This note uses synthetic timestamps. The method comes from an independent, reproducible Python retention audit. That repository is a public submission sample, not a paid client case or evidence of a business outcome.

Pin down the interval before counting

For an elapsed-time definition, let registration be time zero. Define days 8–14 as the half-open interval:

registration + 7 days <= activity < registration + 14 days

This includes the beginning of day 8 and excludes the beginning of day 15. It is a metric definition, not a universal rule for every product. A calendar-day metric in a local timezone needs a different, explicitly agreed definition.

Here is a dependency-free Python check:

from datetime import datetime, timedelta, timezone

registered = datetime(2026, 8, 1, tzinfo=timezone.utc)
start = registered + timedelta(days=7)
end = registered + timedelta(days=14)

events = [
    start - timedelta(seconds=1),
    start,
    end - timedelta(seconds=1),
    end,
]

included = [start <= event < end for event in events]
assert included == [False, True, True, False]
print(included)

Expected output: [False, True, True, False].

If an inclusive upper bound replaces < end, the fourth event changes the result. Aggregate totals alone can hide that boundary error. Four carefully chosen timestamps expose it directly.

Do not give recent users a shorter chance to return

An account registered five days before the snapshot has not yet reached this outcome window. Treating that account as a completed nonreturner changes the denominator for reasons unrelated to the behavior being measured.

For this metric, include a registration cohort only when the observation snapshot covers its complete fourteen-day follow-up. Record the cutoff with its timezone. Keep incomplete follow-up separate from a completed window with no qualifying activity.

Also define what “active” means. A post, comment, login and purchase are different outcomes. If authored events still have visible author IDs and timestamps after moderation, decide explicitly whether those events count; do not let an API's display defaults silently choose the metric.

A missing event is not automatically no activity

Before interpreting an empty outcome as nonretention, check the extraction:

• Did each event stream reach its documented terminal cursor?

• Did a cursor repeat or stop advancing?

• Were records duplicated between pages?

• Did the snapshot stay bounded while new events arrived?

• Do reported totals agree with the collected records, or is the disagreement documented?

The linked sample records page URLs, response hashes, fetch times, cursor state and ID gaps in a manifest. A hash makes a saved response identifiable; it does not prove that the source supplied every event. Likewise, a changing API total is a discrepancy to investigate, not permission to label an incomplete export complete.

Its fixture tests cover time boundaries and a separate pagination-completion case. They can run locally, without calling the public API:

python3 -m unittest -v test_audit.py

Run that command from the downloaded sample directory. Passing fixtures verifies those cases, not the completeness of a future live export.

Before shipping a retention number, keep the metric definition, cohort cutoff, boundary fixtures and extraction manifest beside the result. Another developer should be able to explain why each included event belongs and why each excluded event does not.

Optional scoped review

Free Develop reviews one agreed CSV/JSON metric or transformation: up to two files, 10,000 rows total, and three agreed checks, with a short findings report, a reproducible script where needed, and one clarification pass. Proposed pilot price: 20 USDC on Base, prepaid after scope and acceptance criteria are agreed; delivery timing is confirmed before payment.

For a fit check, email freedevelop.studio@proton.me with subject Data review — retention note and a description of the issue. Start with synthetic or redacted examples; do not send credentials or private production data. The tutorial and linked sample are free to read and run.

Verification note: The Python example and the linked sample’s four fixture tests were run locally.