F-2380Detected Oct 3, 02:30 UTC
What's included in my data access request?
Found in Nightly · v3.4 · 2026-10-03 on support-golden-v4 · Data privacy requests
Model answer vs gold answer
53% of model-answer words appear in the gold answer
Model answer
support-assistant v3.4Your export includes your bills, call records, and the support chat transcripts Unsupported: from the last 10 years.
Gold answer
Human-verified · support-golden-v4The export includes account details, bills from the last 24 months, call and data records from the last 12 months, and support interactions from the last 24 months, as CSV and PDF files.
- claimFlagged by the judge as unsupported or contradicted
Retrieved context
bge-m3 · hybrid (BM25 + dense, RRF) · top 3 passed to the generator
2 of 3 retrieved chunks are relevant · gold passage ranked first
- 1
Requesting a copy of your data
What we include · kb-privacy-002 · v2026.09
RelevantBills (24 months), usage records (12 months), support interactions (24 months).
Similarity0.87 - 2
Requesting a copy of your data
Format · kb-privacy-002 · v2026.09
RelevantDelivered as a ZIP with CSV and PDF files.
Similarity0.80 - 3
Deleting your personal data
Retention exceptions · kb-privacy-005 · v2026.09
Not relevantSome billing data is retained for 10 years for tax purposes.
Similarity0.72
Judge verdict
LLM judge · Claude Opus 5.5 · rubric v3
0.29
rubric score · pass threshold 0.70
The 10-year figure belongs to tax retention, not the export scope listed in chunk 1.
Diagnosis
Unsupported claims · Privacy request timelines
Root cause
Numeric bleed from a retention chunk.
Suggested fix
Route privacy intents to verified templates (same as F-2400).
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