Giving an LLM Pipeline a Memory Without Breaking the Backtest

Contents

• The naive version
• when the call was made
• WRONG
• "the call was made in the past, so it's fine"
• Why gating on entry date destroys the backtest
• The trap this guard exists for: entry is old, but the outcome is not
• yet known on the analysis date.
• The second decision: calibration is reported, never applied
• Below this, a hit rate is noise. Reported with the count attached rather than
• suppressed, but never described as a track record.
• Keeping the backtest reproducible
• What generalizes