Model repeatedly fails to verify day-of-week from date output, causing incorrect data assessments

Summary

Claude (Opus) repeatedly misidentifies the day of the week when reading date command output, leading to incorrect assessments of time-sensitive production data.

Reproduction

1. Ask Claude to check the current time and assess system health
2. Claude runs date which outputs e.g. 2026-04-27 14:46:42 IDT
3. Claude assumes a day-of-week (e.g. "Sunday") WITHOUT verifying — the date command's default format doesn't include the day name
4. Claude then interprets all downstream data through the wrong lens (e.g. "markets closed, weekend gaps expected" when it's actually Monday and alerts are real)

Impact

This has happened twice in production use:

• Apr 21 (Tuesday): Claude said "weekend gap, will refresh at market open" — was actually Tuesday, stale data was a real bug
• Apr 27 (Monday): Claude said "Sunday, markets closed, weekend-expected gaps" and dismissed 3 URGENT pipeline-dead alerts as normal. It was Monday, markets open, alerts were real production issues.

Expected behavior

After running date, Claude should verify the day-of-week explicitly (e.g. date +%A or python3 -c "from datetime import date; print(date.today().strftime('%A'))") before making any day-dependent assessment.

Environment

• Model: Claude Opus (claude-opus-4-6)
• Tool: Claude Code CLI
• Context: Production monitoring workflow where correct day-of-week determines whether data gaps are normal (weekend) or critical (trading day)

Notes

The workaround is a memory/prompt rule forcing day verification, but the underlying issue is that the model confidently infers day-of-week from a date without calculating it — and gets it wrong. This seems like a known weakness in LLM date reasoning.