Methodology
A reproducible research loop
- Define the experiment, version its configuration and record the emulator version.
- Observe game frames or telemetry; log what the controller actually receives.
- Produce valid game inputs and record the action sequence in emulator frames.
- Record score, survival, termination reason, run class and model-call count.
- Confirm significant accelerated discoveries in real time.
- Replay the input sequence and compare captured frames.
- Review the evidence and publication rights before publishing a meaningful result.
Keep the branches distinct
AI perception measures what a model can infer from images. Qwen-controlled experiments use model decisions. Computer vision supports fast observations and controllers. Automated Policy Search discovers deterministic policies, with zero model calls during the run.
Accelerated research speeds up discovery. Real-time confirmation evaluates an execution at normal speed. Replay verification tests whether the recorded inputs reproduce the captured behaviour. These are different claims.
Evidence vocabulary
| Label | Meaning |
|---|---|
| OBSERVED | Recorded directly in the experiment |
| REPRODUCED | Repeated or checked through the reported replay procedure |
| CODE-INFERRED | Deduced from code; not a gameplay observation |
| EXTERNAL SOURCE | Supported by an identified independent source |
| UNKNOWN | Evidence or provenance is not yet established |
The seed site’s labels describe the laboratory’s supplied evidence. They do not mean the website builder independently ran MAME or verified the raw artifacts.
Useful failures
Single-frame perception, model latency, ineffective timing mutations and plateaus are findings worth preserving. We publish limits alongside successes and avoid drawing general model rankings from a single condition matrix.
Hardware and timing
The initial workstation has three RTX 3090 GPUs and 96 GB RAM. E002 cross-model latency was measured on different GPUs. Future comparisons should record GPU assignment, model settings and the complete condition matrix.