E004 — Qwen selection & timing
- Games
- 30
- Overall verified maximum around this phase
- 2,020
- Qwen selector maximum, approximately
- 1,740
Validated checkpoint: 2026-10-05. Data last synchronized: .
Latest record verified from reviewed published run evidence. Registry counts retain the last full dataset checkpoint; record publication does not refresh the complete dataset. This is a snapshot, not a live counter.
Registry count checkpoint: 2026-10-05T08:28:16.420013+00:00. The latest reviewed record and its evidence update separately from full dataset totals.
Higher-level selection
E004 tested Qwen as a policy selector instead of requesting every joystick action directly. It examined whether selecting a controller and its timing could improve gameplay.
What we learned
Giving the language model a higher-level controller role did not automatically outperform deterministic policy search. The Qwen selectors remained around their reported maximum while the overall verified maximum around this phase was higher.
This result narrows the next questions. Classification, strategy, skill selection and failure analysis remain promising model roles. It does not establish that all model-based selectors are ineffective.
The repaired 30-game cohort
All games used real-time execution and a clean reset. Fixed native policies were tested three times each; scheduled skill changes generally had one trial. The best scheduled policy in this cohort reached 2,020, with 92.897 seconds survival and an exact endpoint replay. The page’s phase-level maximum above is a separate checkpoint figure, not a controlled leaderboard.
| Qwen selector | Games | Responses | Selected p0 | Accepted p0 | Stale rejections | Median strategy latency | p90 observation age |
|---|---|---|---|---|---|---|---|
| 8B instruct / single4 | 3 | 88 | 88 | 85 | 3 | 1.296 s | 84 frames |
| 9B | 3 | 48 | 48 | 45 | 3 | 2.313 s | 150 frames |
Both selectors always chose the native fallback p0 and reached 1,740 points, the same score as the fixed p0 replicas. That outcome supplies no causal evidence that Qwen improved the policy. Confidence was model-reported and uncalibrated; an E002 player-location result does not establish enemy or bullet understanding.
What failed / preserved recovery
The first batch failed during Windows CP1252 log counting. Four durable completed rows and one recovered game were explicitly inherited by the repaired UTF-8 batch, retaining both source snapshots. The recovered run has no recoverable monotonic wall time: it remains null. No already-played case was repeated to hide the failure.
Replay coverage is limited: a successful replay verifies the corresponding recorded inputs and endpoint, not every game in the cohort. Small samples and deterministic replicas do not establish generalization. The original sealed summary, analysis, protocol and case registry retain these limitations.