Select an investigation to view its full RCA.
21:02:07alert 'Faultline catalog p99 latency > 1s' started firing
21:02:38ARGUS investigation inv-fcdb95f553 began
21:02:57hypothesis CONFIRMED: An injected/artificial pg_sleep(2.5) call embedded in the…
An autonomous AI SRE for self-hosted SigNoz — it investigates the alert alone, checks every hypothesis against your telemetry, and refuses to report what it cannot prove.
20 investigations recorded here1 verified$0.94
Select an investigation to view its full RCA.
“Faultline catalog p99 latency > 1s” fired on service catalog at 2026-07-17 21:02 UTC. ARGUS took the webhook, opened inv-fcdb95f553, and had a verified answer 50 seconds later.
No prompt, no operator, no runbook. Everything on this page is that run's own output, replayed from the report it wrote.
The rail lists every run recorded here: 1 VERIFIED, 13 NEEDS REVIEW and 6 DEGRADED.
A run only earns VERIFIED when a hypothesis survives a query against real telemetry and confidence clears 75%. The rest keep a badge that admits it.
ARGUS pulled the metrics, the exemplar traces and the logs, then named it: “An injected/artificial pg_sleep(2.5) call embedded in the products SELECT statement is directly causing the ~2.5s query latency that drives the p99 breach.”
It did not stop at plausible. It ran one more query back at SigNoz to check — found 'pg_sleep' in 20 matching rows — and only then wrote the finding down, at 90% confidence.
1 was confirmed against the data. 1 was refuted: the verification query came back with no matching rows. 2 could not be verified at all, because the check itself failed to run — and those are reported too, not quietly dropped.
All of them stay on the page, each with the reason it died. An agent that hid its failures could show you one confident paragraph instead. This one refuses to bluff.
Metric deltas, exemplar traces with span ids, novel log signatures — 8 of the 8 bullets deep-link straight into the SigNoz view that produced them.
Nothing here asks you to take a summary on faith. (The links resolve on the SigNoz instance that recorded the incident, not on this static page.)
1 LLM call on claude-haiku-4-5-20251001 (live), 5,700 tokens in / 1,470 out, plus 22 SigNoz queries · 306,252 rows / 33.8 MB scanned · 151 ms query time. Printed on every report, because an agent you cannot budget for is an agent you cannot run.
All 20 runs together came to $0.94, and 1 of 20 cleared verification. That ratio is on the page rather than hidden — the honesty is the feature.