F-2383Detected Oct 7, 02:36 UTC
I got a text saying my account is locked and to click a link. Is it from Harbor?
Found in Nightly · v3.4 · 2026-10-07 on support-golden-v4 · Account & security
Model answer vs gold answer
28% of model-answer words appear in the gold answer
Model answer
support-assistant v3.4Harbor sometimes sends account alerts by text. Unsupported: If the link starts with harbor, it is safe to use.
Gold answer
Human-verified · support-golden-v4Harbor never sends links asking you to unlock your account. Do not click it. Forward the text to 7726 and report it in My Harbor under Security > Report fraud.
- claimFlagged by the judge as unsupported or contradicted
Retrieved context
bge-m3 · hybrid (BM25 + dense, RRF) · top 3 passed to the generator
1 of 3 retrieved chunks are relevant · gold passage at rank 3
- 1
Recovering your account
Identity verification · kb-account-018 · v2026.09
Not relevantWe verify your identity with a code sent to your email.
Similarity0.74 - 2
Two-factor authentication for My Harbor
Turning on 2FA · kb-account-011 · v2026.09
Not relevantTurn on 2FA for extra security.
Similarity0.71 - 3
Reporting fraud or suspicious activity
Suspicious calls and texts · kb-account-024 · v2026.09
RelevantHarbor never sends links to unlock accounts. Forward suspicious texts to 7726.
Similarity0.70
Judge verdict
LLM judge · Claude Opus 5.5 · rubric v3
0.01
rubric score · pass threshold 0.70
Dangerous advice not supported by context; chunk 3 states the opposite.
Diagnosis
Unsupported claims · SIM swap and fraud
Root cause
Fraud article ranked third behind generic account articles.
Suggested fix
Add a phishing intent route with a verified template answer and boost kb-account-024.
Playbook: unsupported claims
The answer states something that is not present in the retrieved context.
Detection signals
- Claim-level faithfulness below 0.85
- Numbers, fees or durations absent from every retrieved chunk
- Citations that do not contain the cited sentence
Common fixes
- Require a citation after each factual sentence and refuse when none applies
- Reject answers with numeric claims missing from context in a post-generation check
- Lower temperature for billing and policy intents
- Add the query to the regression set with an escalate-or-decline gold answer
Activity
1 event · 0 comments
Relevant ·
Detected by support-assistant v3.4 run and labeled unsupported claims