API: Experiments

Viewer-readable newest-first keyset page. The cursor is opaque and project-bound; using a cursor from another project is rejected. The cursor is rejected above 4096 bytes before base64 decoding. The request has a 10-second deadline plus project rate and concurrency limits.

Stores analysis metadata only. Expected weights are used for the sample ratio mismatch check and never allocate or bucket traffic. The owner must be an active member of the project's organization.

A foreign-project id is concealed as 404.

The project id and experiment key are immutable.

Viewer-readable semantic POST. `from` is inclusive and `to` exclusive; both are mandatory UTC instants, the range is at most 400 days and must stay inside the stored experiment dates. Exposures come only from the canonical `experiment_exposure` event with an exact `props.experiment_key` match and are capped at 100000 rows. Exposure `props` are capped at 4 KiB per row and 16 MiB for the result; actor ids are capped at 256 bytes and fail closed using their full pre-truncation length. `allocation_unit` must be omitted or equal `actor_id`; arbitrary units are excluded and counted under the `unsupported_allocation_unit` warning instead of merging identities. Count/conversion metric queries never select raw props. Metric events use the same query-engine Filter AST, saved-segment resolver and exact actor membership as report queries. Across the primary and at most ten guardrail metrics, one readout is capped at 200000 materialized event rows, 64 MiB of result data and 64 MiB of repeated actor-array data; each ClickHouse query also has explicit result, memory and rows-read limits. All routes have project rate/concurrency limits; readout has one limiter acquisition and a 30-second compute deadline. Revenue is the signed numeric `props.revenue` convention (refunds remain negative). The server-side EnabledRevenue policy makes this available to every project independently of the still-unactivated standalone P4 `experiment_readout` commercial proposal; revision 3 did not add that capability. Every correlation/uplift result permanently carries `meta.estimate=true`.