Asterix failure-conditioned continuation
Derived from the Astra-authored Asterix network in RikiyaT/lm_understand-logs, credited to Rikiya Takehi and collaborators. Gradient-trained here under official MinAtar 1.0.15 rules.
Continue from published run03 weights under the user-requested average-return objective. Former audit 90000-90127 joins training; all its return/survival regressions enter fixed-label memory. Preserve original native game physics. Fresh final audit 110000-110511 is not used for selection.
Development: {'before_mean': 57.90972222222222, 'after_mean': 59.236111111111114, 'gain': 1.3263888888888888, 'wins': 77, 'losses': 70, 'ties': 141, 'lost_survivors': 12}
Selection objective: mean_return. Individual return and survival losses are disclosed in the audit below. A positive observed mean is not a statistical-significance guarantee.
Held-out audit: {'before_mean': 55.34375, 'after_mean': 55.357421875, 'gain': 0.013671875, 'wins': 140, 'losses': 155, 'ties': 217, 'lost_survivors': 68}
Paired mean-gain bootstrap 95% interval: [-0.828173828125, 0.865234375]