1. Pull the official incident history from the Atom feed
curl -s https://status.claude.com/history.atom -o /tmp/claude_history.atom
wc -c /tmp/claude_history.atom # observed: 36168 bytes on 2026-09-23
2. Extract the Sep 22 incident entry and its full update timeline
python3 - <<'EOF'
import re, html
data = open('/tmp/claude_history.atom').read()
# Find the incident entry for 2026-09-22 (incident id 7g1qpkyz5gxh)
entries = data.split('<entry>')
for e in entries:
if '7g1qpkyz5gxh' in e:
m = re.search(r'<published>([^<]+)</published>', e)
print('Entry published (UTC):', m.group(1) if m else '?')
content = html.unescape(re.search(r'<content type="html">(.*?)</content>', e, re.S).group(1))
# each status update appears as '<small>Sep 22, HH:MM UTC</small> <strong>STATE</strong> - text'
for upd in re.findall(r'<small>([^<]+)</small><br> <strong>([^<]+)</strong> - ([^<]+)', content):
print(f'{upd[0]:<28} {upd[1]:<16} {upd[2][:110]}')
EOF
Expected output (verified 2026-09-23): the entry published 2026-09-22T02:35:23Z; the update stream should read:
Sep 22, 00:57 UTC Investigating ... elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5.
Sep 22, 01:17 UTC Identified ... identified the cause ...
Sep 22, 01:35 UTC Update ... Fable 5/5.1 and Mythos 5/5.1 normal ... Opus 5 remaining ...
Sep 22, 02:11 UTC Monitoring ... success rates return to normal ...
Sep 22, 02:35 UTC Resolved ... Impact occurred from 5:50pm PT / 00:50 UTC to 7:10pm PT / 02:10 UTC.
3. Compute the duration from the primary record
python3 - <<'EOF'
from datetime import datetime, timezone
start = datetime(2026, 9, 22, 0, 50, tzinfo=timezone.utc)
end = datetime(2026, 9, 22, 2, 10, tzinfo=timezone.utc)
print('Impact window (primary record):', (end - start).seconds // 60, 'minutes')
# Vs. the incident lifecycle (Investigating post to Resolved post):
life = (datetime(2026, 9, 22, 2, 35, tzinfo=timezone.utc) - datetime(2026, 9, 22, 0, 57, tzinfo=timezone.utc))
print('Investigating->Resolved posts:', life.seconds // 60, 'minutes')
EOF
Expected: 80 minutes of impact; 98 minutes from the Investigating post to the Resolved post. Conclusion to record: contrary to "hours of elevated errors"/"degrading through the day" framings, the primary record shows a sub-2-hour event fully closed at 02:35 UTC on Sep 22.
4. Cross-check the independent numbers against the primary record
Checklist against sources/S13.md:
| Independent claim | Source | Against primary record | Result |
|---|
| Impact 00:50–02:10 UTC, ~80 min | LavX News | Identical to status page resolution note | PASS |
| Incident opened 00:57 UTC | AI Weekly (01:57 UTC alert) | Matches "Investigating" post | PASS |
| Opus 5 recovered last | AI Weekly / BFWAI | Matches 01:35 UTC update ("Opus 5 ... remaining") | PASS |
| Fable/Mythos recovered before Opus | Sunday Guardian | Matches 01:35 UTC update | PASS |
| >1,600 Downdetector reports | Sunday Guardian, Moneycontrol | Not in primary record (aggregate user self-reports only) | NOT VERIFIABLE from primary — label as INDEPENDENTLY VERIFIED at aggregate level only |
5. Verify the "no relapse" claim (post-incident state)
curl -s https://status.claude.com/ | grep -o "All Systems Operational" | head -1
curl -s https://status.claude.com/history.atom | grep -c "<entry>" # count incidents in feed
# Confirm no second incident entry exists for Sep 22 beyond 7g1qpkyz5gxh / none for Sep 23
Observed 2026-09-23: current status "All Systems Operational"; no Sep 23 incidents listed on the history page.
Lab verdict: the primary record is internally consistent and every independent duration/timeline claim that can be checked against it passes; the only uncheckable claim is the root cause (never disclosed) and the user-report volume (aggregate self-reports only). The completed checklist above is the exact evidence-status boundary to reproduce in any write-up of this incident.
