Fewer resets. Still no improvement.
Separate observation batches reduced unnecessary interventions. None of the three candidates improved changing-history damage enough to advance. The original learner remains in the game.
Known-case development · 48 cases · 259,584 cached policy transitions · zero new scored physical fights
All 16 seeds for this enemy
Candidate minus unchanged learning, HP per fight. Changing results equally weight abrupt, gradual and recurring histories. Descriptive 95% paired seed-bootstrap intervals on previously examined cases.
| Candidate | Changing histories | Stable history | Interventions / 100 stable fights | Decision |
|---|
Damage lost to imperfect choices
Expected regret uses the current A/B mixture and the actual choice probabilities. Lower is better. Sixteen-fight averages; all seven arms are included.
Selection, controls and reproducibility
The original overlapping-window rule is the failure reference. Down-only reset changes intervention direction. Down-only softening halves scores and retains counts. Separate batches change the detector evidence while keeping that same soft intervention. Frozen memory and a reset with hidden change timing are controls.
The overlapping-window candidates can react twice to one outlier. Both fail that stress check. The separate-batch candidate passes the stress and stable-performance checks but fails the required 1 HP changing-history gain in all roles. No fresh validation was run.