E003 — Controllers & Search
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.
Controllers and valid inputs
The controller branch uses fast computer vision and deterministic policies to produce legal movement and firing inputs. Timing is recorded in emulator frames so a policy can be replayed without relying on wall-clock scheduling.
Automated Search is a separate branch
Search evaluates movement and firing policies without frame-level model calls. Scores from this branch are attributed to Automated Policy Search, not to Qwen. Discovery can run in accelerated emulation, but a claimed real-time milestone needs its own confirmation.
What we learned
Perception quality, observation latency and gameplay score measure different things. A useful visual model does not by itself imply a competitive controller. E004 explores a higher-level model role; E005 reports the Search milestone.
Measured checkpoint / 217 completed games
The adaptive search evaluated fixed sweeps, pulsed firing, target positions and visual avoidance. Candidate discovery used accelerated emulation; confirmation ran in real time with the same clean reset and DIP settings. A policy seed controls controller choices, not a verified independent game RNG.
| Confirmed real-time policy | n | Median score | Median survival |
|---|---|---|---|
| pulse-sweep-1.2-12-6 | 18 | 1,740 | 77.660 s |
| pulse-sweep-3-24-6 | 18 | 1,210 | 111.378 s |
| pulse-refine-1.4-16-0.25 | 18 | 1,180 | 76.972 s |
Those repeated native policies are deterministic replicas from the same reset, rather than independent trials. The historical 1,970 accelerated maximum belongs to a separate class and earlier pilot run. Full candidate distributions preserve class, agent, source revision and replay coverage.
What failed / what remains unknown
In accelerated discovery, continuous firing and the original sweep both scored zero in their single tested games. Many fast sweep variants also scored zero. This is useful search feedback, not proof those families always fail. Some CV target/avoidance variants did worse than simple pulsed sweeps. Foreground masks are operational hypotheses; they do not independently label enemies, bullet trajectories or death causes.