E001 — Infrastructure & first gameplay
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.
Infrastructure and gameplay
A Windows research workstation with three RTX 3090 GPUs and 96 GB RAM runs local Ollama. A MAME bridge observes the emulated game and sends real joystick/fire inputs. Input recording, automatic restarts, video rendering and experiment logging form the initial laboratory loop.
The measured pilot contains 100 completed runs: 40 heuristic and 40 random accelerated baselines, plus five real-time games for each of vision, history, telemetry and episodic memory. Windows 11, MAME 0.289 and local Ollama 0.35.1 supplied the runtime. MAME refresh is 61.03515625 Hz; observation images are 208×256. Cheats, RAM writes and save states were disabled.
| Class / agent | n | Median score | Maximum | Median survival |
|---|---|---|---|---|
| assisted_research / heuristic | 40 | 310 | 1970 | 45.47 s |
| assisted_research / random | 40 | 175 | 480 | 34.31 s |
| research / history | 5 | 250 | 390 | 50.33 s |
| research / memory | 5 | 140 | 220 | 62.03 s |
| research / telemetry | 5 | 110 | 190 | 67.04 s |
| research / vision | 5 | 60 | 420 | 21.04 s |
The 1,970-point pilot maximum belongs to the accelerated heuristic branch. It is not a Qwen real-time record. Vision/history prompts exclude telemetry; the telemetry variant receives the checked HUD. Memory retrieval is an experimental observation, not evidence of successful learning.
What we learned
The initial numerical-location probes placed the player correctly in 0/2 cases. That small diagnostic exposed a perception problem; it does not estimate general accuracy. It motivated the independently annotated E002 benchmark. Five games per VLM mode establish that the loop runs, but do not establish a stable causal advantage. Controller revisions and run classes differed across this adaptive pilot.
Round identity, collision causes, ending and rollover remain unknown. Death summaries and retrospective AI explanations are hypotheses. Observation, controller decisions and executed inputs must remain separate records.
Reproduce the setup
The reviewed source-only release includes the laboratory orchestration and Lua bridge; configure separately obtained emulator, game and model dependencies. The original 100 run identities and numeric metrics below were exported from a read-only database snapshot. Source revisions and run classes must be retained in future controlled cohorts.