve off.
2. Overrides — a user id or email domain pinned to on/off in the dashboard. First match wins.
3. Conditions — if the flag has conditions and all of them match, serve on. If they don't match, move on to the rollout (they don't serve off by themselves).
4. Rollout — the percentage decides. 100% is on for everyone, 0% off for everyone, anything between buckets the user by userId.

reason	Meaning
flag_off	Off in this environment.
override	A user-id or domain override matched.
conditions	All conditions matched — served on.
rollout	The percentage decided.
missing_user_id	Partial rollout but no userId to bucket by.
unsupported_rule	The flag uses a rule this SDK version doesn't know — served off. Upgrade the SDK.

Experiments

Checked in order:

1. Not running → not_running
2. Off in this environment → environment_off
3. Audience doesn't match → not_in_audience (audience deleted → audience_not_found)
4. No userId → missing_user_id
5. In the holdout → holdout
6. Outside the traffic percentage → not_in_traffic
7. Otherwise → assigned, with a variant { id, name, isControl }

Every assignment also carries an iteration number. Restarting an experiment bumps it and reshuffles everyone — record it with your events so results never mix populations.

Conditions

Traits are compared as strings, the same way on every platform.

Operator	Matches when
is	Trait equals the value exactly.
is_not	Trait differs (or is unset).
contains	Trait contains the value, ignoring case.
gt / lt	Greater/less than. Numeric if both sides are numbers, otherwise text order.
is_set	Trait is present and non-empty.

Stickiness

Rollouts and assignments aren't random — they're a hash of the user id and a salt stored with the gate:

• The same user gets the same answer every time, on every platform.
• Renaming a gate never reshuffles who's in.
• Restarting an experiment uses a new salt, so everyone reshuffles — that's the point of a restart.

Fail-safety

When the SDK meets something it doesn't understand, it answers safely instead of guessing:

• A rule added after your SDK version shipped → the whole gate answers unsupported_rule (flag off, experiment unassigned).
• An experiment's audience was deleted → audience_not_found, nobody is assigned.