Analysis windows & retention

Two different clocks are easy to conflate:

• Analysis window — the date range an experiment is analyzed over. In LLMJury this is the experiment's own duration. It is never artificially capped by plan tier, so a running experiment is never cut off before it can reach significance.
• Retention — how long raw events stay stored (a per-plan limit: 7 days Free, 30 Pro, 90 Business, configurable Enterprise).

When an experiment outlives retention

Only the retained tail of events is analyzable; the result is flagged retention_truncated and window_days records the window actually used. This is a storage bound, never an artificial stop — upgrading extends retention. The business-metric attribution window is bounded by retention too.

Results outlive raw data

Finalized results are snapshotted (Postgres rows + object-store snapshots), so a completed experiment's verdict persists after its raw events age out. Statistical ceilings (permutations, per-arm sample caps) are also per-plan; overrides above the ceiling are clamped and the effective values are recorded on the result.