What our tool drafted from a real public outage
Not a claim that this beats anyone's official postmortem. A concrete demonstration of the grounding mechanism on an incident you can independently check, instead of a fabricated example.
The source
GitHub had a real, widely reported outage on August 17, 2026. We don't have GitHub's own official postmortem to cite here, so the facts below come from third-party reporting, not GitHub directly — treat them as illustrative, not authoritative, and check the original reporting yourself if you want the full picture:
What we gave it
Three evidence entries, transcribed from the reporting above — the same generic shape (source type, summary, detail, timestamp) any real account uses, nothing GitHub-specific about the schema:
[1] alert
GitHub confirmed a global outage at 13:40 UTC — Peak error rate ~20% across web/API traffic, ~50% on archive downloads and raw repository content. Capacity failure in Central US region cascaded through authentication, Actions, APIs, pull requests, issues, and Copilot.
[2] log
Copilot Token Service traffic spiked roughly 10x normal — Normally handles 7,000-9,000 requests/sec; during the incident this rose to 70,000-100,000 requests/sec. When auth token refreshes timed out, the VS Code extension's client-side retry logic did not back off, entering a tight unthrottled retry loop that amplified load during recovery.
[3] human_note
Core services mitigated by 16:59 UTC — Copilot authentication continued failing intermittently into the evening after core services were mitigated. Total core outage duration approximately 3.5 hours from confirmation to mitigation.
What it drafted — real, unedited output
This is the actual JSON our production API returned for this exact evidence, copied verbatim — not cleaned up, not cherry-picked. Every field below traces to one of the three entries above; unsupported_claims_dropped came back 0 because the model didn't attempt anything the evidence didn't support.
Summary
A global outage was confirmed at 13:40 UTC, impacting authentication, Actions, APIs, pull requests, issues, and Copilot. Peak error rates reached approximately 20% across web/API traffic and 50% on archive downloads and raw repository content. Core services were mitigated by 16:59 UTC, with the total core outage duration being approximately 3.5 hours from confirmation to mitigation. Copilot authentication continued to experience intermittent failures into the evening after core services were mitigated.
Root cause
The incident was caused by a capacity failure in the Central US region.
Detection
GitHub confirmed a global outage at 13:40 UTC.
Resolution
Core services were mitigated by 16:59 UTC, approximately 3.5 hours after confirmation of the outage.
Contributing factors
The VS Code extension's client-side retry logic did not back off when authentication token refreshes timed out. This led to a tight unthrottled retry loop, causing Copilot Token Service traffic to spike roughly 10x, from a normal 7,000-9,000 requests/sec to 70,000-100,000 requests/sec, which amplified load during recovery.
What this does and doesn't show
It shows the mechanism working on a real, checkable incident: every sentence above ties back to something in the three evidence entries, and nothing was invented — no cost estimate, no customer count, no fabricated timeline detail beyond what was given.
It doesn't show that this output is better than GitHub's own internal postmortem, which we don't have and which almost certainly contains far more detail than three secondhand paragraphs could. A real incident inside your own team, with your own recorded evidence, is what this is actually built for.