# V14a Forgetting Curve Proof

## Counter-Revision Cannot Model Forgetting

Repeated revision with uncertainty prior (stv 0.5 0.3):

| Round | Prior | Counter-Evidence | Result | Conf Trend |
|-------|-------|-----------------|--------|------------|
| 0 | - | - | (stv 0.9 0.8) | baseline |
| 1 | (stv 0.9 0.8) | (stv 0.5 0.3) | (stv 0.861 0.816) | +0.016 |
| 2 | (stv 0.861 0.816) | (stv 0.5 0.3) | (stv 0.829 0.829) | +0.013 |
| 3 | (stv 0.829 0.829) | (stv 0.5 0.3) | (stv 0.802 0.841) | +0.012 |

## Conclusion
- **Frequency** drifts toward 0.5 (forgetting direction) ✓
- **Confidence** monotonically INCREASES (anti-forgetting) ✗
- Revision always pools evidence — more evidence = more confidence
- **Forgetting REQUIRES pre-inference temporal discounting** (Mechanism 1)
- Formula: conf_effective = conf * exp(-lambda * staleness)
- Applied BEFORE revision entry, not through revision itself