Problem
The agent routinely gets basic arithmetic and counts wrong. Documented instances:
- Said Google Drive free storage was ~5.49 TB (correct: 4998 GB)
- Said 3 spam emails existed when only 2 did
- Said an email was sent "6 hours ago" when it was 35 minutes
- Agent was stuck on wrong date (gave briefing for May 25 when it was May 28)
Root cause
The default cloud model doesn't verify numerical claims against tool outputs before speaking. It paraphrases and "rounds" inaccurately without cross-checking.
Suggested fix
Add a post-processing step — any numeric claim in the response must match the raw tool output, or the agent must cite the discrepancy. Alternatively, add a system prompt rule: "Never paraphrase numbers. Quote exact values from tool outputs."
Evidence
- Multiple Slack reports from Alan documenting inaccurate responses
- Episodic memory entry documenting "Time Awareness / Context Ignorance" pattern
- Pooja requesting hallucination guardrails for wrong-date briefing
Problem
The agent routinely gets basic arithmetic and counts wrong. Documented instances:
Root cause
The default cloud model doesn't verify numerical claims against tool outputs before speaking. It paraphrases and "rounds" inaccurately without cross-checking.
Suggested fix
Add a post-processing step — any numeric claim in the response must match the raw tool output, or the agent must cite the discrepancy. Alternatively, add a system prompt rule: "Never paraphrase numbers. Quote exact values from tool outputs."
Evidence