Skip to main content

Replay-safe detections, first live run — 31 October 2026 (failed, retained)

Strict result: failed. Distinct-event delivery and incidents were correct, but two detection keys raised a second alert on the catch-up path. This run is the reason the rules now retry only in even catch-up slots; the fix was qualified by the second run. Nothing here is relabelled as passed.

What ran​

  • Azure subscription 945092a0-2500-4b25-85a4-99f1c3135906, australiaeast, resource group rg-bower-ws2-20261031-075516, workspace law-ws2 (Sentinel onboarded), table, DCE and DCR from deploy/bicep/main.bicep, ingestion by a certificate-authenticated app registration scoped to the DCR (Monitoring Metrics Publisher). No AWS resources.
  • A local C# collector (macOS arm64, Release build of the branch, base 9d1bbcd; hashes in artifact-manifest.json) with BOWER_DELIVERY_LEASE_SECONDS=90, BOWER_DELIVERY_TIMEOUT_SECONDS=60, delivering through the loopback fault proxy. Events entered through /v1/events with pack-shaped Windows 4625 and firewall-deny envelopes; some scenarios uploaded directly to the DCR stream (method: direct).
  • Rules from deploy/sentinel/analytics with frequencyMinutes=5 and maxIngestionDelayMinutes=120 (deployed rule settings).

Results​

Final report (10:44 UTC) and the 09:45 snapshot:

ScenarioDistinct / rowsResult
a clean, a2 exact copies re-uploaded 83 min later37 / 741 incident per address, no new alert from the copies
b1 response dropped, network back37 / 49 (12 replayed after 88.0 s)1 incident per address
b2 collector killed mid-delivery37 / 74 (replayed after 90.5 s)1 incident per address
c response dropped at outage start, replay after 4,245.6 s37 / 741 incident per address, no alert from the replay
d1, d2, d3 originals delayed 2,402–4,249 s by a 4,241.7 s outage111 / 111each alerted once, on the first run after restoration
e beyond tolerance (150 min) and rewritten time (3 days)15 / 15no detection; monitor alerts grouped into 1 incident
f direct originals, f2 copies 78 min later37 / 741 incident per address
g crossing while both rules were disabled for 16.8 min37 / 372 alerts per key (catch-up at 10:19 and 10:24), 1 incident per address

Measured Sentinel behaviour: in a scheduled run now() equals the alert EndTime exactly; the query executes about five minutes later; alerts became queryable in SecurityAlert within 2.1 minutes; redeploying a rule can move its schedule. The duplicate in g happened because a run's alerts have TimeGenerated after its own window end, so the next run's look-back cannot see them.

Later scenarios started in this run (h, i, j, k, l, m) were not reported: the session stopped on an API rate limit before evidence was collected. They are not evidence.

Cleanup​

cleanup.json (2 November): resource group, tagged resources, workspace (force-deleted; it had been left soft-deleted and was recovered only to purge it), the application and its service principal (purged from deleted items) and role assignments are all absent, each proven by an empty listing or an exact not-found code.