Key takeaway
Keep the assigned population and recorded exposure history distinct. Missing exposure evidence is not proof that a unit never encountered the feature.
Which event started the experiment for this unit?
An experimentation archive can preserve a chosen variant without showing whether the unit encountered the feature. A page may assign a variant while it loads, although the tested element appears only after a later interaction. If the export calls both events exposure, a reviewer cannot tell which population a reported result describes. Resolve that meaning before turning the history into an evaluation package.
Statsig describes assignment and exposure as separate events in some experiments, using an element below the initial page view as its example. Its product update describes qualifying-event filters and explicit SDK exposure logging. Those are implementation options within that system, not evidence that another company recorded either event accurately.
The owner’s immediate decision is documentary: can the archive distinguish assigned units, recorded exposures and subsequent metric events? A complete-looking scorecard does not answer that question. Start with one experiment, its actual unit definition and the logging configuration in force during the period.
Three units, three different evidence states
This hypothetical experiment assigns variant B to three invented units at 10:00. Its declared review question concerns actions after the recorded exposure. All timestamps refer to one stated time basis; the example does not test clock accuracy, causal effects or experiment results. The internal owner has retained the assignment events and the exposure events shown below.
U1 has an assignment but no retained exposure event. That is a gap in the available evidence, not an observation that the person never encountered the feature. U2 has an action before its recorded exposure. U3 has an action after it. Preserve all three histories while recording these different interpretations.
| Unit | Retained events | Completed review decision |
|---|---|---|
| U1 | Assigned B at 10:00; no exposure event; action at 12:00 | Assignment known; actual exposure unresolved; no exposed-cohort assertion |
| U2 | Assigned B at 10:00; exposure at 11:00; action at 10:30 | Action precedes recorded exposure; keep outside this stated post-exposure window |
| U3 | Assigned B at 10:00; exposure at 11:00; action at 12:00 | Action falls after recorded exposure under the declared window; no effect conclusion |
Keep both populations visible
An assigned-population question and a recorded-exposure question can require different inputs. Save both population definitions, their counts and the relationship between them. Do not overwrite assignment with the later exposure time or silently remove U1 because its history is inconvenient. An analyst needs to see how the population was constructed before deciding what a comparison means.
Statsig Warehouse Native documents an analysis flow that identifies enrollment or exposure and associates later metric data with the assigned group. It also describes metric bake windows that exclude cohorts lacking enough follow-up. A retained row count therefore needs its calculation settings: an assigned unit can be absent from a particular metric’s analysis for a documented reason.
These vendor semantics do not make every exposure-based subset a valid causal comparison. The records reviewer should identify the selection rule and ask the experiment analyst to assess it. A cleaner join or more complete log cannot substitute for that assessment. Likewise, an exposure event establishes what the system recorded under its definition; it does not prove a person paid attention.
Write the event contract beside the export
For this example, the completed event contract names the unit, experiment and variant versions; the assignment source; the exposure trigger; the metric event; and the allowed observation window. It states whether the trigger records rendering, a visibility condition or an interaction. Those meanings must come from the retained implementation evidence, not from a reviewer guessing what the event name suggests.
Link the contract to the configuration or release that produced the events. If the exposure trigger changed midway through the experiment, retain the applicable periods and identify which rows use each definition. Keep the original event and a later correction distinguishable. Ask the logging owner whether missing events, duplicate calls or a migration limit the interpretation; do not assume they occurred or assume they did not.
Choose the question the retained history can answer
The completed review holds U1’s exposure assertion, excludes U2’s action from the declared post-exposure window and retains U3 for the analyst’s further review. It approves no uplift claim. If the archive contains assignments only, describe assignment histories. If some exposure definitions cannot be recovered, narrow the period or state the limitation rather than manufacturing an event.
Use Inventory to describe the event categories and Due diligence to request the recipient’s intended analysis. Start with fictional rows. Real product activity may contain personal information, customer restrictions or proprietary features; permission to describe its structure does not authorize disclosure. VOID’s named-recipient metadata introduction is a separate decision from sample access, evaluation and licensing, and it does not certify an experiment’s validity.