When should an enemy forget?
No qualifying case for the preselected seed; no replacement was chosen.
All qualified seeds
| Memory rule | Damage gain vs retained | Alarms | Changes noticed |
|---|
Opportunity lost while learning
Expected regret accounts for the changing A/B mixture and actual tactic probabilities. Lower is better. Four-fight averages; these curves are descriptive.
Locked advancement checks
| Automatic rule | Change-history gain | Stable-history gain | False alarms / 100 | Decision |
|---|
99.583% paired seed-bootstrap intervals; six improvement and six stability checks. Requires ≥1 HP gain, positive lower bound, stable lower bound above −0.5 HP, and ≤1 false alarm per 100 stable fights.
Design and verification
Each enemy has two legal tactics. The task qualifies before any learner runs: the best tactic must change between A and B, with observable rewards and meaningful damage differences. The stable player uses A 90% of the time and B 10%; occasional B fights are not a change in that distribution. Gradual drift changes that mixture over 32 fights. Recurrence switches back after 32.
All branches share their initial memory and random state. Full automatic reset clears both tactic cells in the detected role/context. Selective forgetting clears just the action that triggered. Signaled reset uses the hidden distribution and offline action tables to reset before its best response changes; it is a privileged reference, not a deployable detector or guaranteed ceiling.
Detection windows begin at fight 1, and also 33 for recurrence. A first alarm inside a window counts as operational detection, even though random variation may have caused it. All misses and all later alarms remain in the data. Stable-history alarms are false relative to the stationary distribution.