Analytics prompts: reconcile identities, duplicates, and late arrivals

Last updated: 5 Oct 202611 min read

tutorial

An event stream may retry delivery, arrive out of order, and change identity when an anonymous visitor signs in. A prompt for analysis should name the deduplication key, identity-stitching policy, late-arrival cutoff, and treatment of missing keys. The model can flag suspicious count gaps, but deterministic processing must perform the merge and record what was excluded. A unique event ID removes retry duplicates; it does not by itself establish a unique person. Keep the raw count, deduplicated event count, and eligible-user count as separate ledger lines.

Operational case

Meridian sees 5,170 raw signup-start deliveries. An event-ID pass removes 470 retry copies, leaving 4,700 start events in this synthetic dataset. The analyst then verifies that each represents one eligible user under the chosen identity policy; only then can 4,700 be called unique starts. A late completion arriving after the reporting cutoff is held for the next run or handled by a documented restatement rule, not quietly inserted into a published denominator.

Raw deliveries 5,170
Retry copies by stable event ID 470
Deduplicated starts 4,700
Identity review: 4,700 eligible unique users in this packet
Late arrivals: cutoff and restatement rule recorded

Performance and review cost

Hash-based deduplication of N events is O(N) expected time and O(N) memory. Sorting by event time is O(N log N) when ordered session reconstruction is required. Identity stitching can be much harder if keys are missing or reused, so the report must state its confidence and exclusions rather than claiming exact user counts from an arbitrary event total.

Common Mistakes

Do not equate unique events with unique people without checking identity.
Do not silently mix late events into a locked report.
Do not discard every event with a missing key without recording the exclusion.

Connected lessons

Continue with: Experiment prompts: reconcile missing outcomes by assigned unit.

Continue with: Backfill prompts: account for writes during migration.

prompt engineering

product analytics