Replay-safe detections, second live run — 2 November 2026
Strict result: passed. Every scenario delivered all its distinct events. The function
collapsed every replay path to one row per (CollectorId, EventId). Each expected entity
had exactly one alert and one incident. No detection key alerted twice, and no late event
was missed. This run qualifies the content committed at 4dabe15
(hashes, deployed rule settings). The
first run failed, and its retained evidence explains the
fix. Release packages remain unsigned.
What ran
- Azure subscription
945092a0-2500-4b25-85a4-99f1c3135906,australiaeast, runbower-ws2b-20261102-003156: resource grouprg-bower-ws2b-20261102-003156, workspacelaw-ws2bwith Microsoft Sentinel. The table, DCE and DCR came fromdeploy/bicep/main.bicep. A certificate-authenticated app registration held Monitoring Metrics Publisher on the DCR only. No AWS resources were used. - The analytics content was deployed from
deploy/sentinel/analytics/main.bicep(deployment) withenableRules=true,frequencyMinutes=5andmaxIngestionDelayMinutes=120(look-backPT205M). - The collector was a local Release build of the branch, run with
BOWER_COLLECTOR_ID=ws2b-collector-01, lease 90 s, deadline 60 s andcollector_healthevery 60 s. It delivered through the loopback fault proxy. Events entered/v1/eventsas pack-shaped Windows 4625 and firewall-deny envelopes; scenarios e, f and k uploaded directly to the DCR stream (method: direct). The archive replay ran through two additional local collectors (fault and replay log). - All scenarios were started within 15 minutes of each other, so the outage, the late originals and the late replays overlapped one schedule of 5-minute rule runs.
Results
The verifier report contains exact counts from Log Analytics and
alerts from SecurityAlert. Incidents and their entities came from the Sentinel API.
| Scenario | Method | Distinct / rows | Result |
|---|---|---|---|
| a clean | collector | 37 / 74 | 1 alert and 1 incident per address. The 37 extra rows are archive replay m. |
| b1 response dropped, network back | collector, fault proxy | 37 / 49 | 12 replayed; no extra alert |
| b2 collector killed (SIGKILL) after Azure accepted the upload | collector | 37 / 49 | 12 replayed 94.7 s later after lease expiry; no extra alert |
| c response dropped at outage start | collector, fault proxy | 37 / 111 | replayed 4,099 s later, and again by restore j; no extra alert |
| d1, d2, d3 originals during a 4,029.8 s outage | collector | 111 | first ingested up to 4,039 s after their event time; each address alerted once, in the first owned window after restoration (01:54:57 and 01:55:52 runs). d2 crossed its threshold only with its late half. |
| e beyond tolerance (150 min), rewritten time (3 days) | direct | 15 / 15 | no detection alert; 1 monitor incident (host ws2b-direct) |
| f originals, f2 identical copies 75 min later | direct | 37 / 74 | 1 alert per address; the copies raised none |
| j restored queue backup (133 pending rows) | collector, sqlite3 .backup | 133 / 303 | all 133 redelivered after the restore; no extra alert |
| k crossing while the firewall rule was disabled (22.4 min) | direct | 25 / 25 | exactly one alert, catch-up-2, in the first run after re-enabling; later runs found it in SecurityAlert |
l dead-letter replay (bower queue replay) | collector, dead-letter state simulated in SQLite after acceptance | 12 / 24 | 1 alert, no extra |
| m archive replay of a, replay collector with the same id | bower archive replay | 37 replayed (HTTP 202) | collapsed with the originals; no extra alert |
| m′ the same replay into a collector with another id | bower archive replay | 0 stored | refused (collector-mismatch, CLI exit 2) |
The function was checked against the raw rows with exact counts. 631 raw rows formed 351
identities, the function returned 351 rows whose BowerCopies sum to 631, and
(CollectorId, EventId) and EventId gave the same 351. 176 collector_health rows were
in the table and none appeared in the function. The scheduled-run timing again showed
now() equal to the alert EndTime.
Simulated parts: the fault proxy simulated the ambiguous acknowledgement and the outage, which a firewall rule produced in the native pilot. Dead-letter state in scenario l was written directly to SQLite. Events came through the collector's HTTP intake or the Logs Ingestion API, not native WEF or syslog sources; native sources were qualified in the native pilot.
Not run
- RELP resends were not driven here. The RELP campaign
(
fw-20261101-111911) produced Sentinel duplicates after kill/restart that collapse byEventIdwithin one collector, which is the same key. - No multi-day soak and no load test of the queries at production volume.
Cleanup
See cleanup.json: strict audit, exact not-found codes only.